Intelligence is getting radically cheaper. Useful work is not.

Here is the calculation seducing every boardroom right now: a model completes in seconds a task that takes an employee thirty minutes. The API bill is twelve cents. The employee costs €30 an hour. Multiply across a department and the business case appears to write itself.

But twelve cents buys an output. It does not buy a completed, correct, accountable piece of work. And the gap between those two things, the output and the finished job, is where the entire economics of AI at work actually lives.

The argument cuts against how AI is usually sold. The claim is not that AI is expensive. It is that the token price is the smallest, most visible tip of a cost structure that is mostly invisible, and that for a large share of real workflows, the full cost of an AI system lands above the cost of the person it was supposed to replace. Not always. But far more often than the twelve-cent calculation implies, and in ways that only show up in the second year, on the P&L, long after the pilot deck has been applauded and filed.

We have made a version of this argument before. In The AI Adoption Illusion we argued that counting prompts per employee measures activity, not value. In Tokenmaxxing and the Industry Transition Radar we showed that even the best-instrumented companies can’t yet measure AI productivity reliably: 88% of organisations used AI in at least one function in 2025, yet only 7% called it fully integrated. Put those findings together: if adoption is the wrong input metric and tokens are the wrong currency, then cost per token versus salary is the wrong economic comparison. Here is the right one.

The question worth asking is not “how cheap are tokens?” It is this: what happens when we compare the full cost of an AI system with the full cost of an employee?

Tokens are not labour

Compare what the twelve cents actually covers against what the salary actually covers.

The salary buys a person who receives an ambiguous request, asks a clarifying question, retrieves the context, produces the work, checks it, fixes it, takes responsibility for it, and shifts to a completely different task at 2pm. The twelve cents buys one inference pass.

We explored this task-versus-work distinction in Jobs and AI: Chains of Work: real jobs are chains of interdependent tasks, and every unreliable handoff creates another place where human judgment has to return.

To turn that inference pass into the same completed work, a production system also needs: software and model subscriptions, context retrieval and data preparation, orchestration across multiple agents or model calls, retries and exception handling, human review and approval, evaluation and monitoring, security, governance, integration, maintenance, and the rework caused by unreliable outputs.

Pricing AI labour by token cost is the equivalent of pricing human labour by the electricity that runs the laptop. It is a real cost. It is also the smallest one on the invoice.

The unit that matters is not the cost per response. It is the cost per acceptable outcome.

The energy industry faced a similar accounting problem decades ago and gave it a name. Nobody prices a power plant by the cost of one kilowatt-hour of fuel; the sector uses levelized cost of electricity, which loads capital, operations and lifetime output into one comparable number. A 2025 proposal adapts precisely this logic to AI, a “Levelized Cost of AI” that divides total capital and operating expenditure by valid output, on the argument that token costs and GPU-hours systematically understate what a system costs across its life. Later, we translate that into a better dashboard.

The anatomy of digital labour

Every AI system that does real work carries three layers of cost.

Intelligence. The models, tokens, inference and tool calls required to generate an answer. This is the layer on the pricing page, and the only one falling fast.

Reliability. The evaluations, supervision, retries and guardrails required to trust the answer. Nobody puts this on a pricing page.

Organisation. The infrastructure, integration, governance and behavioural change required to use the system at scale. This layer does not appear anywhere until the second year of the programme.

Why is the first layer cheap and the second one expensive? Two mechanisms, both now formalised in the economics literature.

The first is the convex cost of accuracy. Because models show diminishing returns to data, compute and size, the cost of higher accuracy rises steeply: good performance is cheap, near-perfect accuracy is disproportionately expensive. The practical result: full automation is often not cost-minimising. Partial human–AI collaboration can be the rational long-run outcome.

The second mechanism matters even more. Picture two cost curves. Automation costs fall as compute, data and models get cheaper. Verification costs, the human attention, judgment and accountability needed to trust an output, barely move because they are bottlenecked by human biology rather than silicon. As execution becomes abundant, verification bandwidth becomes the binding constraint on value.

Look at the wedge between those curves. That wedge is where the economics of digital labour actually live. Everything below it gets cheaper every quarter. Everything inside it is your new cost structure, and it is not going away.

Paying subscribers can keep reading on and access our AI Employee Calculator. Consider subscribing to access full essays, our community and more!

Read more

We would love you to join us for the next Work3 Brains Trust on Wednesday, 22 July 2026 at 5 pm UK / 12 pm ET.

The Brains Trust is our space for paid subscribers to come together, test ideas and explore the questions shaping the future of work.

A job is not simply a contract for paid labour. For generations, it has also provided income, learning, identity…

Read more

The CEO of a bank recently announced plans to eliminate 7,800 jobs by 2028.

In his comments about how AI would allow the replacement of back-office jobs, he said,

“It’s not cost cutting. It’s replacing, in some cases, lower-value human capital with the financial capital and investment capital we’re putting in.”

This is, sadly, a fairly normal announcement these days, but he used fairly pernicious language about his workers.

Bill Winters is the CEO of Standard Chartered which employs about 82,000 people. There was a bit of a backlash in the media about his description of workers as ‘lower-value human-capital’.

Whether there will be long term damage from these comments remains to be seen. The share price is up about 12% since the announcement.

The bank will still have around 74,000 employees after the cuts. I wonder how some of the remaining medium and high-value human capital feel about the phrase.

But this article is not really about one CEO or one bank.

Maybe Bill’s real crime was articulating a little too clearly how the top brass talk about the AI shakeout.

Because ‘lower-value human capital’ did not appear from nowhere. It comes from a long line of 20th-century managerial thinking, layered over with 21st-century consulting, finance and marketing language.

We should not be surprised by dehumanising language at work. The surprise is that we still pretend it is accidental.

So here are three phrases I would put straight into Corporate Room 101.

Corporate Room 101 #1 - Human Resources

The phrase ‘Human Resources’ has its roots in the dismal science of economics.

It carries an accounting worldview: people are viewed as productive resources to be developed, allocated and managed in the service of organisational output.

In this mindset, the organisation is a system of inputs: land, labour, capital, materials, equipment, money - and then, eventually, ‘human resources’.

This matters because ‘resource’ is not an innocent word.

A resource is something used to achieve an end. So even when the intent was progressive - the language still placed people inside an instrumental frame.

Referring to your workforce as a resource reduces people to the role of a utility - like a machine, capital, or a donkey.

Corporate Room 101 #2 - Talent Management

The phrase ‘Talent management’ is more recent. It was popularised by McKinsey’s late-1990s ‘War for Talent’ work.

Don’t start me on military metaphors at work.

Of course, organisations have always had workforce planning, succession planning, employee development and career management. But Talent Management was born in consulting and popularised by HR vendors because it made ordinary workforce processes sound strategic, scarce and expensive.

Why call it workforce management, with performance and pay plotted on boring distribution curves, when you can pretend everyone is ‘talent’?

If an organisation has 10,000 workers, they are not all ‘talented’ in the same way.

Some are brilliant. Some are struggling. Most sit somewhere in the middle, doing useful work under imperfect conditions inside systems they did not design.

Talent Management was vendor lipstick on a 20th-century pig: a shinier label for the old machinery of ranking, segmenting, acquiring, retaining, deploying and exiting people.

Corporate Room 101 #3 - Human Capital

Whether lower value or higher value, what a horrible way to describe a person.

Human Capital is more directly rooted in economics and investment logic. It is associated with economists such as Theodore Schultz and Gary Becker, where education, skill and knowledge are treated as forms of capital investment.

So if you want to make the ‘financial people’ point, human capital is the cleaner target than ‘human resources’.

If ‘Human Resources’ says people are resources to be utilised.

Then ‘Human Capital’ says people are assets to be bought, upgraded, sweated and sold.

Watch your language folks.

Moving on Up the Value Chain

The way we talk about work has been shaped by economic and corporate models from the 20th century, with an added layer of American consulting pizzazz.

Now AI is forcing another round of automation, restructuring and repricing.

Automation in work has been going on since industrialisation and the digital revolution of the 1970s. But this one is hitting the office, the laptop class, the back office, the analyst, the junior worker, the support function and the knowledge worker.

Those who lead the AI shakeout need to reflect on the language they use for the people they will need to manage the machines.

This poses a few questions :-

• What workplace phrases do you most dislike?

• In an AI economy, who gets to decide which humans are ‘lower-value’?

• What does your language reveal about how you see work?

Final thought - using language that motivates and engages rather than disenfranchising might just be a differentiator in competitive labour markets.

If leaders talk about people as disposable units of capital, they should not be surprised when people treat the organisation as disposable too.

Moving on up the value chain,

Andy

This edition of Work3 is sponsored by Glean's launch of Independent Agents.

If the job is being unbundled, so is the org chart. Glean's new independent agents are designed as AI coworkers, with their own identity, memory, and accountability, built to take on the kind of work that used to fall to the newest person in the room. Worth a read alongside this essay's question about who trains the next generation when AI gets there first.

Read more about Independent Agents

Will AI take our jobs?

It is an emotionally charged question.

But it is also too narrow.

The deeper question is this:

How do we build a fair society that does not depend so heavily on the job?

A job is not just a contract for paid labour. It is the container through which modern societies distribute income, purpose, status, training, identity, time, benefits, pensions, security, and belonging.

That is why the AI jobs debate cuts so deep.

We don’t just fear unemployment, we are afraid that the whole social machinery built around employment may be lost too.

So what can we do?

Some will wait for the job apocalypse. Others will use AI to eat market share.

But the harder task is to ask what happens when work is deconstructed - and how we rebuild the institutions that once made employment feel secure, meaningful and fair.

The Job Was Never Just a Job

For much of human history, people worked without having what we would now recognise as a ‘job’.

They farmed, traded, cared, cooked, repaired, hunted, gathered, built, taught, healed, raised children, maintained households and contributed to communities.

Remote work is not new. For example in Britain from the 1600s to the mid-19th century, workers made dresses, shoes, and matchboxes in their kitchens or bedrooms.

With mechanisation, from the spinning jenny to the power loom, work that had once been scattered across homes and communities was increasingly pulled into factories.

And the modern job was created.

We created the 12-hour day, fought for the 10-hour day, normalised the five-day week and invented the weekend.

The job was never a natural law. It was designed, contested and redesigned.

That is why today’s debate should not stop at shorter weeks, flexible hours or better benefits. Those may matter, but the deeper question is what happens when the job itself can no longer carry all the social functions we attached to it.

For some, the job gave meaning, identity, status, structure and purpose.

For most, it put food on the plate.

For society, it became a mechanism for redistribution, education and security.

But the job, and many of the institutions we built around it, are beginning to fracture.

No More Absolute Beginners

The danger is not only a world without jobs. It is a world without absolute beginners.

The routes into work are becoming narrower, harder and more fragile.

Graduates in the US have been booing the mention of AI at commencement or graduation ceremonies. Of course this talented generation are using AI and the extraordinary tools in their hands.

This protest is not about technology but the social contract around work.

Earlier generations could plausibly believe that education, hard work and entry-level employment would lead to career progression, home ownership and financial security.

For many younger people, that bargain already looks broken. This is why the death of entry-level jobs matters.

The UK gives us a glimpse of what happens when the routes in start to fail. Alan Milburn’s interim report warns that more than one million 16–24-year-olds are not in education, employment or training - around one in eight young people. Nearly six in ten economically inactive young people have never had a job. Yet 84% of these young people surveyed say they want work, education or training.

So this is not a simple story of young people opting out. It is a warning that the labour market is no longer reliably letting them in.

Young people do not only need skills.

They need routes in.

Join the Work3 Brains Trust

If the job can no longer carry the full weight of income, learning, identity, security and progression, what should take its place?

This is the question we will explore in next week’s Work3 Brains Trust for paid Futurist subscribers on 22nd July 2026 - 5 pm UK / 12 pm ET

We will discuss:

• whether entry-level jobs are really disappearing, or being redesigned

• what young people need if the first rung of the career ladder is weakening

• how employers should rebuild learning when AI absorbs junior tasks

• whether AI dividends and new forms of ownership are realistic

• what a post-job social contract might look like in practice

The Brains Trust is open to paid Futurist subscribers.

For paid Futurist subscribers: continue reading for why the post-job world is already here, what comes after the job, and five functions we need to rebuild - including the AI dividend.

The Post-Job World Is Already Here

The post-job world is not a prediction. It is a distribution problem.

For some, it already looks like freedom: portfolio careers, fractional roles, creative independence, self-employment, AI-enabled micro-enterprises and the ability to work beyond one employer, one office or one job description.

For others, it looks like instability: insecure hours, weak benefits, algorithmic management, low bargaining power and no obvious path forward.

This is why ‘more good jobs’ is an appealing answer, but not a complete one.

Three uncomfortable realities complicate the story.

First, most work in the world already happens outside the secure formal job. Of the roughly 3.4 billion people in the global workforce, around 2 billion work in the informal economy, often without contracts, benefits or meaningful legal protection. A world beyond formal jobs is not some distant future. It is already the daily reality for much of humanity.

Second, flexible and independent work is no longer marginal. The global gig economy was estimated at $557bn in 2024 and is projected to triple by 2032. Around 14.5% of EU workers are self-employed. China estimates it has around 200 million “flexible workers”. Some people are pushed into this world by necessity. Others choose it because they do not want the obligatory boss, fixed schedule or single-employer career. Either way, work is increasingly leaking out of the traditional job container.

Third, even having a job no longer guarantees security. In the UK, almost seven in ten working-age adults in poverty live in a household where at least one adult is in work. In the US, more than 40% of people below the poverty line live in households with at least one worker. The old promise was that employment would be the route out of insecurity. For too many people, that promise no longer holds.

So the old distinction between “employed” and “unemployed” is becoming less useful.

The real divide may be between those who have access to income, learning, assets, networks and agency - and those who do not.

The post-job world is not necessarily a leisure society, as Arthur C. Clarke imagined. It may be a society with more work, but less of it contained inside the traditional job.

It is what happens when the job stops being the main way people get income, learning, identity, security and belonging - and what we need to build next.

Five Functions to Rebuild

What comes after the job is not a world without work.

It is a world where the old job bundle has to be rebuilt.

For the past century, the job has been asked to do far more than organise paid labour. It has carried income, security, learning, status, belonging and a route into the future.

If that bundle is now coming apart, we should stop pretending the answer is simply “more jobs” or “better skills”.

In previous Work3 essays, we explored 16 policies for a post-job world. But underneath the policy debate are five human needs that any serious post-job social contract has to answer. Five stand out.

1. Financial Security after the job

If work becomes more fragmented, income becomes more volatile.

That does not mean everyone needs a universal basic income tomorrow. But it does mean societies need better ways to smooth income shocks, support transitions and prevent temporary disruption becoming permanent exclusion.

The question is whether people can take risks, retrain, care, move, build and adapt without one lost job turning into a lost life.

The most important point is this: financial security has already been partly dismantled from employment.

The old job used to bundle income with pensions, benefits, sick pay, training and progression. As those links weaken, security needs to become more portable.

The question is not whether people should work.

The question is whether they can take risks, retrain, care, move, build and adapt without one lost job turning into a lost life.

Example: Japan’s Fureai Kippu (“Caring Relationship Tickets”) allows people to earn credits by caring for elders, redeemable later in life.

2. Learning after the job

If AI weakens the entry-level ladder, learning cannot remain a one-off event at the start of life.

We will need more apprenticeships, not fewer. More work trials. More project-based learning. More mid-career transitions. More employer investment in beginners. More ways to prove capability outside degrees and job titles.

AI tutors may help. Digital credentials may help. Skills accounts may help.

But the deeper challenge is institutional.

Who is responsible for developing people when employers want ready-made talent, schools cannot keep up with labour-market change and AI absorbs the low-level work that used to train beginners?

Example: Singapore’s SkillsFuture gives adults state-backed credits for approved training.

3. Status after the job

Jobs provide status.

“What do you do?” remains one of the most common social questions. We still use job titles as shortcuts for competence, class, ambition, identity and social worth.

But in a post-job world, many people will not have one neat answer.

They may be building, caring, learning, consulting, freelancing, investing, volunteering, creating or experimenting. They may be between projects. They may be doing economically useful work that does not fit inside a job title.

So we will need new ways to recognise contribution.

Portfolios. Credentials. Reputation systems. Peer networks. Proof of work. Community roles. Public contribution. Ownership stakes. Learning records.

This matters because people do not only need money.

They need dignity.

4. Belonging after the job

The workplace has often been a flawed community.

It can be political, exhausting, hierarchical and exclusionary.

But it has also been one of the main places where people meet others, build friendships, learn norms, find mentors and feel part of something bigger than themselves.

As work becomes more remote, fragmented and AI-mediated, belonging will not happen automatically.

Where do people find community outside the traditional workplace?

Professional guilds may return. Local work hubs may matter. Online communities may become more serious. Alumni networks may become lifelong career communities. Co-operatives, creator networks and member-led institutions may take on more importance.

The future of work is also the future of belonging.

5. Ownership after the job

This may become the biggest question of all.

If AI creates enormous wealth, where will that wealth go?

Without new models, most of the upside will flow to the shareholders of technology companies.

In the US, politicians have floated versions of citizen dividends from AI businesses.

This might be easier in the US or China, where AI champions are domestically based. But what about countries without AI market leaders?

This kind of redistribution is not so far from reality. Many workers are already capitalists, with shares in their retirement accounts and pensions in Nvidia, Alphabet, Apple etc

If income once paid to workers increasingly appears as profits in AI companies, chipmakers, data centres, cloud providers or infrastructure owners, then societies will face a distribution problem as much as an employment problem.

Other ideas include tax AI profits, levies on data centres, higher taxes on capital and lower taxes on labour, sovereign wealth funds.

There are problems with all of these ideas.

Governments may pick the wrong companies. Tax systems are slow. Tech firms are skilled at shifting profits across borders. Public ownership can create political conflicts. Data-centre taxes may be passed on. And countries without large AI champions cannot simply seize shares in foreign firms.

But the ownership question will not go away.

If AI makes a small number of firms and asset owners vastly richer while many workers face weaker bargaining power, unstable income or fewer routes into good work, politics will become ugly fast.

The post-job world cannot only be about safety nets.

It also has to be about who owns the machines, platforms, data and institutions through which future work flows.

Our Most Important Work

The job is not disappearing overnight. Many jobs will survive, some will become more valuable, and new ones will be created.

But the job may no longer be strong enough to carry the full weight of income, learning, identity, security, belonging and social mobility.

That leaves us with the most important work of all: rebuilding the institutions that help people build lives.

The real question is not whether AI takes all the jobs, whether we should slow down technology, or whether everyone ends up idle.

It is what jobs used to do for society - and what happens when they stop doing it.

What comes after the job?

Not one policy, but new institutions, new ownership models and new ways to organise useful human contribution.

Work will not disappear - but the old job can no longer carry the society we built around it.

Post-job since 2006,

Andy

The 2026 World Cup will be the most watched football tournament in history.

It may also be the most analysed.

Every run, pass, press, tackle, shot and goal will be captured, coded, and interpreted.

FIFA’s 2026 Technical Study Group will analyse all 104 matches through multiple video angles and thousands of live data points.

Football remains gloriously human. It is emotion, skill, luck, pressure, exhaustion, resilience and teamwork. The best individual does not always win. The best team does not always win either.

But over time, the best systems usually give themselves a better chance.

So, as the World Cup begins, what can the world’s most analysed sport teach us about the work of us mere mortals?

Subscribe now

The World in Motion: When Work Becomes Measurable

Most professional footballers wear a tracker vest under their shirts. These trackers collect data during a game, which, along with video, includes details of every pass made, tackle made, shot taken and goal scored. The England Lionesses use this technology to have detailed reports on individual players from training sessions.

Managers can track this data, plus heart-rates for an individual player.

