The drag was the organisation, and the supporting operating model. It was roles that had not caught up to what the tools could do. It also was decision rights nobody had redrawn. Or data that was still unavailable, or too immature to use. At the end it was hundred small habits nobody had ever written down, that everyone followed and no one could name.

AI did exactly what we asked of it, but the company was its ceiling.

That was the moment something flipped for me. It became clear that AI is not something you deploy, it is something that inherits the organisation sitting underneath it, like a tenant that moves into the house you already built, with all its wiring and all its cracks.

AI scales your organisation, for better and for worse

AI inherits your workflows, your data, your culture, your org chart, and every unspoken rule underneath it: the undocumented work, the politics, the definitions nobody ever agreed on out loud.

That one line is the whole argument. AI does not create organisational intelligence, it scales organisational reality. It is why two companies buy the same model, run the same pilot, and get results that look nothing alike. Coherent workflows, and AI compounds the coherence. Muddy decision rights, and AI just pours faster output into the mud, so the confusion arrives sooner and in far greater volume.

There is one common pattern I observed: AI succeeds at the level you apply it, and stops at the level above it.

if you gave a team good tools and good training, the team went faster, and you could measure it. But in addition, what you hear instead, more and more, is overwhelm. Nothing went wrong with the AI. The faster work traveled up to an approval gate, a handoff, a decision right that was designed back when judgment was scarce, expensive, and always human, and it stopped there. Not because the adoption failed. Because it worked, and then hit a ceiling. Individuals get faster. The organisation does not.

Here is my own version, still unsolved. My product team went from shipping and innovating in months to doing it in weeks. Real, measurable acceleration. My Go to market teams have not reached that pace, and what they are calling out is not a lack of effort. It is overwhelm, competing priorities, uncertainty about what to stop doing, unclear decision rights. The technology is moving faster than the operating model.

That is where leadership matters most. Not in deploying more AI, but in applying judgment: redefining priorities, redesigning work, and creating the conditions for people to move with the technology rather than drown in it. So the real question is this. Are you running your AI work recomposition at the right level? Do it team by team and you get exactly this: one function sprinting, the next lagging, a company faster in pockets and no faster overall. AI transformation is not the sum of functional optimisations. It is an operating model redesign, and that is a CEO and executive team responsibility.

Where it stalls, and why ownership is the unlock

The clearest proof I have seen was inside a smart mobility company I advise. They ran the same AI program on two engineering teams: same partner, same phases, same materials. Both got faster inside their own walls. And that was the ceiling.

The prize that would have moved the business, the overlap between the two teams, was never captured, because it lived one level above where the program was aimed, and no team can reach up and redraw the org from inside its own boundary. The board’s expected results could not even be articulated. That is what finally pulled the work up to the leadership table.

The teams proved the method worked. What no team could do was decide what the organisation should become. That decision has an owner, and it is not a tool and not a team. It is the leadership. The difficulty is not knowing what to do. It is giving leaders the mandate and the support to make the real decision: what an AI-empowered target organization should actually look like for their company.

What Deutsche Bank understood

Deutsche Bank recently put around 200 of its most senior bankers through mandatory AI training. The easy read is upskilling. It misses the point. Those 200 are not the ones who will be in the tools all day. What training does at that level is not teach a skill. It removes an excuse. Afterward a senior leader can no longer say