The Lionesses also wear GPS trackers during training sessions, which generates another 1.3 million lines of squad data per session. This contributes insights into the volume and intensity of the games each footballer has participated in over the course of a season. Gathering these benchmarks helps the players and coaches understand changes and risks for injuries and peak performance levels.

The smart vests are used by the most successful World Cup team of all time, Brazil.

The tracking technology is used to monitor individual performance through the season to spot injury risk, and also team selection, and substitutions.

So having this player data is one thing, using it to win tournaments is another challenge.

Systems 1, Players 0

Big expectations are not enough to win a long tournament like the World Cup.

In 2018, the reigning champions Germany came bottom of the group stages after losing to Mexico and South Korea. Reigning World and Euro Champions France, in 2002, went home with one point and no goals.

You can’t just put a great player, or team, in a bad system and expect it to work out.

It might be argued that the last World Cup Winners, Argentina, won because Lionel Messi, won the cup almost single-handedly.

However, Argentina did not win because Messi was brilliant in isolation. They won because a robust team was built around Messi’s qualities.

Interesting to read the FIFA match report - Messi does not have the best stats on the pitch.

Overall organisational health is key for a team to be successful.

Skills Ecosystems

Talking of Lionel Messi, he has captained his country to win the World Cup, won a record eight Ballon d’Or awards, and 34 trophies with Barcelona.

He is regarded as the Greatest Player of All Time 🐐 – but how do we know?

Here is a deep-dive player analysis short video on Lionel Messi from Soccerment.

It gives his season stats, spider-charts with percentile rank compared to other players in the league. It shows his open play expected assists, expected threat from passes, completed passes per 100 touches, heat-map analysis of positions.

I am not making the argument that he is the best (I do live in Portugal after all) but that using this type of performance data on individual and team performance is fascinating.

Total Football – Changing the Rules of the Game

So if each top team has a few brilliant players, and a few more workmanlike, how do we develop a successful winning organisation?

We can play by the rules, competing pretty much the same way as the others, make some innovations, and win our fair share of games.

But. then a new approach arises, or a new technology, and slow, slow, quick - people jump ship. Sometimes fresh innovation emerges and makes you fall off your chair.

In the 1974 World Cup, the Netherlands replaced Brazil as the crowd pleasers, and made it to the final of the World Cup. Playing expressive, fluid football - it was an unconventional strategy that effectively did away with positions - players were interchangeable on the field at all times.

Total Football (or Totaalvoetbal for my Dutch friends) was a style of play in which success thrives on a collective understanding of space and movement among all 11 players. Associated with Cruyff, Rinus Michels, Ajax, and the Netherlands national football team.

Total Football wasn’t solely a gameplay style; it embodied a mentality that challenged the confines of convention and urged players to transcend their predefined roles.

As with most innovations, it didn’t spring forth in isolation; rather, it was an adaptation of strategies employed by exceptional teams of the past.

The Cruyff Turn - Total Football is much more than the ‘Cruyff Turn’, but this makes me smile every time…

A Deeper Dive Into Football Network Analysis

This diagram is a schematic illustration of a football passing network. In the plot, players are represented by circular nodes, whose size is proportional to their importance in the network. The position of each player is given by the average of the positions of all passes made by the player during the match. The width of the links is proportional to their weights, which account for the number of passes between players. The example is a match between Barcelona and Real Madrid.

Those who are particularly interested can read the paper in Nature - Defining a historic football team: Using Network Science to analyse Guardiola’s F.C. Barcelona.

And also, for a deeper look at the game, this is a useful resource -> FIFAs Match Report Hub from the World Cup 2022

A short personal note ⚽

My love of football started in Manchester, in the heated red-and-blue rivalry of United and City. Since then, it has taken me from the terraces of Old Trafford to Brighton’s rise into the Premier League, Benfica’s eagles🦅 in Lisbon, a Guatemalan cup final involving giant avocados🥑, the Maracanã in Rio, and La Bombonera in Buenos Aires. So yes, I know football is not really a matter of life and death. But Bill Shankly’s joke still lands because football gives us something rare: a shared emotional language for hope, belonging, failure and leadership.

Extra-Time

Football analytics does not remove the mystery of the game.

It does not explain the Cruyff Turn. It does not measure courage in a penalty shoot-out. It does not tell us why a team suddenly believes, or why a favourite freezes when the pressure arrives.

But it can help us see patterns we used to miss. It can show how players connect, where systems fail, when fatigue rises, and how a team creates space for talent to flourish.

That is the lesson for work, and people analytics.

The organisations of the future will not win because they collect the most data. They will win because they use data to build better systems for human judgement, trust and performance, or hope they have a genius like Cruyff or Messi on the team.

The dashboard is not the manager. The model is not the match. The data is not the team.

As the final whistle blows, football’s blend of skill and chance still mirrors the world of work. Some moves are planned. Some are improvised. Some hopeful crosses fly out of play. Others find the back of the net.

The ball is now at your feet.

Good luck!

With fingers crossed,

Andy

Every organisation right now is trying to answer the same question: where are we on the AI curve? Most are measuring the wrong thing to find out. The tools they're using seat counts, token dashboards and usage leaderboards tell us how much AI the organisation is consuming. They don't tell whether the structure of work has actually changed around it. That distinction is becoming expensive to ignore, which is why this week we want to focus on surfacing these mistakes and providing some guidance and framework on how to fix it.

Let's dive in!

Tokenmaxxing

On May 29, Amazon shut down KiroRank. The internal leaderboard had tracked AI token usage (basic units of AI computation) across the company's developer workforce, ranking engineers by how much they used Kiro, Amazon's own AI coding platform. The problem wasn't the platform, it was that employees had figured out how to game the ranking by running autonomous AI agents to generate activity, inflating token counts, climbing the board. The SVP responsible sent a message to staff that should have been unnecessary to send: "Please don't use AI just for the sake of using AI."

The same week, Meta took down Claudeonomics, a grassroots leaderboard built on the company's intranet. It had tracked token consumption across more than 85,000 employees, awarding titles like "Token Legend" and "Session Immortal" to the top 250 users. Total consumption in the preceding 30 days: 60.2 trillion tokens, roughly $900 million at API prices, likely over $100 million at Meta's negotiated rate. Most of it was deliberate waste: engineers ran agents for no outcome, built throwaway features, defaulted to AI when hand-coding was faster, because being seen as "not AI-native" felt like a job security risk. Some production incidents were traced to AI-generated code from developers who had prioritised volume over quality and the leaderboard came down one day after the story went public.

Microsoft has had a comparable system running since January: one developer described asking AI questions already covered in internal documentation, prototyping features intended never to ship, and always defaulting to agents even when slower, specifically to avoid being flagged for low token usage in performance reviews. Salesforce enforces a $100-per-week minimum spend on Claude Code via a Mac widget that updates every 15 minutes, with colleagues' spend visible to anyone. Uber burned through its entire 2026 AI coding budget in four months, then imposed a $1,500 monthly cap per employee per tool after its COO admitted he couldn't draw a line between rising expenditure and useful features shipping to customers.

The only visible outlier has been Shopify. They renamed their leaderboard a “usage dashboard” and added circuit breakers that cut access when spend spiked unexpectedly to catch both runaway agents and infrastructure bugs, as well as having leadership personally review the accounts of the highest spenders.

What that review found was a ratio that inverted everything: the number that mattered wasn’t total token spend, it was tokens consumed per unit of meaningful output produced. The ratio runs counter to intuition: the highest cost-per-unit indicates the most valuable work. Genuinely hard problems like novel architecture, complex debugging and non-trivial design, require many iterations and long reasoning chains. They’re expensive per unit precisely because they’re real. On the other hand, cheap-per-unit means many low-stakes throwaway interactions, each generating tokens, none generating value. High volume was the signature of gaming, while high cost-per-unit was the signature of genuine complexity. So when Shopify sorted by the adjusted metric, the leaderboard inverted: the people who had looked most productive on raw spend rose to the bottom, and the developers doing the hardest work who were mid-table on volume, rose to the top.

But cost-per-unit is a better metric than volume, not a complete one. It exposes gaming but it can’t see the two failure modes that sit just underneath. The first is rework cost: AI-generated code that ships, passes the denominator test, and then causes a production incident absorbed by a teammate in the next sprint. That cost is real, but it appears elsewhere in the system, attributed to something else, invisible to any dashboard tracking the original developer’s output ratio. The second is the efficiency ceiling: even when a feature ships cleanly, the question remains whether it required the spend it consumed. A developer who spent $600 on a feature that could have been built with $200 of targeted prompting is neither gaming nor exemplary — they’re just not yet good at using the tool. The ratio captures output but not optimal output. It tells you whether value was produced; it doesn’t tell you whether value was produced at the right cost.

Shopify found the denominator everyone else had been ignoring. It was the right move, and it caught the worst behaviour. But it didn’t solve measurement, it got closer to it. If even the best-instrumented companies can’t yet measure AI productivity reliably, what does it mean to claim that your sector is “ahead” on AI? The answer requires a different map.

Why the standard Map is wrong

Every organisation trying to locate itself on the AI curve is using the same map. The map has one axis: adoption. Who’s ahead, who’s catching up, who’s been left behind. By that measure, the story is straightforward: Information technology, finance, and professional services are the early movers. Manufacturing and retail are accelerating. Construction and agriculture are laggards. The league table is updated quarterly by analysts, cited in board presentations, and used to calibrate urgency.

The problem is that the axis is wrong.

Adoption rate — the percentage of an organisation’s workforce using AI tools, or the number of functions where AI has been deployed — is a measurement of input. It tells you how much AI your organisation is consuming. It is, in this sense, the sector-level equivalent of the token leaderboard: a number that is visible, trackable, and only loosely connected to the thing you actually want to know.

The data makes this difficult to ignore: 88% of organisations used AI in at least one business function in 2025, yet only 7% said it was fully integrated across their organisation. That gap between “we use AI tools” and “AI has changed how work gets done” is the central fact of the current moment, and the adoption league table has no way to show it.

Technology readiness alone is insufficient for operational success: organisations need people readiness, process readiness, and data readiness alongside it. As we have seen, technology is the easiest of the four to buy and most C-suite hope that it can be a sort of a magic wand.

The reasons sectors lag on AI aren’t uniform, and solutions that work in one sector are structurally non-transferable to another: Finance’s readiness problem is different from healthcare’s is different from construction’s, not just in degree but in kind. A single axis that ranks them all by adoption speed doesn’t just oversimplify, it actively misleads, because it implies that the slower sectors need to do what the faster sectors did, only faster. Often they don’t, but the constraint is entirely different.

The distinction that matters is between adoption intensity and structural readiness. Adoption intensity is what dashboards measure: seat licences, active users, token consumption, functions touched. Structural readiness is something else: whether the organisation has redesigned its work around AI rather than simply layering AI on top of existing work, like whether roles have been redefined, whether workflows have been rebuilt from first principles rather than augmented at the edges or whether the data infrastructure exists to make AI outputs reliable rather than plausible. Last but not least, if governance frameworks have been developed to determine who builds what, who owns context, and who is accountable when something goes wrong.

The Transition Radar

The gap between these two dimensions is where most organisations currently live: an organisation can have very high adoption intensity and very low structural readiness, and it can have low adoption intensity and meaningful structural readiness. The two axes produce four distinct positions, each with a different risk profile and a different intervention requirement. That is the Industry Transition Radar.

The four quadrants that follow describe distinct positions on the radar, each with its own risk profile, its own failure mode, and its own intervention requirement.

Equipped — high adoption intensity, low structural readiness

The most populated quadrant in enterprise AI right now. Tools are deployed, usage dashboards are green, and leadership is receiving weekly reports on seat utilisation and active user counts, while underneath it the work design is largely unchanged. Workflows have been augmented at the edges, a Copilot here and an AI assistant there, but not rebuilt. Roles have been retitled but not redefined. Data infrastructure has not been addressed because it was not visible on the procurement order, and governance exists on paper, applied as a brake rather than as a structure.

The Equipped organisation is not failing. It is producing real individual productivity gains, and the 5x super-user effect is real and shows up consistently in surveys. What it is not doing is capturing those gains at the organisational level, because the value is concentrated in a small tier of power users and diffuses rapidly as you move toward the median employee. When the board asks what the AI investment has actually produced, the honest answer is individual efficiency gains that have not yet compounded into structural change, and that answer is becoming harder to defer.

Embedded — high adoption intensity, high structural readiness

The target state, and the rarest position in 2026. Work has been redesigned rather than augmented, roles reflect what AI actually does versus what humans need to do, data governance is operational rather than aspirational, and governance frameworks have real accountability structures where someone owns context, someone audits outputs, and someone decides when an agent has exceeded its remit.

Almost no sector has reached Embedded at scale. What exists are pockets: individual firms within professional services and finance that have redesigned their delivery models around AI rather than equipping their existing models with AI tools. Amazon’s pivot from KiroRank to normalised deployments, measuring code that actually ships rather than tokens consumed, is the first public signal of a major enterprise attempting the transition from Equipped toward Embedded. It is one metric change inside one workflow, and the distance from that to genuinely Embedded operations is significant, running through the three readiness factors that technology spending alone cannot buy.

Exposed — low adoption intensity, low structural readiness

Low on both axes. The Exposed organisation has neither deployed tools at scale nor prepared the organisational ground for them, which is not necessarily a crisis position given that the competitive window has not yet closed, but the gap is compounding. Industries with more exposed sectors are watching peers accumulate productivity advantages that will become structural over time, and the risk is not that the Exposed organisation is currently losing but that the cost of catching up rises every quarter, and that some of the structural work, particularly data architecture, takes years to build rather than months.

The important distinction within Exposed is between sectors that are exposed because they have not prioritised AI and sectors that are exposed because their structural conditions make standard AI adoption strategies non-viable. Construction is the clearest example of the latter: the industry’s fragmented, project-based model destroys data continuity at the end of every project, meaning individual firm investment in AI tools hits a structural ceiling that internal spending cannot resolve. For those sectors, the path out of Exposed does not begin with tools. It begins with industry-level coordination on data standards, a precondition that no single organisation can create alone.

For free subscribers, that’s it for this week, but tune in for our upcoming essay on the World Cup⚽ and people analytics,

For futurist subscribers read on for the 2nd part of this essay on the transition radar, where each sector fits, and what moves you on the radar. Watch this space for invitations to our 24th of June Work3 Community Brains Trust Discussion - What comes after the job?

Read more

A note from Andy and Matteo

If you’ve been reading Work3 for a while, you know we don’t do many housekeeping posts. We’d rather send you something worth thinking about than fill your inbox with announcements.

But this one is different. Because what we’re about to share isn’t really about us - it’s about what we want to build with you this year.

So here it is: the full picture of what Work3 will be for the rest of 2026.

Why now, and why this matters

Work3 started as a place to think clearly about work: not the LinkedIn-motivation version, not the techno-utopian version, but the real, complicated, sometimes uncomfortable version (with a touch of humor, and human perspective).

What happens to people when their organizations change faster than they can adapt?

What AI is actually doing to jobs, power, and meaning, not in five years, but now?

That hasn’t changed and if anything, we are trying to sharpen these questions.

We’ve spent the last few months honestly assessing where we are: 30,000 of you read Work3 every week. For us, this is both exciting and humbling.

What hasn’t kept pace is the structure around the content, publishing cadence and structure has sometimes been inconsistent and our paid and community offer has been vague. There’s been no clear signal of where Work3 is going next.

That’s what we want to change now.

What you can expect, and when

Starting in June, Work3 publishes on a more reliable schedule.

Here’s exactly what that looks like:

Three original essays per month. These are the core of what we do: long-form, researched, argument-driven pieces on the themes that matter most right now. They won’t be rushed, and they won’t be filler. Each one will have a clear aim and earn its place in your inbox. We will also be differentiating them further, with an editorial calendar and structure that we outline below.

Work3 Monthly Round-Up. A curated, opinionated look at what happened in work, AI, and organizational life that month. Not a link farm, our actual views on what it means and a way for you to be always up to speed on what's happening.

One monthly Work3 Community Brains Trust. A live conversation with members of the Work3 community. We get leaders from different industries and parts of the world chiming in on what we write about, trends, real case studies and experiences. This is the human part of our work, that we started in 2025 but will be running with an improved format, schedule, to make it not just useful, but also actionable.

Quarterly deep-dives. Every three months we’ll publish a substantial research report - 20 to 30 pages - on a single theme. The first one, launching in June, is a PDF Report on “Nine Global Mega Trends Shaping the Future of Work”.

We did this in 2025, but we are relaunching in 2026 with updated trends, developments and insights. If you’ve wanted a document you can actually use in a strategy conversation or share with a leadership team, this is it.

What we’re going to write about

Today, everyone has an opinion on the ‘future of work’.

However the term is about as meaningful as the ‘future of weather’ - your prospects very much depend on where you are.

We don’t aim to predict the future, we want to a build a better one.

We plan to share our research on how people are building a preferable future for specific areas, skills, and, industries.

Work3 is about the redesign of work as a system, not the inspirational layer on top, but the operating system underneath: how work gets organized, who holds power, what technology is actually doing to the firm, the job, the career, and the person doing the work.

This year we’re writing across five connected territories:

Work as a system. The unbundling of the firm, Agentic AI as a new operating model, What replaces the manager, the job description, and the org chart as the primary units of coordination.

Work as a human experience. Trust inside organizations. The emotional economy. What it feels like to work when the tools around you are changing faster than the culture.

Careers and skills. How to build something durable when ladders disappear, credentials erode, and the definition of “good at your job” shifts every eighteen months.

Work and society. The policy layer. The demographic layer. What breaks when work systems change faster than institutions can adapt. A little Toffler, a little Luddism, a little of what science fiction got right.

Industries in transition. How specific sectors are reshaping before the labour market catches up — because industry precedes work, and work precedes policy.

Each month will be anchored to one of these territories. You’ll always know what kind of conversation we’re in.

New membership plans: what’s changing

We’ve redesigned the paid offer from scratch, because the old one wasn’t clear about what we were giving you.

Here’s how it works now:

Free - You will continue to read one of the original essays per month, plus the new Work3 Monthly Round-Up. If you never pay us a penny to support us, we still want you to keep reading!

Futurist - $8/month or $80/year. This is for readers who want all of our content. You get all three original essays, the Work3 Monthly Round-Up, invitations to the monthly Work3 Brains Trust live discussion, and a Work3 community chat later this year. You will also get early access to two tools we’re building: the WorkForce3 dashboard (a live signal tracker of AI displacement and labour market shifts by sector) and the Work3 GPT, which lets you search and converse with everything we’ve ever published. Pay annually and you get a discount of wo months free. Plus the exclusive PDF Report on “Nine Global Mega Trends Shaping the Future of Work”, and future quarterly reports.

Founding Member - $250/year, limited to 100 people. This is for the people who want to shape what Work3 becomes with us. Everything in Futurist, plus access to 1:1 with Andy and Matteo, direct input on our editorial and product roadmap, and beta access to every new tool we build before it goes public.

If you’re currently a paid subscriber on the old plan, nothing changes for you automatically - we’ll be in touch personally. We appreciate so much your early support, that we will be moving you all to the founding member plan. Our goal for you is to live up to our promise and really pay back your initial support as we launched last year as everything wasn’t 100% clear.

If you’re ready, please use this button to subscribe/upgrade!

Subscribe now

One last thing we’re asking of you

We’re sending out a short survey to the full Work3 community shortly. It takes two minutes. It will help us understand who you are, what you want more of, and whether the new membership structure makes sense for you.

There’s also a free gift attached - if you complete the survey, you’ll get 2 months of membership on the Futurist paid plan.

We’ve been building this thing for a few years now, mostly by instinct. This year we want to build it with more information. You are, after all, the reason it exists.

Help us shape Work3

New here? Start with these reader favourites

If you’re still deciding whether Work3 is for you, here are a few of the essays that have shaped the conversation so far.

1. Death of Entry Level Jobs - Are we building a world of work with no way for beginners to begin?

2. The Last Stand of the Corporate Peacock🦚 - Why return-to-office mandates may be less about productivity - and more about power, status and old habits refusing to die.

3. The Decentralised Workforce - How freelancers, creators and one-person empires are reshaping work, wealth and the future of the firm.

4. The Global Mega Trends Shaping the Future of Work - A big-picture guide to the forces unbundling jobs, organisations, careers and the social contract.

5. Future of Jobs Report 2025 - What the World Economic Forum’s latest jobs forecast gets right - and where leaders should read it with caution.

6. AI-geism - Will AI help older workers stay relevant - or become another excuse to write them off?

If you've made it here - you're a champion of attention and a true fan! Again, thanks for your support, engagement and feedback so far. We are humbled and honored, and will be pushing to make Work3 more useful every day.

Any comments or questions then let us know.

Matteo and Andy

Subscribe now

NOTE FOR WORK3 ONLY

• This is vehicle for distributing PDF reports to paying subs.

• Set to Paid Only

• Link to this page on Welcome Emails

• Pin to home-page - FOMO for non-payers

Read more

Welcome to this week’s edition! We are continuing a series on Enterprise AI adoption in the workplace, this week looking at how we can move from AI experiments to AI Systems.

Let’s dive in!

When Agents leave the Demo

There is a strange moment that happens after a good AI demo: the agent works, it finds the right document, drafts the right answer, pulls together the right context, and does in a few seconds what would normally take someone twenty minutes of searching, copying, checking, and rewriting. Everyone in the room has the same reaction: this is useful, we should have more of this.

In 2026, most large organizations have moved past the question of whether they should experiment with AI agents. Some have dozens, others have hundreds. McKinsey has talked about running 20,000 AI agents internally. Agents exist across HR, sales, customer support, finance, IT, and knowledge management. Some are built centrally by technical teams, but many emerge from the edges of the business, created by people trying to solve the specific, visible friction in their daily work.

This is genuinely a good thing: for years, enterprise software has been something done to employees rather than with them: bought centrally, implemented slowly, and pushed into workflows that didn’t always match how people actually worked. Agents change some of that energy when the people closest to the work can now shape the tools around it.

But once the agent leaves the controlled environment of the demo, it enters the much messier reality of the company it’s supposed to serve, and that’s where things tend to get complicated.

Enter the agent graveyard: every organization has a graveyard of internal tools. Dashboards nobody opens anymore, some automations built by someone who has since moved teams, knowledge bases that were useful for six months, then slowly drifted out of date as the processes they documented evolved. Some are workflows that technically still run, even though nobody can quite remember why they were created or whether they still reflect how the work is done. Most of these failures are not dramatic - the tools didn’t announce themselves as failures. They were useful once, then gradually underused, forgotten, or broken. The more common pattern is quieter: the tool keeps existing, but nobody is quite sure whether it still reflects how the business actually works.

I suspect many agents will follow exactly the same path. They won’t stop working in any obvious way. The underlying business process will shift and the documents they rely on will be updated, duplicated, or contradicted somewhere else in the organization. Six months later, someone will ask whether the agent is still accurate, and the honest answer will be that nobody has been watching closely enough to know.

This is what AI sprawl (uncontrolled proliferation of AI models, tools, and agents across an organization, creating security, governance, and efficiency risks) actually looks like in practice: a gradual accumulation of promising tools that never quite became reliable systems. It’s a pattern (or vicious circle) enterprises have lived through before. Useful tools appear at the edge of the business, solve real problems, spread faster than the governance structures around them, and eventually become another layer of organizational complexity to manage. Shadow IT wasn’t born from bad intentions, it came from good intentions meeting the wrong infrastructure, and agents risk following the same trajectory, except with considerably more surface area.

The CIO’s problem is not “can we build agents?” It is “can we have hundreds or thousands of agents operating across the enterprise without creating unmanaged risk, duplication, shadow AI and another layer of digital clutter?”

Five ways an agent fails

The failure usually doesn’t happen at the model level, it happens at the five levels underneath, and in my experience they tend to chain rather than appearing separately: an agent built with poor context gets used less, which means nobody notices the problems accumulating, which means nobody owns fixing them, which means governance gaps go unaddressed, which means when someone finally asks whether the program is working there is nothing to measure it against. Each failure makes the next one harder to catch.

1. The first is context. Most agents are built and tested against a simplified version of the business. In the demo, the documents are clean, the use case is narrow, and the permissions work. In production, the agent meets the actual company: siloed data, inconsistent metadata, outdated policies, undocumented processes, and knowledge that lives partly in systems and partly in the memory of whoever has been around longest. Without a shared context architecture, every new agent starts from zero, with the same data sources getting reconnected, the same permissions getting negotiated, and the same business logic getting reconstructed by a different builder in a different team.

2. The second is adoption. Many agent rollouts are still treated as a distribution problem: announce the tool, show the demo, share the link. The pattern I see most often is this: the agent gets announced in a team meeting, four or five people try it in the first week, usage drifts back to two by week three, and by month two it’s one person occasionally checking whether it still works. The announcement created awareness but it didn’t create habit. Without structured enablement, feedback loops, and visible integration into the workflows people already use, most agents end up bypassed by everyone except the people who built them.

3. The third is ownership. The builder moves on, the workflow changes, the underlying source gets updated, and the agent keeps running, still available, still answering, but no longer fully aligned with the process or the data it was built around. A particularly common version of this: the organization goes through a reorganization, a process changes, and the agent keeps confidently answering questions about the old workflow because nobody updated its knowledge sources. Nobody notices immediately because the people who knew the original process have moved on too.

4. The fourth is governance. Even when IT and security are involved, they’re often trying to govern an expanding surface area with limited visibility. If every team makes its own decisions about permissions, access, and risk, governance becomes inconsistent at exactly the moment the organization needs more consistency, with no org-wide guardrails and every team making its own security decisions at speed and in isolation.

5. The fifth is measurement, and this is the one I find hardest to explain to people because it seems so obviously fixable and yet it almost never gets fixed in time. Agents get launched without a clear definition of what good looks like, and a few months later, when someone asks whether the program is working, there’s no baseline, no measurement framework, and no business case to interrogate. The agent either feels useful or it doesn’t, and that’s not a sufficient basis for deciding whether to maintain it, improve it, or scale it.

What makes these five problems particularly hard to solve is that they don’t belong to the same person: the builder owns context and adoption, nobody clearly owns maintenance. IT owns governance, when asked. The CIO owns measurement, when someone thinks to involve them early enough. Each group does their job reasonably well within their own lane, but none of them do it in sequence with the others. This is less a collection of operational failures and more a single coordination failure that shows up in five different places.

Not that this is a new kind of scenario: Software has had a development lifecycle for decades: define the problem, design the solution, build it, test it, ship it with governance, monitor it in production, and improve it over time. You don’t deploy code into a business-critical environment and then walk away. But agents feel lighter than traditional software: the interface is conversational, the build is fast, and the use case seems obvious, so organizations sometimes treat them more like experiments than production systems. That informality is useful in the early stages of discovery and becomes a problem when agents begin to touch real workflows, real data, real users, and real decisions. Enterprise software has tried to solve versions of this before. Agile addressed fast iteration. DevOps connected development and operations. MLOps emerged to govern model training and monitoring. Security frameworks cover access controls and risk management. But none of these were designed for the specific challenge agents create: dozens or hundreds of them being built by non-technical business teams, operating inside complex permissioned environments, touching real data and decisions. Agents are too easy to build to warrant major software project treatment, but too consequential to be treated as casual experiments. They currently slip through all of these frameworks simultaneously.

The question then is not whether organizations should build agents more carefully, but where that care actually has to start, and the answer is usually much earlier than anyone involved in the build expects.

Running the sequence backwards

The order most enterprises follow: build something, launch it, and figure out whether it worked. After all, you need something to exist before you can govern it or measure it, so the natural instinct is to build first and sort out the rest afterwards.

The problem is that this ordering, however logical it feels in the moment, is also why so many agent programs end up with tools nobody can honestly evaluate. By the time the CIO asks whether the agent is delivering value, the builder has moved on and IT is still finding out the agent existed.

The order that works is the reverse: Measure, Govern, Build. Before the builder writes a single prompt, the CIO’s office agrees on what success looks like and how it will be tracked, and that baseline becomes the brief. Before the agent launches, IT and security define the access policies, the system boundaries, and the failure protocols, not as a review gate at the end but as constraints that shape the design from the beginning. Then the builder starts, with a clear problem definition, guardrails to work within, and a measurement framework to build toward.

What I find interesting about this sequence is that it doesn’t actually require new tools or new roles to implement, it requires the most senior stakeholders to do their work before there’s a polished demo to react to, and that’s a much harder cultural shift than any technical one. The CIO naturally wants to see the agent before committing to a definition of success, IT naturally waits to be asked. The builder, who has the clearest mandate and the most momentum, fills the vacuum that this creates, which is how organizations end up in the situation they were already trying to avoid.

Reversing the sequence is fundamentally a question of discipline rather than technology, and it requires a shared operating model that works across teams, vendors, and business functions, much like a software development cycle.

Context as infrastructure

There’s a reason we have started talking about context engineering rather than prompt engineering: the prompt is a relatively small part of what determines whether an agent works in production. The larger part is everything around it: what company data the agent can see, what it knows about the process it’s operating inside, what it remembers from prior runs, and what it’s blocked from touching. Getting that right is an information design problem, not a language one.

Good context engineering has two qualities: it’s dynamic, meaning the agent fetches what it needs automatically rather than relying on a human to pass it every time. And it’s multi-player, meaning it’s designed for the whole organization rather than one person’s carefully configured setup.

Most enterprise AI efforts struggle with both. What passes for context engineering in most teams is one builder spending a weekend connecting data sources, writing system prompts that encode their own understanding of the business, and configuring an environment that works well for them and nobody else. When that builder moves on, the context architecture they built tends to go with them, because it was never documented or shared in a way that others could maintain or build on.

Making context engineering genuinely multi-player requires agreement across teams on what the agent should know, what it should have access to, and what the standards are for how it gets built. That agreement doesn’t happen at the builder level. It requires the same coordination across the three roles described earlier, which is why the sequence matters as much as the tooling.

The Enterprise Agent Development Lifecycle

This is the problem Glean, whose Enterprise Graph we covered earlier in this series, has formalized into what they call the Enterprise Agent Development Lifecycle. I found it interesting not because it invents lifecycle management, but because it applies that discipline to the specific problem AI agents create: fast, distributed creation inside complex, permissioned, constantly changing organisations. This also matters because it provides the shared knowledge infrastructure foundation, and this is the operating model that runs on top of it and turns agent building into something repeatable.

The framework spans seven stages.

1. Opportunity – Start by spelling out the business problem the agent is intended to solve. Use plain language: who is affected today, how work happens without the agent, and what tangible change you expect if the agent succeeds. This anchors everything that follows to outcomes the business already cares about, not just an interesting demo.

2. Design – Describe what the agent is actually responsible for. Define the unit of work (a ticket, an incident, a call, a workflow), when it should run, what information it needs, and what it should produce every time. Be explicit about what’s in scope, what’s out of scope, and which assumptions need to be tested early.

3. Performance – Turn that intent into a small, concrete set of success metrics: the business KPIs you expect to move, plus agent‑centric quality and safety signals. Agree on baselines and target ranges, and decide up front what would cause you to expand, pause, or roll back the agent.

4. Context – Identify the minimum set of permission‑aware data sources, tools, examples, and feedback signals the agent needs to do its job and to be evaluated fairly. This includes which systems it can read from, which actions it is allowed to take, and what telemetry you will use to understand how it’s performing.

5. Develop – Turn the design into a reliable agent. Choose the right execution model (for example, a structured workflow versus more flexible auto‑mode), then test against golden examples, run it in parallel with the existing process, and pilot with design partners until the agent behaves predictably, not just on a hand‑picked demo set.

6. Launch – Treat rollout as a change‑management exercise, not a switch flip. Decide who will see the agent first, how it will show up in their workflow, what training and communication they need, and which guardrails, SLOs, and kill switches must be in place before you broaden access.

7. Monitor & Improve – Operate the agent like any other critical system. Use dashboards, alerts, and runbooks to track impact and quality over time, handle incidents, and feed real‑world signals like user feedback, overrides, drift in key metrics, back into earlier stages of the lifecycle so the agent keeps improving.

The practical question for most people reading this is how to actually introduce this sequence in an organization where the builder already has momentum and the CIO hasn’t been asked. The honest answer is that you rarely reverse the sequence across the whole organization at once. What tends to work is picking one upcoming agent build and treating it as a pilot for the new process: asking the business owner to define the success metric before the configuration starts, getting IT in the room during the design rather than the review, and documenting the context architecture in a way that survives the builder moving on. One agent built this way is more persuasive than any argument about methodology, because it gives you something concrete to point to when the next build starts and someone asks why the process is different this time.

What makes this more than a framework is that each stage requires concrete tooling behind it. The platform must be designed to lower the barrier for non-technical teams without sacrificing the context quality that makes agents useful in production. Builders should be able to describe what they want an agent to do in plain language and have the platform handle execution across company data without predefined workflows. The key differentiator for any platform is its foundation: a robust, permission-aware enterprise graph ensures the agent isn’t searching a curated subset of documents but is accessing a continuously updated map of how the company actually works across all connected systems. When something fails, the platform must offer tools that surface what happened at every step: inputs, tool calls, model decisions, and output, giving teams a basis for diagnosis rather than guesswork. For instance, Glean provides features like Debug and Trace Views to achieve this visibility. Furthermore, sophisticated architectures should support coordination at runtime, allowing agents to react to enterprise events automatically rather than waiting to be manually invoked.

The launch stage is built around the premise that governance designed after deployment isn’t really governance so much as an attempt to retrofit control onto something that was already built without it. An Agent Library gives IT teams the controls to verify agents, organize them by department, and manage their lifecycle, turning informal distribution into something the organization can actually see and track. Agent Access Policies sit above that, applying org-wide guardrails consistently regardless of who built the agent or which team owns it, so that the same controls that apply to the tenth agent apply to the thousandth.

The Monitor & Improve stage closes the loop that most agent programs leave open. An Agent Insights dashboard tracks adoption rates, top use cases, estimated hours saved, and ROI trends over time. The business value isn’t just tracking a single agent’s performance but being able to compare impact across the portfolio, build a defensible case for continued investment, and identify which use cases deserve more development and which should be retired. That’s what turns monitoring from a reporting exercise into an operational decision.

The next divide

So far, the conversation about enterprise AI has been dominated, understandably, by what agents can do.

But questions shift quickly: early on, the question is whether the agent can work at all. Then it’s whether it works reliably. Then it’s whether anyone is actually using it. Then it’s whether it’s still accurate six months later. Then it’s whether it can be trusted with something more consequential. Each of those transitions requires a different kind of organizational response, and most enterprises are still working through the first few while the more mature programs are already navigating the later ones.

The organizations that handle those transitions well probably won’t be distinguished by how many agents they’ve created, they’ll be the ones where someone, at some point, decided to run one build properly before scaling the approach. That’s usually how the sequence gets reversed in practice: not through a top-down mandate but through one team demonstrating that the slower, more disciplined approach produces something the organization can actually rely on.

The next divide will not be between companies that use AI and companies that do not, it will be between organizations that can turn AI experiments into trusted systems, and those that simply accumulate another layer of digital clutter.

TL;DR - We’ve built an economy where ownership drives wealth - but a political system still optimised around work. Policy needs to shift from supporting earners to enabling ownership, especially for younger generations.

In Part 1, I argued that the defining divide in today’s economy is no longer class, but generation - between those who earn and those who own.

Work no longer guarantees financial security. Ownership increasingly does.

This second part asks the next question:

What would it take to rebuild access to ownership?

You can read Part 1 of the essay below.

Lower Barriers to Ownership

So where does that leave younger cohorts of workers with ambitions to live just as fully as their parents’ generation?

The cost of living is increasing, there is AI hype and anxiety in the workplace, even those with jobs feel financially insecure.

No wonder in recent years we have seen greater worker demands for

• Autonomy – aka who wants to work for a corporate peacock?

• Flexibility – enforcing rigid work patterns feels very last century.

• Equity – “I might not want a 10% cut of every deal, but at least some recognition.”

Maybe an aspiration for most is to achieve two out of three of this trinity. These aren’t just workplace preferences. They are signals that people want a stake - not just a salary.

At Work 3 we have described how people are increasingly working in non-traditional forms of employment, with the emergence of The Decentralised Workforce and the creator economy - freelancers, solopreneurs, digital nomads, and playbourers.

One way to see the interest in Web 3 wasn’t a fascination for slow databases, but an opportunity to collaborate together on projects with some skin in the game, i.e. equity.

Few are building million-dollar one-person businesses - yet.

Building Preferable Futures for the Next Generation

The 20th-century social contract assumed that work would provide security.

That assumption no longer holds.

Unfair social contracts don’t change overnight, with or without democratic mechanisms.

We don’t make glib predictions based on ChatGPT task extrapolations, but we do expect volatility and insecurity, as jobs fragment.

The 20^th Century Social Contract is clearly broken,

the question is no longer how we redistribute income - but how we distribute ownership.

There are multiple fault-lines in society, Owners Vs Earners, Young Vs Old is just one of them.

Generational divides are one of several fault-lines in society, not a master explanation.

Another divide is from those who can ‘work anywhere’, and those who must ‘work somewhere’ - remote knowledge workers vs location-bound service workers is also politically charged.

We are still organising politics around the job - even as the job itself is fragmenting.

However, millions now work as freelancers, gig workers, creators or part-time carers. AI agents and task platforms are unbundling what we used to call a ‘job’; and that invisible labour - care, emotional support, community work - still goes unrecognised.

We have postulated how we might Build a Better World Without Jobs.

Most of our heritage political parties, and policies, are impotently barking for the wrong goals. Full employment and great quality jobs might have been feasible in the last century, but not today. Our political platforms still frame policy through the lens of job creation, job guarantees, job security, and job protectionism. If the job is fragmenting, then organising politics around either working-class solidarity, young earner grievances, or protecting elders investment streams may be missing the structural shift.

Throughout history your prospects have always been partly determined by when and where you were born, and who to. For many this has become more entrenched, with less mobility than previous generations.

AI is accelerating the generational problem specifically at the entry level for the newest members of the workforce, the newly graduated. Work3 has highlighted what it calls the “death of entry-level jobs”

A generation facing a ‘hopeless’ job search, applying for hundreds of positions with little response, while confidence among Gen Z tumbles and student debt rises.

AI is accelerating a system that was already breaking.

Gen Z is rewriting the employment contract from below - 44% of Generation Z have turned down an employer based on personal ethics or beliefs. One survey found that 75% of Generation Z have no plans to go full-time in their whole lives.

This is politically significant: it suggests the generational divide isn’t only something happening to young people (precarity, housing lock-out), but also something young people are actively choosing - rejecting the Boomer social contract of stable employment in exchange for security.

That makes the politics more complicated: it’s not simply young people as victims of a system designed by the old, but a generational divergence in what work is for.

What Can We Do?

Most policy responses today aim to rebalance the system. What’s needed is redesign.

Around the World there have been some policies to attempt to fix the intergenerational fairness problem. You can read more here…

In general the policies lean toward rebalancing, not redesign, for example, investing in education and lifelong learning and to encourage fair taxation and social protection.

Denmark has gone furthest in building intergenerational fairness into the pension system structurally. Retirement ages are indexed to longevity developments.

Norway’s sovereign wealth fund — saving windfalls for future generations - as of mid-2025, it held over $1.9 trillion in assets, equivalent to over $340,000 per Norwegian citizen.

Taxing housing wealth rather than earned income has been explored by several countries, including New Zealand. This would shift the tax burden from labour (which young earners pay heavily) to property (which older asset-owners hold).

We need to redesign access to ownership.

We need to rethink work.

We need to build systems that deliver financial security beyond employment:

• Build solidarity between generations

• Improve access to housing

• Expand participation in ownership

For Employers - move beyond paying for time towards sharing in outcomes. Mastercard has built an internal talent marketplace that matches employees to projects, mentoring and opportunities across the organisation, giving them exposure and growth beyond their formal role.

For Individuals – increasingly, security comes from owning even small pieces of value - whether that’s equity in a company, a portfolio of assets, or a side venture built on platforms like Shopify or Substack.

From Earners to Owners

For most of the 20th century, work was the path to prosperity.

In the 21^st Century, ownership is.

Prosperity has become inversely correlated to age.

• Owners (assets, pensions, property) are structurally advantaged over earners (labour income, early-career workers).

• The ‘young vs old divide’ makes more sense now that the job is fragmenting and asset ownership is what separates security from precarity.

• The ‘working vs middle class divide’ made more sense when the job was stable and unionised, but not now.

Heritage political parties are not redesigning a fairer system, they are just protecting historical incumbents and vested interests.

Work is being deconstructed. Assets are compounding faster than wages. Politics reflects voters, not workers.

Younger generations are Less wealthy, Less secure, Less represented

If the 20th century built a middle class through work, the 21st will be defined by whether we can rebuild it through ownership.

We’ve seen early signs of how this might evolve, from new forms of work and equity to emerging policy debates around assets, taxation and access. The direction is clear, even if the model is not. This is a theme we will continue to explore.

To end, a question for all of us,

How do we distribute wealth and security in a world where paid employment is no longer a reliable way to do it?

This is the second part of a two-part essay.

The first part explores how we arrived at this generational divide — and why work no longer guarantees security - you can read the full article here →

From Class War to Generational Divide (Part 1)

Further Reading

• Intergenerational Fairness - Recommendations in report from European Commission

• Intergenerational Fairness - Resolution Foundation

• Select Committee on Intergenerational Fairness and Provision
  Tackling intergenerational unfairness - UK Parliament

• Inequality according to Thomas Piketty, in 10 graphs

• The Bank of Mum and Dad - BBC

• What drives poverty and inequality, and how can governments tackle them? Joseph Rowntree Foundation.

Learning, earning - and owning,

Andy

TL;DR - We’ve built an economy where ownership drives wealth - but a political system still optimised around work. Policy needs to shift from supporting earners to enabling ownership, especially for younger generations.

In the last century, the JOB provided most people in industrialised countries with food and housing for their families.

Our societies were funded from taxes from companies and workers earnings.

Our politics galvanised around workers’ interests, and owners’ interests.

Today this model is broken.

Most of the poor in the UK and US actually have jobs.

Getting on the housing ladder has become nearly impossible for thirtysomethings.

Financial security has become disconnected from employment.

Millennials are the first generation to earn less than the preceding one.

We have built an economy where ownership drives wealth - but a political system still optimised around work.

This essay - in two parts, argues that we need to reframe public policy, and work, through a different lens, that between the young (the earners) and the old (the owners).

Work Don’t Work

Work provides many meaningful benefits from a sense of purpose to social cohesion and importantly financial security for our families. Well-paid work is critical for ensuring people can afford the essentials and have a good standard of living. However, almost 7 in 10 working-age adults in poverty are in a household where at least one adult is in work. (Joseph Rowntree Foundation). Over 40% of Americans below the poverty line live in households with at least one worker.

Even people who are in employment often cannot rely on work to lift them out of poverty.

And looking more globally, around six in ten workers don’t even have a formal job.

That is two billion people.

The idea that employment alone provides financial security is becoming harder to sustain.

For some, work doesn’t work anymore, for many, it never did in the first place.

Your Generation

The connection between work and financial security is fractured, and is a factor between some of the visceral divisions in our societies. One that is overlooked is the divide between generations.

The older have always been richer than the youngers for the obvious reason that they have been on the planet longer to acquire wealth and assets.

There are societal fairness adjustments between generations such as inheritance, but I would argue that the contract in our society between the older and younger has crumbled. Increasingly there are much more favourable policies for elders than youngers and this is playing out in the world of work.

Older generations benefitted from cheap housing, final-salary pensions, and free University education.

Younger generations face unaffordable property, defined contribution pensions and student debt.

As Scott Galloway puts it,

This divide between older and younger is not just found in the US.

Intergenerational wealth inequality has deepened significantly since the financial crisis; between 2006-08 and 2018-20, median wealth for Britons in their 60s rose by 55% in real terms, while wealth for those in their 30s fell by 34%.

An EU report on the “Intergenerational Contract”, noted that,

r > g, capital > labour, owners > earners

The idea that society divides between those who work and those who own is not new.

It’s always been better to own the factory, than to work in it.

Economists from Ricardo to Marx framed it as labour versus capital.

The 20th century temporarily softened that divide, as rising wages, home ownership and pensions created a broad middle class. But the underlying tension never disappeared.

As Tomas Piketty showed, when returns on capital outpace growth, wealth concentrates.

Owners of capital see their incomes grow faster than ordinary laborers.

r >g OR returns on capital exceed economic growth

My translation of this would be

If you own assets → you compound wealth

If you earn income → you fall behind

You can see the trend in the main graph, as a source of income, the trend is clear

Capital > Labour

The divide here is between earners versus owners,

Which tends to be younger generations vs older.

The Silver Voting Wave

We have an economic situation, and policies, that suits owners more than earners, the old more than the young, and vested interests mean that people will vote with their feet.

In the UK for example this tension plays out on protecting national expenditure on state pensions and health services, offering tax shelters for investments, which directly benefit the older, versus spending on education and welfare for younger working-age adults.

Historically, people might have voted by long-standing political class voting patterns but in 2026, whether you are an owner or an earner can mostly be predicted by your age.

In recent elections across the UK and US, age has become one of the strongest predictors of voting behaviour - arguably stronger than occupational class. In the 2019 UK general election, the crossover point was striking: voters under 39 favoured Labour, those above favoured the Conservatives. This wasn’t simply reducible to class.

A rigorous study of Swiss referendums across four decades provides unusually clean causal evidence. As voters of the same cohort age, they become less politically liberal and less supportive of policies that protect the environment, support young workers and families, or redistribute wealth.

Explicitly age-based political parties exist, from the Grey Panthers in Germany to 50PLUS in the Netherlands, and Gil Pensioners Party in Israel, but they remain marginal. Age is not a political identity in the way class once was. Yet modern politics increasingly behaves as if it is.

This provides hard evidence for the “gerontocracy” concern: that in ageing democracies, older voters structurally block policies that the young would prefer.

The Bank of Mum and Dad

As baby boomers hold an increasing share of wealth, younger generations are finding it harder to build financial security, often relying on family support for major milestones like homeownership.

Under 30s hold just 2% of UK property equity, the over 60s now own 55%. (Savills).

More younger adults live at home with their parents until later than in previous cohorts, and it has been estimated that a worker renting in London saving their disposable income would take 30 years to save a typical deposit for a house. (ABC Finance)

An argument against ‘age vs class’ as a dominant political narrative would argue that if we looked more deeply at age effects we would see class underlying it. Over half of those with university-educated homeowning parents received transfers when buying for the first time, with receivers getting around £35,000 on average. This compares to 29% of those with renting parents receiving transfers, with an average transfer of £11,000. (IFS)

This is Part 1 of a two-part essay. You can read Part 2 below, “Beyond the Generational Divide”

Learning, earning - and owning,

Andy

Around one in six people in the world now use generative AI tools.

AI promises transformation.

In reality, many of us are using it to rewrite emails, overthink text messages, and outsource uncomfortable conversations.

Inside organisations, a new metric is emerging: AI adoption.

Dashboards measure prompts per employee.

Readiness scores appear.

Leaders ask whether teams are “using AI enough”.

This misses the point.

AI adoption is an activity metric.
It tells us people are interacting with a tool.

It does not tell us whether anything meaningful is changing.

And confusing activity with progress is a familiar mistake.

Economists have a name for it.

Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure.

If employees know their AI usage is being tracked, the incentive is simple:
use more AI — not use AI better.

A recent viral chart shows where we are in the AI adoption cycle.
Each dot represents millions of people.

The story is simple: most of the world has still never interacted with ChatGPT or Claude.

It looks like the early stages of a global technology shift.

But the obvious question is: so what?

Another statistic tells a very different story.

Learning poverty, the share of children who cannot read and understand a simple text by age 10, sits at around 70% in low and middle-income countries.

Hundreds of millions of children are entering an AI-shaped world without the foundational capability required to use it.

If you cannot read, you cannot meaningfully use AI.

AI usage measures exposure to a tool.
Literacy measures the ability to benefit from it.

AI adoption is a growth metric for technology companies, not for society — and not for employers.

Firms Are Enforcing AI Adoption

Inside companies, the pressure to adopt is intensifying.

Boards expect progress.

Vendors push urgency.

Competitors announce initiatives.

Leaders worry about falling behind.

So organisations start measuring usage.

🎡 Prompts per employee.
🎡 Tokens consumed.
🎡 Adoption dashboards.

The logic is implicit: more usage signals progress.

It doesn’t.

Research is starting to show exactly why.

In one study with Boston Consulting Group, consultants using AI completed tasks faster and produced higher-quality work — until they crossed a boundary.

Beyond that point, performance actually declined.

The researchers called this the “jagged technological frontier” — the invisible line where AI shifts from improving work to quietly degrading it.

Most organisations do not know where that line is.

And they are not measuring it.

There is a second risk.

When people use AI, they tend to trust it — even when they shouldn’t.

Research on automation bias shows that individuals often defer to AI outputs, even when those outputs contradict evidence or their own judgement.

The problem is not just that AI can accelerate bad processes.

It is that AI-assisted decisions can be worse than unaided ones — because people stop checking.

Adoption metrics capture none of this.

Adoption Is Not Transformation

PwC’s global chairman Mohamed Kande recently noted that more than half of companies are seeing little value from AI investments — largely because foundational groundwork is missing.

Organisations are scaling usage before they understand where value actually comes from.

As a shareholder, I don’t care how many times employees used Google, Excel or ChatGPT this week. I care about performance - and ultimately, profits.

None of this underestimates the power of AI.

Productivity gains are real. Experimentation matters. Learning requires deployment.

But history suggests something important: adoption alone does not create transformation.

The question is not:

“How do we implement AI faster?”

Better questions are:

• What are we trying to achieve?

• What capabilities do we need?

• How are customer expectations changing?

Organisations repeatedly confuse deployment with change.

We Have Seen This Before

In HR, we once had paper application forms pages of questions filled in with a black biro.

When the internet arrived, we put the same forms online.
When applications came in, we printed them out and sent them to hiring managers.

With mobile arrived, we shortened the forms - but kept the process.

We used new technology to make old processes faster.

We did not redesign the system.

This is happening again with AI.

Implementing chatbots without redesigning service ends in tears.

Productivity Is Not Transformation

There is a useful distinction here.

Some changes improve existing work.
Others redefine it.

The first is productivity.
The second is transformation.

Most organisations are doing the first — and calling it the second.

History suggests this takes time.

When electricity first entered factories, many simply added it to existing layouts and saw modest gains.

The real transformation came later — when organisations redesigned how work was done around the new technology.

With AI, we are still at the wiring stage - not the redesign.

Doing Smart Things Slower

So the real risk is not that organisations adopt AI too slowly.

It is that they adopt it quickly in the wrong places.

Accelerating processes that should not exist.
Making decisions with less scrutiny.
Measuring activity instead of outcomes.

Doing the wrong things faster is one of the most reliable ways to create the illusion of progress.

So what does a better approach look like?

• Experiment — but don’t turn usage into a performance metric

• Invest in the few people willing to redesign how work happens

• Question the purpose of a process before automating it

• Redesign workflows, roles, and decisions before selecting tools

Adoption should follow clarity.

It is not a substitute for it.

Change What Matters

AI will reshape industries, organisations, and learning in meaningful ways.

But the constraint is unlikely to be access to tools.

It will be capability, judgement, and the willingness to redesign work.

We are already seeing the pattern.

We celebrate adoption before capability.
We scale tools before building foundations.

The organisations most likely to succeed with AI are the ones least focused on measuring its adoption.

That is not a comfortable thought when the board wants a dashboard.

The risk is not that organisations fail to adopt AI.

The risk is that they succeed - and nothing that matters actually changes.

Postscript June 2026 → we have taken this idea to a deeper level in our June article for Futurist subscribers

Drafting slowly,

Andy

Today we’re hosting a very interesting guest post from Dr James Ransom, Honorary Senior Research Fellow at UCL Institute of Education. He brings a thorough, analytical lens on how AI is impacting (or will) jobs by looking at a lot of research and stitching it together with lessons learnt from history. You can catch more of his work at Reskilled.

Enjoy!

What is a chain?

As AI advances, it will stitch together multiple tasks, reshaping the role of humans in jobs. The problem: if you leave the loop, you won’t get back in again.

You are an advertising specialist, and you’ve been called into your boss’s office.

“Right, I need you to own the full campaign lifecycle for the SmartBrew Pro launch – starting Monday, you’ve got two weeks to vacuum up every scrap of consumer data we can lay our hands on, then a week to interpret it all and tell me where the connected kettle market is heading, then about two weeks max where you’ll need to simultaneously advise Sarah’s team on strategy and build out the campaign plan, and finally three weeks to get scripts, copy, storyboards and media bookings nailed down with the agency. I think that’s it, I gotta join a call, keep me posted.”

What you probably won’t think about (especially as you’re going to be pretty busy for the next eight weeks) is that this is a chain of work. You have five tasks to complete, and they build on each other, albeit with some overlap in the middle.

You might consider how AI can speed this up, especially as your boss is pretty bad at spotting AI-generated output. Perhaps ChatGPT can sketch out a campaign plan. If you remember to feed it the right documents, Copilot should be able to help schedule some meetings, draft agendas and summarise next steps.

Most discussion of LLM tools and agents takes a similar approach. Think tank reports look at how AI is replacing or augmenting tasks, the building blocks of jobs. Academic work helps us understand the dynamics of what happens when you automate tasks, but the approach is usually atomistic: the relationship between the tasks is missing.

In real work, tasks are not separate or discrete. They run into each other, overlapping in a messy flow of projects and activities and responsibilities. Why would this be any different for the LLMs of the future? What happens if you skip all the intermediate steps and just request the final deliverable – “produce a complete advertising campaign for the SmartBrew Pro”?

The atomistic fallacy

Most serious attempts to understand AI’s impact on work share a common approach: break jobs into tasks, assess which tasks AI can handle or support, and add up the results. This is sensible, and has produced useful insights. Frey & Osborne’s landmark 2013 study estimated the probability of computerisation for 702 occupations. More recent work by Eloundou and colleagues mapped GPT-4’s capabilities against task descriptions in the US occupational database. Felten and colleagues built exposure scores based on AI capabilities mapped by the Electronic Frontier Foundation; similar approaches have been adopted by organisations from the IMF and ILO. (I’ve also created my own measures for India and the UK). In each case, the logic is the same: examine tasks one-by-one, score them, aggregate at the occupation level.

The problem is that real work doesn’t present itself as a neat list of separable tasks. Our advertising specialist isn’t completing task (a), filing it away, and moving cleanly on to task (b). She is running several in parallel, feeding the output of one into the next, making judgment calls about when something is good enough to build on. The tasks form a chain, and the connections between them matter as much as the tasks themselves.

Recent academic work has started to recognise this. Autor & Thompson argue that occupations ‘bundle a range of tasks of different expertise levels’, and that the overall composition of those bundles shapes what automation does to a role. Gans & Goldfarb show that automating tasks in bundles is often more economically rational than doing so one at a time. ‘A method that classifies tasks one-by-one can miss situations where automation arrives as a bundle – for example, when a platform or integrated system automates multiple steps at once’, they conclude.

But even bundling doesn’t quite capture what’s coming. A bundle is a cluster, whereas a chain is a sequence. Once AI can reliably execute a sequence of dependent tasks – where each step builds on the last – something qualitatively different happens. Our advertising specialist isn’t just removed from one task: she’s removed from the entire process.

To see what’s coming, we need to go back in time.

A lesson from the assembly line

The assembly line is given the credit for the explosion of productivity at Henry Ford’s Highland Park factory in 1913. But, as David Hounshell describes in his landmark study From the American System to Mass Production, 1800-1932, there were major bottlenecks to overcome before this – above all, reliability.

For most of the nineteenth century, ‘interchangeable parts’ were an aspiration: at the Singer sewing machine factory, as late as the 1880s, components were still being hand-fitted together by skilled workers who filed, adjusted and coaxed each piece into place. Every item needed a fitter. The US armouries at Springfield and Harpers Ferry had cracked interchangeability for muskets, but at enormous unit cost, and only at a scale of thousands, rather than hundreds of thousands or millions. Singer simply couldn’t afford that level of precision for a consumer product.

Ford solved the reliability problem by investing heavily in precision machining until parts were fully interchangeable, no fitting required. Only then did the assembly line make sense. Once reliability was sorted, speed followed: after April 1913, chassis assembly time fell from over 12 hours to 93 minutes. As Ford himself put it, ‘in mass production there are no fitters’.

Today’s LLMs have yet to overcome their reliability issues. Their outputs require a fitter – a human who checks, adjusts, corrects and decides whether the result is good enough to use. We are at the Singer stage: producing impressive work, but with someone leaning over every piece. As reliability improves, the pressure to chain tasks together without human intervention will become economically irresistible. What does that look like?

When the machines start talking

In January 2025, a video from an ElevenLabs hackathon went viral. Two AI voice agents, set up to negotiate a hotel booking with each other, quickly abandoned English and switched to a stream of rapid, garbled noise – a machine-to-machine protocol that was far more efficient than human language. The project was called Gibberlink, and some comments on the video suggest it unsettled people. “This is actually terrifying”, said one. “That’s… actually kinda scary”, another agreed. “This will be the last sound many of us hear”, said a third.

When AI agents operate in a chain, they need to pass information between steps. They can do this in English (or your human language of choice) – and for now, mostly they do – because the task is handed off to a human. But natural language is wildly inefficient for machine-to-machine communication. It’s verbose, ambiguous, and expensive: you’re paying for tokens on both ends, to generate and to comprehend. A structured format like JSON, or something more compressed still, strips out the redundancy and lets the chain move faster.[1]

Faster chains are great for efficiency, but they bring accountability concerns. If the chain communicates in formats humans can’t easily read, how do you audit it? One answer is to force every step to produce a human-readable output – a natural language summary of what it did and why. This sounds reasonable until you think about what you’re actually getting. LLMs are, by design, systems that produce plausible-sounding text. A reasoning trace or a summary of an intermediate step might look like a faithful account of the model’s process. But there’s no guarantee it is. It may be a post-hoc rationalisation – a tidy narrative wrapped around a process that was nothing of the sort.[2]

And would a manager who is broadly happy with the chain’s final output want to slow it down so a human can step into the loop and try to audit each step?

For basic transactional auditing – did the right data get passed? was the output in the correct format? – structured logs are fine. But for the kind of judgment-heavy, ambiguous work that characterises most knowledge work, the audit problem is largely unsolved. And as chains get longer, it gets worse, not better.

Toyota’s corrective

Ford’s assembly line is an imperfect analogy for modern work. Whilst technically brilliant, it was designed for high volume and zero variation: millions of identical Model T cars. Knowledge work is the opposite – high variation, with bespoke outputs, and highly context-dependent. A legal brief is not a chassis, and a marketing campaign is not a musket. (As an aside, Ford faced a ‘wrenching nightmare’ in the years that followed, according to Hounshell, as he tried to increase the flexibility of the plant and introduce variety).

Toyota faced a version of this problem in post-war Japan. The market demanded small quantities of many varieties of car, under conditions of low demand – circumstances closer to a modern consulting firm than to Ford’s Highland Park. The Toyota Production System, developed by Taiichi Ohno from the 1950s onwards (and articulated in a book of the same name), presents four principles that may help us better understand how AI chains could function.

One operator, many machines. Before Toyota, the assumption was one worker per machine. Ohno rearranged the factory so that a single operator could attend to three or four machines, intervening only when one stopped. There was understandable resistance – craftsmen didn’t want to become generalists – but it doubled and tripled productivity. The parallel for knowledge work is a human overseeing several concurrent AI chains across different functions: marketing, analysis and finance, for example.

Autonomation: automation with a human touch. Toyota’s machines were designed to detect problems and stop themselves automatically. Any worker could pull the andon cord to halt the entire line if they spotted a defect. The equivalent for an LLM chain would be built-in checkpoints – steps where the system self-evaluates and flags anomalies rather than blindly passing flawed output downstream. A human may not watch every step, but monitors the flow of work, tweaks the process, and responds when the system signals something is wrong.

Kanban as meta-information. Toyota’s kanban cards were packaged in vinyl envelopes and carried only the production information needed so the earlier and later processes could talk to each other. In a multi-LLM chain, this maps onto disciplined context management, passing only the relevant outputs and specifications between steps, not flooding each model with the entire history.

Pull, not push. Ford pushed products through the system based on forecasts. Toyota pulled them through based on actual demand. An AI chain triggered by a specific prompt (produce this campaign, for this client, to meet this goal) is inherently a pull system. The output is bespoke, drawn through the process by the request, not mass-produced and distributed.

Ford concentrated thinking within management. Toyota discovered this was economically inferior in complex, variable environments. Kaizen – continuous improvement driven by the workers themselves – outperformed top-down optimisation because the people closest to the work understood it best. If we build LLM chains on rigid sequences, lacking feedback loops, with no capacity for self-correction – we may get impressive throughput on routine work while creating brittleness at the edges.

What this all means

Today AI is mostly augmenting, rather than replacing, tasks. Humans are in-the-loop. But my sense is that, for many tasks, augmentation is a pit stop on a journey where full automation is the final destination. And as we chain these tasks together, humans begin to slip outside the loop: instead of steering it from within, they oversee it.

This means work is reorganised around several functions, and these do not necessarily sit with the same person: designing the chain (what agents, in what order, with what instructions and context), overseeing the implementation of the chain (and tweaking as needed), and judging the output (is this campaign any good? does it meet the client’s needs? would I stake my reputation on it?). The first two of these in particular are perhaps what the World Economic Forum and others mean by ‘agent orchestrators’.[3]

METR’s measure shows LLMs complete more complex tasks over time, with a doubling time of around 7 months. This chart is from 6 February 2026, and shows the duration at which an agent is predicted to succeed 80% of the time (a 50% success version shows longer time duration). Note the logarithmic Y axis. The measure is human expert completion time.

We’re on a trajectory of rapid improvement, but there is a caveat. METR’s research shows a steady increase in the ability of models to handle longer and more complex tasks.[4] However, the confidence ratings – in this case, models are assessed on whether they can complete a task successfully 50% or 80% of the time – tell us that reliability remains an issue.

For Ford, interchangeable parts was the prerequisite for reliability, and he solved this by 1913. The Toyota system emphasises a continual reduction in defective goods – a continual increase in reliability. For both, solving reliability meant the whole process ran smoothly.

What does this all mean? I think there are four key implications.

1. It means that for now we will have humans in the loop, shepherding tasks and part of the chain of work. When, or if, we increase the reliability of AI systems, those humans disengage from the loop.[5] Until then, audit is difficult to scale: a five-step chain with human review at each stage is manageable, but a fifty-step chain is not.

2. It means that new roles (chain designer, process auditor) are real but scarce. Ohno’s principle of one operator, many machines applies: the human who once executed a five-task chain becomes the human who monitors five chains running concurrently. But this does not necessarily imply one-fifth the headcount, as it depends on the other tasks left after automation.

3. It means that management skills are, as Ethan Mollick puts it, an AI superpower. I’d summarise this in two concepts. The first is compound engineering: most of your thinking takes place before you start work. Before, you might have spent 20% of your time planning and documenting what you want, and 80% executing. This now flips. Second, like any effective manager, you ask, ‘what do you need so you can do your job better?’. (This is captured nicely by the term context engineering).

4. It means a repeat of deskilling and reskilling cycles. Craftsmen who hand-fitted Singer sewing machines were displaced by assembly line workers, who were eventually displaced in turn. But workers in Toyota’s factories had to build up a broader range of skills. As Taiichi Ohno put it, applying human intelligence to machines was the only way to make machines work for people.

Back to the boss’s office

Let’s return to that meeting: your boss wants the SmartBrew Pro campaign, and you’ve got eight weeks. Only now it’s two years later, and the agency has deployed a chain of work, pulling through everything from marketing materials to consumer data analysis.

You write the brief on Monday afternoon. On Tuesday morning, there’s a full campaign to review, and media slots to approve. You didn’t collect the data, you didn’t interpret the trends, and you didn’t draft the scripts. You judged the brief going in, and you’ll judge the output coming out. Everything in-between happened without you.

(Quick aside: the five tasks chained together are actual tasks for 2431: Advertising and Marketing Professional, as per the International Standard Classification of Occupations (ISCO). There are nine tasks in total.)

There’s a question nobody in your office is asking yet: when you’ve been outside the loop for the entire process, how well can you actually judge what comes out the other end? The advertising specialist who spent years learning the craft – who could spot an unconvincing headline or a lazy script because she’d written hundreds – is precisely the person qualified to audit the chain. But she’s also the person the chain is designed to replace. And the junior who never built that expertise? They’ll inherit the auditor role without the knowledge to do it well.

If you leave the loop, you won’t get back in again. That’s the lesson of every previous wave of automation, from the Springfield Armoury to Highland Park to Toyota City. AI can already do some of your tasks. What we need to ask is whether it can chain them together, and what’s left of your job when it does.

[1] JSON, or JavaScript Object Notation, is at least a human-readable format, albeit not the way most humans would naturally communicate to each other. Other formats, such as vector embeddings, are not human-readable.

[2] Of course, this same problem applies when you’re only using one LLM for one task. But it is magnified when (a) tasks get longer, (b) are chained together (and call on external APIs), and (c) you can no longer micromanage the outputs. For example, the overall chain might produce outputs that no individual model ‘intended’, and debugging this becomes very complex.

[3] To stretch our car analogy, manufacturing roles all have their equivalent in the chain. Product designers describe and define the outcome; Process engineers are the architects of the chain; the Foreperson or Supervisor monitors for failures and exceptions; the Line Worker is an LLM executing a task; the Quality Inspector is in charge of validation/verification and could be a human or (if the human is just spot-checking) another LLM.

[4] METR’s study focuses on software engineering tasks. OpenAI’s GDPval looks at a broader range of tasks. Quite separately, OpenClaw demonstrates what AI delegation could look like, albeit in a form that would give any IT manager nightmares.

[5] The METR study does hint at another option, and one that was not feasible for car manufacturers: running a chain of work in parallel five or ten times, and selecting the best outcome.

There are moments we remember forever.

One of these, for me, is the final scene in Fight Club. When “Where Is My Mind” by The Pixies starts, and the corporation buildings that Tyler Durden equipped with dynamite begin to explode and crumble down.

I’ve seen the movie about 28 times (another music quote for my millennial friends - but more of a guilty pleasure), but it was mostly in my teen years. I was very sensitive to the angst and fascination with decay, though not grasping its full-blown accusation of capitalism, consumerism, and corporatism.

So I find it funny that in 2026, the film feels even more relevant than it was in 1999. It popped back in my mind this weekend, as I was chipping away the legs for my game table in one of my precious woodworking sessions, and the song came along on the radio (yes, on the radio). Fresh off having published the last article on OpenClawd (Red Lobster), a question emerged:

“What are we really delegating when we delegate to software?”

It’s been something I already tried to tackle in Enterprise Graph, then in You are training AI, but who gets paid for it?, but that only started to gain ground once I had the opportunity to start really testing this type of technology.

This article is not a tutorial. It’s a small piece of research on what it means to work, think, and decide in a world where AI starts to do it for us. Some things are good, some are bad. What changes, as usual, is how we use it, and if/how governments and corporations decide to build guardrails around it.

The OpenClawd Moment: Buildings Come Down

In the movie, the buildings coming down are Tyler Durden’s way of destroying the infrastructure of a system he despises. In my mind right now, those same buildings resemble something different: jobs and identities being redefined.

I’ve been talking about this shift for a while, but 2026 is the first year I can see things moving for real. Right now, SaaS companies are watching their stock prices correct as markets price in the possibility that AI can substitute their features. I don’t think it spells the death of all of them, but I do believe we’ll see far more homegrown products and software, especially for small and medium businesses. The building blocks are becoming accessible enough that the middlemen start to look expensive.

But the more personal tremor is this: once you delegate enough, it stops feeling like assistance and starts feeling like outsourcing parts of yourself. Not just the boring parts, not just the admin but the decisions you didn’t make, the messages you didn’t write, the reminders you didn’t remember.

Right now, the experience can be satisfying: it removes friction that was created by two decades of normalised performance and hustle culture (and too many SaaS products). But here’s my prediction: there will be a moment, not so far off, where in your day-to-day you won’t be entirely sure which part of your day you did, and which part your agent handled.

That’s what I mean by the mind being split.

We think of delegation as something external, like “I handed this off.” But it’s going to be closer to internal fragmentation: “Part of me did this, but I didn’t touch it.” Which brings us to the central question (sorry, had to use this image to quote my favorite fictional author here):

Tasks? Attention? Judgment? Identity?

The deeper I go, the more I think it’s all four.

The Four Layers of Delegation

1. Tasks

This is the most obvious layer. We let agents take over scheduling, drafting, filing, searching, summarising. Classic productivity moves, but even here, there’s more at stake than time saved. Research on cognitive offloading (Tsai et al., 2024) shows that repeated delegation of routine tasks leads to decreased recall and weaker task engagement, the so-called “Google Maps effect” applied to work. You save time, but you lose the muscle memory that made the task yours.

A 2025 MIT study saw students who used ChatGPT to write essays couldn’t produce a single accurate quote from their own writing afterwards. Their brain scans showed weaker connectivity between regions compared to students who thought first and used AI later. The researchers called this “cognitive debt”: you defer effort now, but pay later with weaker critical thinking and poorer recall. David Epstein, writing about the study, framed AI-first work as the opposite of what cognitive psychologists call “desirable difficulty.” He called it an “undesirable ease.” The work gets done, but nothing sticks.

What’s interesting is that students who did their own thinking first and then used ChatGPT performed well. Their essays were better structured, they remembered what they wrote, and their brain activity looked healthy. The tool wasn’t the problem, the sequence was: brain first, tool second. This distinction alone could reshape how we think about delegation at work.

2. Attention

What we automate, we stop paying attention to. When agents filter our inputs, we lose the ambient awareness that comes from skimming an inbox, scanning a feed, or casually checking in. Liu & Pelowski (2023) showed that reliance on LLMs for triaging and prioritisation can lower attentional resilience, which is our ability to maintain context over time. The small, forgettable acts of paying attention turn out to be load-bearing.

There’s also a subtler attention tax that I don’t see people talking about yet. Even when we delegate to AI, the more complex the task, the longer the agent takes to respond. And in that gap, we do what we’ve been trained to do by a decade of smartphones: we context-switch. We check another tab, glance at notifications, start a different thread. So the tool that was supposed to free our attention actually fragments it further, because we can’t sit still while it works (and it takes an average of 30 minutes to refocus for the human brain). That means that potentially, we’re not saving attention, we’re scattering it across more surfaces.

3. Judgment

The more we delegate prioritisation or decision-making, the more we risk eroding our intuitive judgment. A 2023 Stanford study (But et al., 2023) found that over time, users relying on AI for decision support in ambiguous tasks began deferring more quickly and questioning their own reasoning less. That menas that delegation doesn’t just outsource effort, it can atrophy discernment.

Seneca wrote that “no man is ever wise by chance.” Ryan Holiday, building on this in his recent book Wisdom Takes Work, argues that wisdom isn’t gifted by age or intelligence but earned through repeated effort - it’s a verb, not a noun. The same is true of judgment; you don’t develop it by reading a summary of someone else’s reasoning. You develop it by sitting with ambiguity, making calls, getting some wrong, and refining your instincts over time. If we skip that process and let the tool decide, we don’t just lose time. We lose the reps.

But here’s a question worth sitting with: are all of these layers sequential? Does delegation always start with tasks and slide toward identity, or can someone jump straight to delegating judgment without realising it? I suspect the answer is yes, and that’s what makes this harder to design for than it looks.

4. Identity

This is the most subtle and maybe the most profound: when your agent sends the email, files the doc, writes the first draft, whose voice is it? Work used to be a reflection of your craft, your thinking, your presence. Now, as we shift to directing rather than doing, authorship becomes distributed. When value is produced in invisible layers of automation, how do we measure contribution? And more importantly, how do we maintain a sense of ownership over work that increasingly isn’t done by us?

Each of these layers has a trade-off, and the consequences appear downstream, not immediately. This is why they’re hard to see, but also why it’s worth paying attention now.

So when we say we’re using AI to save time, we should ask: which part of ourselves are we asking it to stand in for?

The Case For Delegation (Before we blow it up)

Before going further into what delegation erodes, it’s worth being honest about what it enables, because the optimistic case is real too.

For many people, delegation doesn’t diminish them: it can free them. A designer who offloads invoice management isn’t losing part of their identity, they’re reclaiming time for the work that is their identity. A founder who automates meeting summaries isn’t diffusing their cognition, they’re concentrating it where it matters most. A neurodivergent professional who uses AI to structure their day isn’t becoming dependent, they’re gaining access to a kind of executive function support that didn’t exist before.

The risk isn’t delegation itself. It’s undiscerning delegation, handing things off without asking what you’re losing in the process. I think the people who will thrive with AI tools tend to be intentional about what they keep and what they release.

A Simple Filter: The Delegation Compass

When I started being more intentional about what I hand off to AI and what I keep, I found it helpful to run each task through a quick filter. Not a rigid system, more of a compass. Four questions:

1. Is this a skill I want to keep sharp?

If yes, do it yourself, even if AI could do it faster. Writing, strategic thinking, negotiation, creative problem-solving. These are muscles. Outsource them long enough and they atrophy. The convenience isn’t worth the cost if the skill matters to your identity or career.

2. Is this a task where my judgment is the point?

Some tasks exist because they require your specific perspective: reading the room in a client situation, deciding what to prioritise this quarter, choosing the right tone for a difficult message. If the value comes from your judgment, delegating it isn’t efficiency.

3. Would I notice if this was done poorly?

If your agent summarises a meeting badly or sends a slightly off-tone email, would you catch it? If yes, delegation works, you’re still the quality filter. If no, you’ve created a blind spot. The more invisible the delegation, the more dangerous it is.

4. Does doing this connect me to people or context I’d otherwise lose?

Sometimes the “inefficient” task, like scanning the inbox, joining the standup, writing the update yourself, is actually how you stay plugged into what’s happening around you. Automate it away and you might save thirty minutes but lose the peripheral awareness that makes you effective.

In practice, I’ve found that most people over-delegate on questions 2 and 4, and under-delegate on the genuinely mechanical work where AI shines and nothing meaningful is lost.

From Delegation to Dependency

We start by automating friction, but friction often served a function.

Just like GPS replaced our need to build spatial maps in our head, and in turn weakened our ability to navigate unaided, software agents may dull our sense of sequence, priority, and even memory. We begin to use the tool not just to do more, but to remember how to do at all.

A 2025 study published in The BMJ examined death certificates from nearly 9 million people across 443 occupations and found that taxi drivers and ambulance drivers, who do constant real-time wayfinding, had the lowest risk of dying from Alzheimer’s. Bus drivers and pilots, who follow predetermined routes, showed no such benefit. The cognitive effort of navigating in real time, the daily difficulty of figuring out where to go, appeared to protect the brain. When David Epstein asked one of the study’s co-authors whether GPS might erode this benefit for future drivers, the response was telling: “The now ubiquitous use of GPS could, in theory, mitigate the spatial memory based benefits that prior generations of taxi or ambulance drivers may have developed.” The convenience removes the challenge, and with it, the benefit.

I notice this in small ways. I catch myself reaching for the agent before I’ve even tried to think through a problem - not because I can’t, but because the path of least resistance has shifted. Dependency grows not from a single choice, but from repeated relief.

We also become dependent not just functionally, but psychologically. We get used to instant access, always-on assistants, smart recommendations. I see it happening when my son scrolls endlessly through shows, not because he’s engaged, but because he’s come to expect that something better is always one swipe away. In us adults, it takes other forms: jumping between tabs, cycling through to-do apps, refreshing notifications. These aren’t failures of discipline. They’re symptoms of abundance. Our brains evolved for scarcity, and the sudden overload of delegation, information, and options doesn’t create clarity. It creates craving.

The Multiplication of Selves

Once delegation is normalised, and dependency starts to take hold, something new kicks in: We aren’t just outsourcing parts of our cognition, we are multiplying them.

One you drafts strategy. Another executes ops. One handles outreach, another files reports. You’re still in the loop, but your cognitive bandwidth is split across different contexts, each running in parallel.

In Fight Club, Tyler Durden and the narrator are the same person, a split mind acting across incompatible realities, fractured by contradictory roles and desires. And that was 1999, a metaphor for industrial-era alienation.

What happens when work actually mirrors that? Because here’s where I think we’re heading: as AI drives productivity up, expectations rise in parallel. People are asked to manage more workstreams, shift faster between strategic and tactical tasks, switch hats more often. Not because it’s healthy, but because now it’s possible.

And this creates a vicious cycle that I’m already starting to feel.

The faster you work, the more you’re expected to produce in the same amount of time. So you delegate more to keep up. But now you’re delegating things that didn’t need to be delegated, things you would have been better off doing yourself, because the pressure to output has outpaced your ability to be thoughtful about what you hand off. The quality of your thinking degrades not because you’re lazy, but because you’re running to keep up with a pace that your own tools set. You start producing what I’d call slop: work that technically got done but that nobody, including you, actually thought through. And the worst part is, the faster cycle makes it harder to notice, because you’re already on to the next thing.

When your workday is a rapid shuffle between six mental states (editor, analyst, manager, recruiter, planner, fixer) you start to lose track of which version of you is the real one. Or if they all are. Or none.

This fragmentation intersects with something we don’t talk about enough in productivity conversations: loneliness.

We are already living through a social crisis, exacerbated by digital platforms, remote work, and growing individualism. If your systems know you better than your colleagues do, and if your assistant is your primary collaborator, what happens to the communal fabric of work?

There’s a question here that genuinely unsettles me: when you get used to asking AI for everything, and expecting immediate, exhaustive answers, can you still switch to human communication with its pauses, its ambiguity, its need for empathy and patience?

Because human interaction is slow - it’s inefficient, it requires tolerating not-knowing, sitting with discomfort, building understanding over time. If our baseline shifts to the speed and precision of AI responses, the friction of human conversation might start to feel like a bug rather than a feature.

Who Controls the Extended Mind?

In 1998, philosophers Andy Clark and David Chalmers proposed a radical idea in their essay The Extended Mind. The thesis was simple but profound: if we regularly use external objects, like notebooks or devices, to store and process information, those objects should be considered part of our cognitive system. Thinking doesn’t stop at the skin. It stretches into the world.

When that world includes always-on agents, auto-completing systems, and task-oriented memories, the boundary between brain and tool gets even blurrier. If your assistant finishes your sentence, prioritises your day, and summarises your calls, is that still you thinking? Or are you co-thinking with a machine?

And if part of your cognition happens in a tool you don’t fully own, like an enterprise LLM or a rented platform, who owns the thinking?

We are entering a phase where parts of our minds are embedded in rented software, shaped by algorithms, monitored by employers, and retrained on our patterns of thought. A note you wrote in Notion might be semi-private. An email draft written with AI might live in training logs. A decision made through a delegated workflow might leave behind a traceable record, even if you never typed it yourself.

Designing the Split

We are living in the equivalent of what 2007 was for social media. The beginning, but also, potentially, the beginning of the end of something we haven’t yet learned to name. The end of undivided attention, perhaps. Or the end of unsupervised cognition.

Can we stop, think, and design things right this time around? Or, much like all of human history, will we need to hit a wall before we start to scale back?

I don’t think those are the only two options, but I do think the window for the first one is smaller than we’d like to believe. The tools are moving fast. The habits are forming faster. And the questions this piece has tried to raise, about what we delegate, what we lose, and who owns the parts of us that live inside machines, are not ones we can afford to answer by default.

So here’s what I’d leave you with. Not a prediction, but an invitation:

Pay attention to what you’re handing off. Not just the task, but the thinking behind it. Notice when delegation stops feeling like a choice and starts feeling like a reflex. That’s the edge worth watching.

Until next week

Matteo

Welcome to this week's edition!

Unless you have been living under an AI-generated rock, you’ve probably seen your social media feeds filled with a red lobster emoji: the story of Clawdbot, now OpenClaw, and its accompanying AI agent social network.

Unlike the videos you may have seen, my goal is not to show you how it works technically. It’s to understand why it caught on like wildfire and what that reveals about the future of work.

Let’s dive in.

What the hell is OpenClaw?

Unlike chatbots that wait for prompts and respond with text, OpenClaw takes action on your behalf on your own computer. It can control parts of your system, use your apps and files, trigger workflows, and execute multi-step tasks.

Technically, OpenClaw isn’t a new AI model: it uses existing large language models as its brain. What makes it different is that it gives those models hands: system access, tool use, and the ability to operate software. You can get Claude and ChatGPT to do many of these things through MCP connections, but it requires several setup steps: OpenClaw just made it simpler.

Over the last week, people have experimented with three broad categories of use:

Personal automation and coordination - Many use it as an always-on executive assistant: checking inboxes, preparing daily briefs, organizing reminders, and flagging what needs attention. The appeal is offloading the small coordination work that fills the day - or like someone else did, automate sending messages to your partner (🙃).

Work and ops monitoring - Others have delegated operational tasks: monitoring release workflows, watching GitHub or issue trackers, updating to-do tasks, or alerting when jobs finish.

Custom workflows and experimentation. The rest have built custom skills, connected APIs, automated home setups, or created personal routines. Some plugged it into their notes or knowledge bases, asking it to recall information, summarize data, or connect ideas.

Obviously, this tool is a security nightmare. You’re essentially giving it keys to all your information, which it can store and potentially share or breach. But that’s not really the point: over 2 million people visited in one week, and many will gladly take the risk until something happens or until someone releases a more secure version.

What caught my attention is how it spread. For a long time, marketing has been about persuasion: telling a compelling story, building awareness, and convincing people something is valuable before they touch it. In AI, that model is shifting as the center of gravity moves from persuasion to participation.

The question is no longer “Did people understand it?” but “Did people try it?”.

So overall, I think one of the reasons OpenClaw spread more because it was easy to try, easy to show, and socially rewarding to talk about. Using it became a small signal of being early, curious, and AI-literate. It functioned as experiential marketing disguised as a tool.

Part of the viral moment also came from the AI agent social network that launched alongside it, Moltbook. People were fascinated, but then news broke that it was mostly staged (the installation included a Markdown file that let humans direct these agents to write and post about certain things. There was automation, but also human orchestration). Someone then built “MoltHub,” a clever PornHub parody, which added more fuel to the hype.

The marketing "lesson” here is that FOMO now acts like a distribution engine. When a new AI tool appears, people don’t want to be the last to test it and trying becomes a way of signaling relevance: I’m up to date, I’m exploring, I’m not falling behind. The irony is that this often says more about identity than utility.

Early success in AI can look less like product-market fit and more like hype-market fit: a tool can capture attention long before it becomes part of real workflows. That’s not entirely bad, since attention accelerates experimentation and surfaces new ideas, but it also shortens reflection cycles and increases the risk of confusing virality with value.

Why delegation tools are catching on

The reason this spread like wildfire is partly about what it can do today, but mostly about what it promises to change about how we work tomorrow.

Over the past decade, knowledge work has quietly become heavier. Not necessarily harder, but more fragmented, more layered and more tool-dependent. We’ve seen an explosion of productivity tools that we need to connect, switch between, mentally set up, and constantly tweak in everyday work.

This mirrors something we’ve discussed before in Work 3: that learning and managing AI and tools can start to feel like a second job in itself, adding cognitive load even when the promise is efficiency (https://wrk3.substack.com/p/learning-ai-feels-like-a-second-job).

A few tensions stand out:

Paradox of choice - We drown in tools. Personally, I have over 100 saved from Product Hunt that I never get to try. Every workflow begins with deciding which tool or stack to use. Choice creates friction before real work even starts.

Fragmented work environments - Work lives across tabs, apps, and channels. Moving between them resets attention and breaks flow. The day becomes navigation between systems rather than progress within them.

Coordination tax - A large share of work is coordination, not creation. Scheduling, aligning, following up, confirming. It’s invisible labor that consumes time but rarely feels meaningful.

For the past couple of years, AI has helped, but mostly as an assistant. It drafted, summarized, and reviewed, but it still waits for us and depends on our constant input.

Agentic tools like OpenClaw hint at something different: they have full control and access, they add memory so interactions accumulate context, and they add hands so the system can execute actions, not just suggest them.

That combination creates a new proposition: not help, but delegation. The fatigue with the current state of work is so real that many people are willing to trade a surprising amount of control for relief. They’re willing to hand over the keys to parts of their digital lives in exchange for cognitive breathing room.

Some enthusiasts are even running multiple agents in parallel, chasing levels of leverage that used to require entire teams.

This made me question: how will this change the way we work?

Is the real story not how powerful these tools are, but whether we actually need them and whether we’re just exhausted by how we work now?

The layers beneath delegation

The concept of delegation sounds simple: let machines handle more so humans can focus on higher-value work, but its implications unfold in layers. Yes, like an onion.

The first thing people notice when they start using agentic tools is apparent relief, the kind that comes from realizing they no longer have to personally carry every small action in their day. Something that used to hum in the background of the workday, that low-level operational noise, begins to fade.

What’s interesting is how quickly that relief evolves into a different experience.

Once delegation works reliably, the pace of work changes almost on its own. You find yourself researching faster, coordinating faster, and executing faster, not because you’re rushing but because the system removes friction. Output rises, sometimes dramatically, and for a while it truly feels like you’ve unlocked a new level of personal capability. We start rushing into things at 2x.

Then, gradually and without much fuss, the baseline shifts.

What looked impressive a few months ago begins to feel normal. The extra capacity that once felt like breathing room becomes something others factor in. If your AI enables you to handle twenty things at once, people around you start planning as if twenty things at once is simply what you can do. Not as an explicit demand, but as an adjustment to a new reality.

We’ve seen this pattern before in other technologies: email reduced the time it took to communicate and then multiplied the amount of communication expected. Messaging platforms lowered friction and, in doing so, normalized near-constant availability. Tools that save time rarely give that time back for long. More often, they reset the pace at which work moves.

Delegation may follow a similar trajectory, where the benefit is real but the system recalibrates around it.

Somewhere in the middle of this shift, another change begins to surface, one that has less to do with speed and more to do with how we relate to our own output.

When your AI drafts most of your emails, organizes your calendar, prioritizes your tasks, and follows up in a tone that sounds like you, the boundary between your work and your system’s work starts to blur. It doesn’t happen all at once, and it doesn’t feel dramatic. It’s gradual and almost easy to ignore.

You still feel responsible for the outcome, but you’re no longer touching every step. At some point it becomes harder to answer a simple question: did I do this, or did my setup do this?

That distinction matters more than it first appears. Much of how trust and reputation are built at work relies on the sense that a person stands behind what they produce. When more of that production is mediated by systems, authorship doesn’t disappear, but it becomes less visible and more distributed (this connects closely to another Work 3 discussion about how humans are already training AI systems and contributing value in ways that are not always visible or compensated).

The bottom line is that delegation doesn’t just redistribute tasks, it potentially redistributes identity.

If delegation comes to the workplace

Today, tools like this feel experimental, risky, and not enterprise-ready. Most companies wouldn’t let an autonomous agent touch their internal systems, and for good reason.

But it’s worth separating today’s implementation from tomorrow’s direction: If safer, more governed versions emerge, delegation could start to reshape how organizations think about work.

Some roles today are largely coordination roles: following up, routing information, tracking progress, making sure things move - which aligns with a broader idea we’ve explored about work becoming increasingly programmable and orchestrated, where designing and managing systems becomes as important as executing tasks.

The real question is not whether delegation will enter organizations, but how.

Because it likely won’t start with official rollouts. It will start quietly, bottom-up, with individuals using agents to manage their own workload, automate small tasks, and gain leverage where they can. In other words, a new wave of Shadow AI, or even Shadow Agents.

If those individuals perform better, respond faster, and ship more, organizations face a dilemma: enforce strict rules, or tolerate the behavior because it improves outcomes.

We’ve seen this pattern before with email, smartphones, and collaboration tools. Each promised efficiency and delivered it, but also quietly redefined what “normal” output looked like.

Which leads to the harder question.

The deeper issue isn’t whether delegation tools will enter the workplace, but how organizations will draw boundaries around them, and whether those boundaries will be proactive or reactive.

Not everything that can be delegated should be, and deciding what shouldn’t be may become one of the defining management challenges of the next decade.

Open Questions We’re Just Beginning to Ask

After the hype, the demos, and the experimentation, what stays with me are not firm conclusions but open questions, which I take a stab at - but would love your thoughts on in the comment section.

If everyone has AI leverage, does leverage still differentiate anyone?
My sense is that leverage doesn’t disappear, it moves. Execution becomes cheap, so differentiation shifts toward judgment, taste, and deciding what to pursue in the first place. The scarce resource may not be output, but direction.

Are we reducing work, or just raising the expected output per worker?
History suggests the latter. Many productivity tools saved time locally but increased expectations globally. Delegation might give breathing room at first, then reset what “normal productivity” looks like.

If early-career workers delegate too much, where do they build mastery?
A lot of expertise is built in the repetition of small tasks, the very tasks we’re eager to automate. Delegation could accelerate learning for some, but it could also remove the friction that teaches. We may need to rethink how mastery develops in an AI-rich environment.

Will companies trust human judgment less if systems appear more consistent?
Systems are often more predictable than people, and organizations like predictability. The risk isn’t that humans disappear, but that their discretion narrows. Human judgment may become the exception layer rather than the default.

If your agent does the work, what exactly are you being paid for?
Perhaps for defining goals, setting constraints, and taking responsibility for outcomes. Work might shift from doing to directing, from producing to being accountable. That’s still valuable, but it’s a different identity than many roles were built on.

What about the people who can’t or won’t adopt these tools?
Delegation creates new forms of leverage, which means it also creates new inequalities. Not everyone has the technical literacy, resources, or desire to manage AI agents. As these tools become more common, does a new divide emerge between those who can multiply their output and those who can’t? (we discussed these broader structural shifts in how work is distributed and modularized when discussing the decentralised workforce).

Maybe the real signal in the OpenClaw moment isn’t that AI can now act for us, it’s that many people are ready to let it. After a decade of managing tools, dashboards, and notifications, the appeal of delegation feels as much emotional as it is practical.

The challenge we will have is whether we use delegation to create more meaningful work, or simply to fit more work into the same hours. That, in the end, isn’t just a technology choice, it’s a cultural one.

That's why I think that in a world where trying tools becomes a signal and delegation becomes easier, the real differentiator may not be speed of adoption but clarity of intention.

Until next week!
🦞

This week, we’ll be introducing each of the five finalist research teams for the Future of Work Global Research Prize.

These proposals come from leading universities across Europe and the US, and they all tackle questions that sit right at the heart of Work 3:
how work is changing as AI scales, identities fragment, power dynamics shift, and coordination becomes more complex.

We care deeply about closing the gap between academic insight and organisational reality. Too often, rigorous research arrives years after leaders need to make decisions. These projects feel different: ambitious, and grounded in real-world work.

A big thank you to Professor Jochen Menges, the University of Zurich, and the Adecco Group for highlighting and supporting this work.

At the end, I’d encourage you to cast your vote for the research you find most compelling.

🗓 Voting deadline: 30 January 2026

1 - Can a generative AI chatbot improve the wellbeing and performance of office workers over time?

This project tackles one of the most practical questions organisations are quietly asking: does everyday AI support actually help people?

To answer it, the team is running a large-scale field experiment inside a single firm with 40,000 employees across 96 countries, tracking real wellbeing and performance outcomes over 18 months. For example, one hypotheses being tested is does greater use bring larger impact?

Led by Dr William Fleming, with Jan-Emmanuel De Neve and Cherise Regier from Oxford’s Wellbeing Research Centre, this study stands out for its scale, duration, and realism.

From a Work 3 perspective, this matters because we are moving beyond AI pilots. If AI is becoming part of the daily work environment, we need evidence not just on productivity, but on human sustainability over time.

2 - In a world of “perfect” AI output, does imperfection become the new creative advantage?

As generative AI produces increasingly polished work on demand, this research asks a provocative question: will creativity be redefined not by perfection, but by deliberate imperfection?

Prof. Daria Morozova (Leiden University) and Dr Sumeet Malik (University of Amsterdam) explore how human creators may differentiate themselves by curating flaws, ambiguity, and idiosyncrasy — precisely the things AI tends to smooth out. With some examples from Daria’s own poems Vs LLM poets.

As tasks become automated, human value shifts from execution to judgment, taste, and intention. Understanding how imperfection functions as a signal of humanity may become central to future creative work.

3 - What happens to your digital past when you’re looking for your next job?

In a labour market shaped by online traces, Jaylon Sherrell (Harvard Business School) examines how workers manage identity during job search.

People change. Opinions evolve. Yet digital footprints often persist for decades, resurfacing at precisely the moment of highest scrutiny: hiring.

This research speaks directly to Work 3 themes of identity, ownership, and control. Who owns your past self in a world of searchable histories? And how should organisations interpret digital traces that may no longer reflect who someone is today?

Subscribe now

4 - Why is delegation riskier for women — and how can organisations change that?

Delegation is meant to reduce workload and enable leadership. But this research starts from an uncomfortable truth: delegation is not experienced neutrally.

Anisha Singh (LSE), with Laura Giurge (LSE) and Cydney H. Dupree (UCL School of Management), show that women managers face a “double gaze” when delegating — judged both by those above them and by those they lead. This pressure can intensify further when race is taken into account.

Based on 30 qualitative interviews and pilot experiments, the findings suggest that delegation can increase leadership strain for women rather than relieve it. The next phase, a large-scale experiment with 800 women managers, will test these dynamics even further.

From a Work 3 lens, this research will highlight how work design and power dynamics still shape outcomes.

5 - How do humans and AI actually work together under pressure?

Studying human–AI collaboration in real time is hard. Professor Steven Zhou (Claremont McKenna College) takes a bold approach: AI-enabled escape rooms.

By placing human teams and intelligent agents into time-pressured problem-solving environments, the research captures leadership, coordination, and communication as they unfold using cameras, microphones, and wearable sensors.

The questions are very practical:

• When does AI help teams perform better?

• When does it get in the way?

• What leadership behaviours enable effective human–AI collaboration?

What makes this especially compelling is the blend of experimental control and realistic pressure exactly the conditions future teams will face.

Vote for Your Favourite Research Idea Here!

Vote for Your Favourite Idea Here

We highlighted last years winner of the prize, from Max Bogaert et al, with our article

Dystopia or Design? Imagining Work Futures Through Science Fiction

His thesis: dystopian science fiction films are cultural artefacts that help us understand how people imagine the future of work.

These projects offer a glimpse of where Future of Work research is heading — and where leaders will soon need answers.

Joining the dots,

Andy

Welcome to this week’s edition!

This week, I want to address what I think is one of the biggest mistakes we keep making about AI and work:

We think the battle is about jobs.

It isn’t.

The battle, is going to be about ownership.

AI systems are trained through everyday work. We correct them, guide them, refine them, and teach them what “good” looks like. That knowledge doesn’t disappear when the task is done, it gets embedded into systems that scale.

Previous technologies (think of the industrial revolution) didn’t have this immediate and total absorption mechanism. Yet, 100 years on, we pay for time just like we did when people worked in factories.

That’s why today I want to deep dive into how we need to redesign ownership and compensation in the era of AI.

Where AI creates value - what people get wrong

Right now, 99% of the debate is framed around jobs, productivity and efficiency, as if we’re still inside the industrial system where value is created by hours worked and outputs shipped.

The catch is that AI doesn’t just accelerate work (which we have already argued is not always a good outcome, when it scales the wrong workflows faster, or optimizes for the wrong outcomes "AI Productivity vs Performance”, "The Impact of AI on Work (Isn’t Jobs)"). It absorbs it and turns work into training data.

Most new and native AI systems (as we talked about in "The Enterprise Graph") are explicitly designed to rely on continuous human input, feedback, supervision, and correction to function at scale: every prompt, judgment call, document, and online meeting is not just “using AI.” It teaches a system how to behave, and it turns everyday work into training signal.

This is also why “agent skills” are becoming such an important concept. In our piece on Anthropic’s Agent Skills, we argued that a Skill is essentially a digital representation of procedural knowledge, a reusable module that captures how work is done so agents can execute it reliably. In other words, the work is not just performed. It gets packaged, versioned, and encoded into systems. Once work becomes programmable in this way, the ownership question becomes unavoidable, because what you are producing is no longer just output.

This dynamic is likely to intensify over the next decade.

Research firm Epoch AI estimates that the effective stock of high-quality public human text is finite, and that frontier models could fully utilize it between 2026 and 2032. If public data becomes a bottleneck, the next training frontier will shift from “the open internet” to something else: proprietary workflows, human feedback loops, and enterprise context. Once that happens, the question is no longer whether AI uses human knowledge, but who owns the intelligence that gets captured. Researchers increasingly describe this as instructional labor, data labour, or human-in-the-loop work: work that produces durable economic value, but remains largely underwater in our compensation models.

Yet we continue to evaluate and compensate people as if their contribution disappears the moment they log off. That’s because knowledge work has not been redesigned for the AI era, it simply inherited the rules of industrial labour (time-based pay, task-based productivity, zero ownership of accumulated knowledge).

Those rules worked when work was ephemeral, but they break when work becomes cumulative, reusable, and scalable without the worker.

We’ve already lived through one great wave of value extraction: Social media trained us to give away personal data in exchange for free products. Search became one of the most profitable business models in history precisely because people accepted this trade-off: convenience now, monetization later. But AI at work is not the same thing. When a worker uses AI to be faster, more accurate, or more creative, the value does not accrue to the them individually, it compounds inside the company. The system learns from professional judgment, edge cases, and institutional context, all while the worker is paid as if nothing durable was created.

Data as labour

The truth is that labour is becoming equivalent with data.

Across academia and industry, the same pattern is described in different ways, but it shows up in very specific roles inside real companies:

Data labor - At companies like Scale AI and Sama, this work appears under titles such as data annotator, labeling specialist, or AI trainer. These roles involve tagging images, classifying text, ranking outputs, and correcting model behavior. While often framed as low-skill or entry-level, this work requires sustained attention, contextual judgment, and consistency. It is also frequently outsourced to workforces in Africa, South Asia, and Southeast Asia, where wages are a fraction of those in Western markets.

Feedback labor - Large model developers like OpenAI and Anthropic rely on roles such as AI trainers, model evaluators, and alignment reviewers. These workers compare outputs, flag failures, enforce tone and safety norms, and teach models what “good” looks like. Despite shaping how systems behave at scale, this labor is typically contractual, anonymized, and disconnected from any long-term ownership or upside in the systems being trained.

Human-in-the-loop work - In enterprise automation platforms like UiPath and Palantir, humans are embedded as process supervisors, operations analysts, or exception handlers. Their job is to intervene when automation breaks, resolve edge cases, and gradually encode those decisions back into the system. This is often presented as oversight or QA, but in practice it is incremental system design, where human judgment is steadily converted into machine behavior.

Invisible labor - A significant share of moderation, safety review, and model supervision is handled by large outsourcing firms such as Accenture and Cognizant. Job titles here include content moderator, trust and safety analyst, and AI operations associate.

In all cases, people are performing instructional work that improves systems over time. The output is not a task completed, but intelligence accumulated.

And while this is increasingly becoming a recognizable category of work, it is still compensated as if nothing enduring is being created (though one could argue that at least it is compensated, if you think about how most LLMs have been trained on stolen copyrighted data).

Another way to see this shift is through the industry’s upcoming obsession. The emerging reality is that models are becoming commoditized, and context is the moat. Prompt engineering is giving way to context engineering, the discipline of curating and maintaining the right information at inference time. In enterprise settings, the limiting factor is often not the model, but operational context, the live signals, workflows, and metadata that tell systems what is actually happening. This is where the ownership question becomes unavoidable. Context is increasingly the moat, and most of that context is generated by work.

As we expand usage of this technology, we increasingly start to capture any type of work and intellectual output as part of what we called the Enterprise Graph: basically a digitalized version of every interaction, thought, and output of knowledge work with its history, connections, evaluations and so on.

This means salary, which was built for work that ends, is not sufficient anymore.
And this won't be a question about fairness, but about work design.

Intellectual Equity: a better mental model

So what can be a better option? Not higher wages, but intellectual equity. The idea is simple: when human contribution continues to generate value through AI systems, compensation should be recurring, attributable, and durable.

In practice, this can take four main forms.

1. Royalties for AI-generated value

One prominent alternative model is to compensate contributors through royalties tied to the ongoing use of an AI system.

Instead of one-time payments for data, labeling, or training, contributors receive recurring micropayments whenever the system generates value. This mirrors how artists earn royalties from music streams or book sales and reframes human contribution as a long-term asset rather than a consumable input.

A concrete example is Codatta. Codatta turns expert human knowledge into structured data assets and tracks how those contributions are used during model inference. Every time an AI model powered by that data produces an output, contributors receive a micropayment via smart contracts.

This model aligns incentives across the entire AI lifecycle. Contributors are rewarded for quality and durability, not volume, and are treated as partial owners of the value the system creates over time.

At a conceptual level, this shifts AI work from a gig economy of tasks to an ownership economy of intelligence.

2. Inference-time compensation for domain expertise

A related proposal focuses on when value is actually created.

Much of an AI system’s economic value emerges after deployment, through millions of inferences, API calls, and downstream decisions. This brings the argument that compensation should therefore occur at inference time, not just during training.

In this model, contributions from skilled workers are tagged, logged, and traced. When a system applies patterns, heuristics, or judgment learned from a specific contributor’s data, that contributor receives a micropayment.

In effect, every time a system uses techniques learned from a worker’s expertise, the worker earns a share of the value generated.

While technically complex (inference from models is hard, but not impossible), this approach has clear precedents in industries like music streaming and software licensing. Most importantly, it reflects a deeper shift in how we think about work: from being paid to perform tasks, to being paid for the reuse of intelligence. I would also argue that if we cannot trace contributions precisely, companies will default to zero compensation - so an even stronger argument for redesigning work accordingly.

3. Equity, tokens, and collective ownership

Beyond royalties, some approaches treat contributors as stakeholders, not just licensors.

One proposal gaining traction among labor advocates is collective equity. For example, unions or pension funds could acquire minority stakes in AI or data-labeling companies, allowing workers to benefit from long-term growth rather than short-term contracts.

In decentralized ecosystems, this logic is already being tested. Platforms such as Streamr Data Union DAO issue tokens to users who contribute data. These tokens represent governance rights and a share of revenue when the data is licensed or used.

The mechanism differs, but the principle is consistent: those who supply training data or system improvements should hold an economic claim on future returns, whether through profit sharing, voting rights, or ownership units.

This reframes contributors not as outsourced labor, but as capital partners in AI systems.

4. Attribution, licensing, and IP as compensation layers

Money is not the only form of equity: attribution and intellectual property rights can become a second pillar of this shift.

In creative fields, there is growing consensus that AI training constitutes a distinct use of intellectual property and requires explicit consent and compensation.

LUISS Professor Christophe Geiger has argued this approach protects authors’ rights while preserving innovation, and has proposed statutory remuneration schemes for AI training data. This would imply that instead of forcing creators into litigation, AI developers would pay into collective funds whenever copyrighted works are used, with proceeds distributed directly to creators.

Similar ideas are appearing globally. Brazil’s draft AI Bill proposes mandatory remuneration tied to company size, aiming to prevent large platforms from extracting cultural value without proportionate compensation.

At the contractual level, industry groups are adapting: the Authors Guild has made clear that AI training rights are not covered by standard publishing agreements and must be negotiated separately. Their recommendation is to treat AI training as a subsidiary rights deal, with a defined revenue split. Creative contracts are beginning to include AI-specific clauses that regulate whether work can be used to train models, under what conditions, and with what royalty structure as derivative use expands.

Who wins, resists, and why this still happens

Is this utopia? Maybe. The main resistance is of course that this type of model would reduce value extraction and concentration from who is deploying the technology itself. Plus, it would also slow the current arms race, because it would require redesign, approvals and so on. On top of this, there are three other reasons:

It breaks the accounting model - Salary is simple. Royalties, usage-based compensation, and equity are not. They introduce long-tail liabilities, variable costs tied to success, questions about attribution and measurement. From a CFO perspective, this turns predictable labor costs into open-ended obligations. Even when companies acknowledge the fairness of the model, they struggle to fit it into payroll systems designed for industrial work, not compounding intelligence.

It weakens ownership and control - Traditional employment contracts assume a clean transfer of value. Intellectual equity challenges that assumption: if contributors retain economic claims on trained systems, companies no longer fully own the intelligence they deploy. That creates discomfort around IP, governance, and exit scenarios. In practice, many firms prefer to treat training, feedback, and supervision as usage rather than contribution, because usage does not come with rights.

It exposes uncomfortable power dynamics - Much of the work that trains AI is outsourced, anonymized, geographically distant, poorly paid. Recognizing it as value-generating intellectual labor would force companies to confront how much of their AI capability depends on invisible global workforces. Avoiding intellectual equity is often a way to avoid that conversation.

And yet, some organizations are moving and will move in the opposite direction - not out of altruism, but out of necessity.

The companies that experiment with intellectual equity tend to share a common constraint: they cannot succeed without sustained, high-quality human contribution. When an AI system depends on ongoing expert input, supervision, or refinement, contributors cannot be treated as disposable. One-off payments fail to align incentives.

This is why platforms like Codatta adopt royalty-based models. Their value proposition depends on experts continuing to contribute over time. Royalties reward durability and quality, not just participation, and create a direct link between contribution and downstream value.

A second group of early adopters operates in regulated or rights-heavy domains, where attribution and licensing are already part of the economic fabric. Creative industries understand residuals, secondary use, and long-tail value. Stock media platforms moved first not because they were idealistic, but because the logic was familiar. Shutterstock created a contributor fund to share revenue from AI training datasets. Adobe introduced bonuses for creators whose content trained its generative models. These companies did not invent new compensation models. They extended existing IP logic into AI.

A third category competes primarily on trust. In domains like voice, likeness, and creative identity, consent is not optional. Platforms such as ElevenLabs offer royalty-based voice licensing because without it, they would struggle to attract high-quality contributors willing to let their identity be replicated at scale. In these cases, intellectual equity is not a moral stance, but it is a market requirement.

Other actors can also step into the picture.

For example, labor movements and cooperative models are emerging as a parallel response to AI-driven value extraction. Across Europe and the US, data labor unions and cooperatives such as the European Data Union aim to give individuals collective bargaining power over how their data and training work are used, shifting compensation from isolated gigs to shared revenue and ownership.

In the Global South, where much of AI’s invisible labor resides, grassroots efforts like the Kenyan Data Labelers Association highlight how outsourced annotation and moderation work underpins advanced AI systems while remaining underpaid and unrecognized, pushing the conversation from wages toward profit-sharing, royalties, and IP claims.

TLDR: Who owns the future of intelligence of work?

As long as contributors are replaceable, anonymized, or distant, the system holds. As long as AI improvements appear incremental, the imbalance stays hidden.

But three forces are converging:

First, AI systems increasingly depend on expert judgment rather than raw data. The closer systems get to real decision-making, the harder it becomes to substitute the humans shaping them.

Second, contributors are becoming more aware of the value they are transferring. Creators, knowledge workers, and operators are starting to ask not just how AI helps them work faster, but who benefits from what they teach it.

Third, regulatory and contractual pressure is rising. From statutory remuneration proposals to AI-specific licensing clauses, the legal system is slowly catching up to what the technology already enables.

Together, these forces make the current model fragile.

By knitting human incentives into AI systems, we not only address issues of justice, but also potentially improve AI outcomes through sustained human engagement. In the long run, an AI ecosystem that rewards its human contributors is one that is likely to be more inclusive, trusted, and robust.

We’re only at the beginning, but these early frameworks and case studies provide a roadmap toward an AI future where value is shared – and that may be the key to unlocking AI’s full societal benefits.

Until next week!
Matteo

For most people, work isn’t designed - it’s just assigned.

How do we design work?

As new tools and intelligent systems reshape how work gets done, design matters more. Any successful organisation needs the capabilities, systems, and culture to design work.

Empowering successful teams should be the remit of HR.

It is envisioned that a successful HR function in 2030 will be smaller, less process-orientated and architects of a healthy work ecosystem.

As Amber Grewal recently wrote in her Eightfold essay ‘A New Recipe for HR Transformation,’ the ingredients for success have changed. The familiar layers of HR - people, process, and policy - are being reimagined around data, intelligence, and experience.

So how should we approach People Management?

Like motherhood, everyone has an opinion about HR - even when they have never done it. This article will be of interest to those working in HR, and also to those whose working lives are impacted by it.

Old Recipes, New Ingredients

A confession. I have designed, built, improved, tweaked, and dismantled more HR Operating Models than there are pastéis de nata in a Lisbon pastelaria.

There have been many flavours - from force-fitting HR generalists into business partner roles, installing self-service systems that nobody really used, and migrating HR into shiny shared-service centres from Glasgow to Mumbai - each iteration chasing efficiency, not always effectiveness.

Some went down a treat, others went a bit stale.

I have worked with many great teams and seen careers flourish, but have also seen untapped potential, and survived some toxic workplaces.

These experiences shaped a simple goal for me - to make work better, much better.

I believe HR has a big role to play in this, and here are some learnings I have shared recently at events in the Netherlands, London and Lisbon.

The first rule of HR Transformation is - Don’t mention HR Transformation

Nobody actually wants to be transformed.

After many organisational changes, most things will stay the same anyway. Use inclusive language, involve employees in this change from the beginning - they are closer to the action and often know what will actually work.

Beware Borrowed Cookbooks

The standard HR operating model was developed around the turn of the century and influenced from external factors happening at the time. This included the move to Account Managers in support functions, e.g. IT, Finance. ERP Systems. Shared Service Centres.

IT and Finance went first, and HR copied this model.

But if you bake with the wrong ingredients you will get mixed results.

Primarily because the implementations did not use organisational design techniques, from first principles and incorporating the business context.

Secondly, many were trying to fit a square peg into a round hole. A model that works in an IT company might not work in a factory.

I no longer show my slide, “here is the answer, now what is the question?” with respect to the implementation of rigid HR Operating Models - it ruffled certain academic feathers apparently.

Sometimes, we designed HR models that had no Business Partners, because they were not needed, with others we hired BPs from outside of HR to be commercial account managers. It depends on the business context and industry.

The external factors that once influenced the shape of HR in the last century, have now changed substantially, from SaaS to AI, and a workforce increasingly outside the payroll.

It’s time to throw out the old HR cookbook and rebuild the people function.

Evidence Over Technology

Why do we continue to use certain twentieth-century management practices?

One reason is that we continue to use the software licenses we inherited.

If there was a business case, we don’t know whether the process worked, because we didn’t measure the results.

Use evidence-based design to solve organisational problems - and avoid the worst mistakes.

Use the best available evidence to make decisions, this includes academic meta-studies (if you are lucky), asking those who have done it before (if they will speak to you), and your own stakeholders (much easier).

For more resources on evidence-based management use this excellent resource → CEBMa

Start with Better Questions

From the start of your journey – obsess about asking the right questions.

I sometimes hear from CHROs, “fascinating how India and Africa will change work this century, but my CEO wants to know, when are we implementing AI and can we cut 20% off the payroll in the next 3 years?”

So a hint,

“how do we implement AI?”

is the wrong question.

Instead ask about which capabilities your organisation needs to succeed, and how to create a system that will adapt successfully to changes.

So what’s a modern HR Leader to do in 2025?

Look to the Future using Scenarios

The next decade isn’t about digital transformation - it’s about organisational redesign, as industries restructure.

Teams will increasingly manage themselves, powered by smart tools and intelligent systems. Firms will get smaller, more fluid, mixing employees, contractors, and digital agents. HR’s role won’t vanish; it will orchestrate agility, culture, and capability across these ecosystems.

Read our article, “The Global Mega Trends Shaping the Future of Work”, download PDF for your team to read, and pick out some scenarios to test against your workforce strategy.

Focus on What Matters

Every HR leader is being sold the latest platform, dashboard, or productivity hack.
But the best-performing teams keep their eyes on purpose and capability, not shiny tools.

Start with organisational goals, design for how value is created, and help teams build the skills to deliver it.

Evaluate new technology — but only after you’ve answered: what problem are we solving?

Amber describes this shift as ‘deconstructing and rebundling’ HR capabilities — a recipe where technology, data, and design come together to elevate the employee experience.

Get stuck into conversations about the future of the organisations – the workforce strategy will quickly follow.

Think Work, Not Jobs

Work is being unbundled and reassembled in new ways.
Talent is no longer bound by geography, and tasks are no longer trapped inside jobs.
Intelligence is shared between humans and AI co-pilots; teams are networked rather than hierarchical.

Security, too, is shifting — from employment to skills and reputation.
HR’s mission now: engineer how work gets done - not just enforce process.

Trust Over Technology

Trust remains the true currency of change — for employees, customers, and leaders alike.

AI without trust is just automation.

Keep people first, stay curious, and remember: technology amplifies what already exists - it doesn’t fix what’s broken.

Empowering Teams to Design Work - From Recipes to Actions

Some HR operating models are on their eighth recipe iteration; others are still half-baked.

Any successful organisation needs the systems and culture that enables the process of designing work to respond well to the industrial context they are in.

Some suggested actions for Work Designers :-

• Start small, design boldly. Choose one workflow or team and redesign how work gets done using evidence and data - not hierarchy.

• Map capabilities, not jobs. Use talent intelligence tools to understand which skills drive performance and where the gaps lie.

• Build feedback loops. Treat every HR process as a prototype - test, learn, and refine.

• Empower teams. Train managers to co-design roles and workflows with their people, not for them.

• Measure what matters. Track capability growth, trust, and adaptability - not just headcount or time-to-hire.

Don’t just add AI to an old recipe and expect a better flavour.

The Cherry on Top

HR can play a key role to create systems that enable teams to design and manage work better, much better.

Those who get this will power ahead.

Challenge the 20th-century modes of management and reinvent them for your teams, your organisations and your industries.

Embrace the role of the work designer. Leaders must shape systems that flex as fast as the world changes.

Work design isn’t a project - it’s a craft.

HR’s next act isn’t about rolling out another transformation project. It’s about designing the systems that let people and technology work brilliantly together.

Your cinnamon-dusted futurist,

Andy

I want to write a letter full of optimism for 2026. However, my honest feeling is that this year will be hard for a lot of people.

Today I explore what shifting jobs, fragile tools, and volatile markets might mean for work this year.

In the week before Christmas I met with Matteo in Rome to plan the Work 3 themes we will focus on in 2026 which we are very excited about.

With some inspiration from the Eternal City, todays article is about building resilience.

Subscribe now

Jobs Change Shape but Salaries Are Still Worth Their Salt

At recent festive gatherings I have met people making a living in diverse industries and there seems to be anxiety about job security, much attributed to the rise of AI. For example, the movie special effects conference delegates downbeat after another round of job-losses.

What does the actual job data tell us?

US unemployment is at its highest level since September 2021, but not historically high at about 4.6%.

The LinkedIn Workforce Report shows that over the past two years, hiring rates across major economies have fallen sharply.

We are not seeing mass job losses, but job slowing and job thinning. There is a lot of variation across skills, industries and geographies.

That’s not job destruction. It’s job deconstruction.

This graph looks like a simple story: markets rising while job openings fall.

The data is correct — but correlation isn’t causation.

Markets move on expectations; jobs move on budgets and time.

As Derek Thompson explains, for the last few decades, job openings have mostly tracked the rise and fall of the stock market. Nothing like the current divergence between equity returns and job openings has ever occurred in the history of the Job Openings and Labor Turnover Survey (JOLTS).

In fact the reduction in hiring can be better explained by interest rate hikes than AI. Higher rates are meant to lead to less investment, less spending, and less economic activity, which altogether can have the effect of reducing hiring.

Jobs won’t vanish all at once; they might crumble, get patched, get rebuilt, and work will continue with different arrangements, contracts, and, payment terms.

I have envisaged what a world with less formal jobs might look like. However, in an uncertain environment, a regular salary is still worth its salt.

As jobs become less predictable, people are compensating by embedding AI tools directly into how they work and think.

Don’t Build Your Thinking on Sinking Sand

AI tools are no longer just productivity aids. For many people, systems like ChatGPT, Copilot, Claude, Gemini and others have become part of how they think: drafting ideas, learning new topics, practising languages, preparing difficult conversations, organising knowledge, and making sense of complexity.

The scale of adoption is striking. I think it is reasonable to estimate that around a billion people use LLMs each week.

At the same time, the economics behind these tools are still being worked out.

OpenAI, for example, is widely reported to be valued at $500 billion USD, with revenue roughly $12 – $13 billion in 2025. Despite its popular tools it remains unprofitable, with losses likely in the $7 – $9 billion range as it invests heavily in compute, research and growth.

Similar patterns exist across the AI sector: extraordinary capital investment, rapid user growth — and a long road to durable profitability.

None of this means the tools are going away. But it does mean that the way we access and pay for them are likely to change.

Some plausible futures include:

• Repricing: free or low-cost tiers becoming more limited as providers push toward sustainability

• Consolidation: tools merging, being shut down, or absorbed into larger platforms

• Enterprise capture: the best features moving behind corporate paywalls

• Model churn: frequent shifts between models, behaviours, and interfaces as competition accelerates

Many people are now building prompt libraries, personal knowledge bases, workflows, and even emotional habits inside specific AI tools. If a product changes, reprices, or disappears, that work — and the thinking embedded in it — can be hard to retrieve.

What individuals can do now?

• Keep your thinking portable. Store core ideas, notes, and frameworks outside any single AI platform.

• Build a small portfolio of tools, not a dependency on one.

• Treat prompts as patterns, capture reusable structures and reasoning approaches, not just verbatim text.

• Use AI to amplify thinking, not replace it - If a tool vanished tomorrow, could you still explain your reasoning?

The winners in the next phase won’t be the people with the most AI subscriptions.

They’ll be the people with robust knowledge systems, adaptable workflows, and the ability to think clearly when the tools change.

Use AI deeply. Just don’t build your precious knowledge base on sinking sand.

Tools change. Platforms come and go. What matters is understanding the economic cycle underneath them — because that cycle has always shaped work more than any single technology.

When Markets Catch a Cold, Work Gets The Flu

AI has been a powerful tailwind for financial markets. Nowhere is that more visible than the Magnificent Seven. No this is not Rome’s seven hills, but what investors call the seven US technology companies that now dominate the skyline of global markets:-

Apple, Microsoft, Alphabet (Google), Amazon, Nvidia, Meta Platforms,
Tesla

Together, these firms account for around one-third of the entire S&P 500, and well over half of the Nasdaq 100. From 2019 to 2025, the Magnificent Seven delivered a combined total return of over 1000%, compared with around 180–200% for the S&P 500 over the same period.

These companies make genuinely useful products, platforms, and infrastructure. Markets don’t price what companies did last year — they price what investors believe they will deliver in the next few years.

For example, Nvidia has recently carried a market valuation of around $4.6 trillion - more than the entire annual GDP of Germany. In its most recent financial year, Nvidia reported roughly $73 billion in net profit. At that pace, it would take around 60 years of profits at today’s level to match its current valuation. To put it another way, the market is assuming hundreds of billions of dollars in cumulative profits over the next decade.

The valuations might be realistic, but rely on near-perfect execution in an industry that will be impacted by competition, disruption, consumer demand, wars, sovereign debt crisis, recessions, and geo-political chess moves.

For now, I think we can safely say there is a chance that there will be a significant re-rating.

For history buffs, or those who still remember the sting, in the Dot Com crash in 2000/2, the Nasdaq fell 78% peak to trough, In the Global Financial Crisis of 2008/9 – the S&P fell 57%.

So you might be reading this, saying, well I don’t own any of the Magnificent 7 Stocks, so what’s the problem?

The problem for the rest of the economy is contagion. Typically when there is a market crash, all sectors sink. It will not discriminate between good and bad companies in the short term.

So what might be the impact on work and jobs?

History suggests that when the stock markets crash, work follows but on a delay.

After the dot-com crash, markets fell quickly, but hiring froze for years. During the Global Financial Crisis, unemployment continued rising well after share prices had bottomed out. And in 2020, while markets rebounded rapidly, more than 22 million US jobs were lost, with many workers taking years to regain stability.

One lesson is consistent: US market cycles shape hiring, investment, and job security far more than the latest AI product launch or marketing campaign. Macro-economics, markets, geo-politics move the labour market over the years. You can read more about the Nine Global Mega Trends Really Shaping The Future of Work here, and download our 30 page PDF.

When markets tighten, organisations pause, investments stall, projects stop, hiring freezes.

The real question isn’t what the market does next — but how can we prepare?

Strong Foundations for Uncertain Times

Just maybe….the markets are calm, the $50 trillion debt is ignored, the labour market generates wealth and security, we enjoy abundant innovation with little extra cost.

But, as 2026 emerges, this is a moment to think about the risks mentioned today calmly.

Resilience isn’t about predicting the future. It’s about staying regulated, connected and adaptable as life unfolds.

Here are some thoughts on building up your resilience :-

Stabilise your habits, processes, networks and health.

Reinforce your existing habits and processes to stay mentally and physically healthy. Focus energy on what you can control - your skills, and your habits. Rather than things you can’t control e.g. the job market or stock prices. Nurture your existing support network.

Create more options for your income, your ideas, and your learning.

Identify and reduce single-point failures – one employer, one skill, one revenue stream. If your work network is mostly people who work for your current employer. Cast the net more widely on ideas and people you can learn from. Keep Learning – people, agents, workflows – not just courses. Think about second and third income streams, more freelance work – from fractional execs, to design, coding, writing and task outsourcing. Start to experiment – it takes courage, but small experiments might lead to big reinventions later.

In 2026 Rome still stands, as it has done for over 2000 years. Families still enjoy the Christmas Market at Piazza Navona. There is resilience built into the system.

Rome has had its weak spots and cracks, but that is how the light comes in.

And the world of work will move forward in a similar way, unevenly, imperfectly, and very much in public.

Wishing you a steady year ahead.

Andy

In past tech revolutions, access to tools and information was limited — only a few could afford early computers. Today, AI arrived essentially for free and for everyone, even before we fully understood what to do with it. The result is a much narrower learning divide—but also a new kind of pressure, where even early adopters don’t feel ahead. LinkedIn data reflects this tension: 41% of professionals feel overwhelmed by how quickly they’re expected to ‘get AI’, and posts about AI overwhelm are up 82% year over year.

This is where our earlier work connects: in the first part of this series, AI feels like a second job, we explored what many professionals were quietly experiencing: anxiety, pressure, and the sense that AI has added a whole new layer of work. In Learning in the Age of AI we focused on the antidote that works in real life: learning together and building confidence with other people, because that is how we are truly wired to progress.

Today, we shift from the emotional and social side of adaptation to the structural one. We look at:

- How AI is reshaping careers: what current adoption looks like, what the data shows about the roles changing first

- What it means when job titles, the labels that defined work for decades, begin to lose relevance.

We’ve Been Here Before, But This Time Feels Different

Each technological shift has arrived with predictions of disruption, job loss, and new forms of productivity: the ‘web’ has been dying since 1997, - and yet each time, the story unfolded slower and more unevenly than predicted. As Reinhart and Rogoff argue in the book This Time Is Different, we routinely believe that a moment is unprecedented, even when the underlying pattern repeats. The real story is not that something fundamentally different is happening, it’s about the familiar mismatch between expectations and actual readiness.

For this reason, I spoke with Catherine Fisher, Career Expert and VP of Communications at LinkedIn, to try and get some insights from her vantage point - right at the intersection of companies and employees. During our conversation, she captured this clearly and simply: “First of all, employees are still learning AI, but companies already expect them to be fluent. Then, the disruption is creeping in at the task level, while expectations rise at the role level. Workers can feel the shift before they fully understand it.”

This is the gap that makes this moment feel faster, more visible, and more personal than previous ones and that sets the stage for what we should really focus on: how people learn, experiment, and adapt in small steps, long before job titles start to shift (remember what we talked about extensively The Impact of AI on Industries: disruption starts at the industry level, then organizational level, before it goes to the individual level).

Skills before Jobs: What’s changing, and where value is moving

Just as scientists must study matter at the atomic level to understand real change, we need to examine work at the level of skills, not job titles. Job titles are lagging indicators - the work inside them shifts long before the label does. Skills show us where change is actually happening.

There is lots of data that predicts this shift: according to LinkedIn data, 85% of U.S. professionals could see at least 25% of their skills change as AI advances. Yet the fastest growing skills are not the ones you might expect, they are human. Of the 10 rising most quickly only one is technical (LLM Development and Application, possibly joined by Process Optimization, depending on how you classify it). That is the real story underneath the numbers: the skills rising fastest are not technical certifications or model-specific capabilities; they are the human skills that become more valuable as AI reshapes work (don’t forget, we are still looking at how to use it as a tool, but it will eventually change the whole of work design itself). Judgment, coordination, originality, and communication will become the scarce inputs, because they are the skills that hold teams together, move projects forward, and turn AI output into real outcomes while keeping it in check. They are also the skills that differentiate people in a market where everyone has access to the same tools.

Conflict Mitigation - Already very valuable, but as AI speeds up decision making and shortens project cycles, misunderstandings can multiply faster. Teams collaborate across functions, AI agents, and shifting roles means that more decisions happen in ambiguous contexts. This will increase friction and people who can calm tensions, clarify expectations, and prevent escalation will become organizational stabilizers.

AI Literacy - On this, Catherine Fisher from Linkedin stressed how “It's one of the top skills that companies are looking for, and certifications help you show it, but what really matters are concrete examples of how you have used” - which ties in strongly with everything we talked about proof of work.

Adaptability - When tools evolve often and industries are more dynamic, adaptability becomes a core market survival skill. It is difficult to measure, and it lives as much in emotional regulation as in technical ability but at its heart, it’s about responding constructively to change, staying grounded, and focusing on the parts of the work that remain under your control.

Process Optimization - Once AI handles individual tasks, the real leverage comes from improving the system around them. Companies need people who can redesign workflows, spot inefficiencies, and integrate tools in ways that multiply output.

Innovative Thinking - AI inevitably makes MOATs incredibly hard to have for companies, and at the same time, for individuals themselves (always remember, that both inside the company and in the job market, you are a ‘product’ that needs to be evaluated by supply and demand). This means, that thinking outside of the box becomes essential. Anyone can now do basic (maybe not professional) design, edit, or prototype with minimal expertise and this pushes the value of originality far higher. With execution becoming cheap, fresh thinking and novel combinations of ideas become the scarce advantage.

Public Speaking - As AI automates “cheap” written communication, spoken communication becomes a differentiator. Teams and leaders rely on people who can explain ideas clearly, persuade groups, and make decisions in real time. An AI can generate a script or deck, but only a person can deliver it with conviction and move a room.

Solution Based Selling - Buyers now arrive informed, with AI powered comparisons and research already done. What they need is interpretation from professionals who can frame problems, translate complexity, and guide clients through options become trusted advisors. The role shifts from pitching features to helping customers make confident decisions in a very competitive world.

Customer Engagement and Support - If AI handles routine or repetitive questions, that leaves humans with the complicated, emotional, or high stakes cases. Empathy, clarity, and accountability become premium skills. Customers escalate beyond chatbots because they want someone who can truly understand the situation and resolve exceptions with judgment.

Stakeholder Management - AI powered work touches more teams than ever before, making alignment harder. Projects move faster, involve more roles, and create more dependencies. People who can maintain relationships, negotiate needs, and keep groups rowing in the same direction become essential. They are the coordination layer that makes complex work possible.

These skills are not something you acquire through courses alone: they emerge through experience, through practice, and through the kind of human interaction that shapes judgment. When I asked Catherine Fisher about this, she made the point in a way that stayed with me. “As people start using AI in their daily routines… it becomes second nature, but what rises in importance is our ability to interpret, explain, and guide - not just use the tools, but make sense of them for others.”

Once we understand which skills are rising in value, the next step is to look at how people combine them in ways that create new advantages in an AI driven career world. This is where Thomas Pueyo’s idea of skill stacking becomes particularly relevant: the notion that developing two or more distinct skills, and blending them, can create a kind of edge that is very hard to replicate. Every skill has diminishing returns, and there is a point where going from great to world class requires exponentially more effort, while adding a new complementary skill often produces far more leverage.

You can be in the top 10 percent of F1 engineers, for example, or the top 1 percent of people leaders who understand F1 engineering teams - and the second profile, while less obvious, is far rarer and often far more valuable. Many of the most impactful innovators operate this way. They don’t win by burrowing ever deeper into a single discipline, but by combining fields into a lens that allows them to see patterns, solve problems, and produce work that specialist paths alone could not have produced.

In an AI powered workplace, where the boundaries of roles are still being drawn, skill stacking becomes one of the most practical ways to shape the careers that are emerging, not the ones we are leaving behind.

Right now, Experimentation means everything

AI adoption rates are still very timid in business functions versus individual use. Most companies are in the experimentation and pilot phases, something LinkedIn’s data confirms:

- Two thirds of U.S. workers say they are still in “experimentation mode,” using AI only for simple tasks

- Almost half say they are not using AI anywhere near its full capability, and one third say they rarely or never use it at work.

This is important because it reframes the fear: the pace of change feels fast, but the real transformation will unfold over years, not months. Expectations have surged, but practical capability is only beginning to develop. Not to say this won’t be growing or accelerating, but my argument is that even though the pace of changes is (or feels) faster, it will take several years for AI to be massively adopted, functional, and have the scary outcomes that people are fearing. Side-note: right now, we have seen how most AI-pocalypse layoff announcements have come as a cover up for COVID hiring sprees, and general macro-economic weakness.

What does this means for careers? It means that people should not be worried about their job 5, 10 years from now. We have already been living (especially as Millennials, and GenZ) in an era of disillusionment and detachment from long-term and linear careers (wrote extensively about in Planning for Non-Linear Careers). AI is accelerating a shift that was already underway.

This is why the smartest thing we can focus on right now is not predicting where jobs will be in a decade, but building better habits of small, consistent learning.

James Clear’s well known concept, “The Power of Tiny Gains” shows how 1% improvements compound into something transformational over a year, and the same logic applies to AI adoption. Tiny experiments lower the anxiety, build confidence, and turn ambiguity into action. The moment you try something, the fear starts to shrink. And as you practice in small steps, you also become more visible to your team and your company. You begin to close the gap between the expectations companies place on workers and the reality of how people are actually learning these tools today.

Catherine Fisher called this the “Life Curriculum”, emphasizing the importance about building a cadence where each week (or even each day) you try one new thing, refine it, and fold it into your practice. The only addition I would make is that companies must adapt their expectations too: instead of imagining instant transformation or treating AI as a magic switch, they should value and evaluate the frequency and consistency of these small experiments, because that is where capability compounds.

That brings us to Learning Velocity: not just the willingness to learn, but the pace at which someone can absorb something new, apply it, and move forward again. Unlike formal credentials, this is a muscle you build through repetition which happens through four steps:

Consume: Ingest high-quality information and examples.

Analyze: Reflect on what you’ve learned, test your assumptions.

Create: Apply it by build something, try a workflow, solve a problem.

Teach: Share it with someone else - teaching forces clarity and deepens expertise.

This loop turns scattered experimentation into something compounding. It creates visible evidence of learning - something you can point to in interviews, performance reviews, or even within your team. And it replaces the outdated idea of mastery with something far more realistic: adaptability, curiosity, and the ability to build new skills faster than the environment changes.

(This is a long post, if you prefer to read in the browser click here)

I also think that people who experiment the most will be the ones who end up shaping the new roles. At this stage, even companies do not fully know what the next generation of jobs will look like. They are still figuring out the tools, the workflows, and the impact on teams. If roles are being rewritten task by task, then the people who test the tools, document what works, and share what they learn become the ones who define what the new role actually becomes. In a moment when the job market is fluid, the experimenters can become the architects. This can become a ‘mission’ for individuals or teams, something that also communicates ambition and eagerness to innovate and evolve inside of the company.

Last, but not least - experimentation also creates the conditions for career pivots. Across job functions like legal, engineering, accounting, and marketing, many are moving across fields that seem unrelated at first glance:

- Lawyers are moving into research and operations

- Engineers are moving into education and community roles

- Marketers are moving into project management and even real estate

- Accountants are shifting into education and operations

These shifts come from people understanding their strengths and goals, and then testing possibilities through small experiments, both in their jobs and in the work they explore also in ‘side quests’.

Rise of Side Quests, Solopreneurship, Portfolio Careers

Everyone I know who builds something in their free time, or uses their spare time to explore a curiosity, ends up much further ahead in one or more of the skills we discussed earlier. Side projects become labs for experimentation because they sharpen communication, deepen expertise, stretch creativity, and build confidence in ways that formal training rarely does.

They learn faster and discover new strengths because they step outside the comfort of their job description and the limits a role quietly imposes. As people take more control of their learning and experimentation, something important happenes: they feel more control over their entire career:

LinkedIn’s Workforce Confidence data reveals that only 45% of Americans feel in control of their career, but that number jumps to 65% among the self employed.

From side quests, the leap to solopreneurship seems much smaller. And with AI being such a powerful tool to prototype and experiment, we are seeing unprecedented growth in bootstrapped (no venture capital involved) startups being founded.

According to MBO Partners’ 2025 report, nearly 73 million Americans, about 38% of the U.S. workforce, now freelance in some form. At the same time, 78% of companies say they plan to rely on freelance or contract talent rather than hire full-time staff. Add to this shift the fact that more than 22% of U.S. employees now work partly remotely, and the stage is clearly set for a rise in portfolio careers. Flexibility is, or should, become the design principle of work.

The next step from here is already emerging: as AI lowers the cost of coordination and execution, work begins to break apart into smaller, more personalized units. Instead of every responsibility being absorbed into a single job description, more tasks are becoming project-based or bounty-like, where individuals contribute to specific problems rather than to abstract roles. Companies benefit from more precise access to talent, while individuals participate in work that matches their skills and availability more closely. Work can become something you assemble, not something you inherit.

The concept of portfolio careers fit where the reality that careers are becoming fluid, modular and shaped by a mix of income streams, side projects and learning loops. But beware: this is not about doing more work (that is the concept of Overemployment). Like we said in skill stakcing it is about combining multiple capabilities, each one opens the door to a different type of project or collaboration.

In this sense, AI does not push people into ever narrower specialisations. It pushes them into becoming adaptive generalists with spikes of depth (David Epstein has a great book on this called Range). It rewards people who can learn quickly, combine fields, translate between domains and move fluidly where the work requires them. As work fragments, the worker becomes the integrator and as the market tilts toward flexibility and optionality, portfolio careers become less a fringe choice and more the default operating system for how modern careers evolve.

Skilling me softly - What happens when Job Titles lose meaning

I think the broader picture is starting to emerge.

The Latin root of the word “career” means journey, yet for most of the twentieth century that journey followed a narrow, predictable path: roughly 45 years of linear progression, a dozen employers, and a clear endpoint. Today careers resemble a 3D lattice made of learning, earning, caring, moving, pausing and restarting. Some work is paid, some not. Some defines our identity, some simply carries us to the next chapter. Job titles lose meaning in this environment. What matters is the composition of skills and ability to demonstrate them in motion.

This is why Google DeepMind’s Demis Hassabis is right when he says that “learning how to learn” becomes the most essential skill for any professional. As tools evolve faster, meta skills become the real differentiators.

This is also why making predictions is a weak response to uncertainties. We need to think like investors, not a forecasters. We won't know which jobs will exist in ten years, but we can know how to compound skills, habits and visible work in ways that keep creating opportunity. We need to ignore advice built for the last generation.

For companies, this shift requires clarity at the foundational (and executive) level. Every organisation should be doing four things exceptionally well: mapping their skills, understanding how work actually flows through their networks, designing learning programs that build capability rather than credentials, and setting goals that help people see how their growth connects to the company’s mission. These are not “future of work” initiatives. They are simply the infrastructure of a redesigned modern workforce.

For individuals: build range, make your work visible, strengthen the human skills that are rising in value, experiment often enough that you learn faster than your environment changes. Invest in relationships because careers are built inside networks, not job descriptions. And above all, treat learning as a lifelong operating rhythm, not a course you take when everything feels stable again.

Whether or not we ever see the mythical “one person billion dollar business” is beside the point. What matters is that the job market we know today will age quickly, just as the pre-gig economy world looks unrecognisable in hindsight. And whatever comes next will be shaped not only by AI but also by geopolitics, macroeconomic cycles and climate realities that may prove even more disruptive.

In the meantime, we focus on what is in our control. Catherine Fisher said something in our conversation that I keep returning to: “It is not always about productivity. It is about how the work becomes better.” If we orient ourselves toward making the work better, the productivity and the outcomes tend to follow naturally.

That may be the simplest and most durable career strategy of all.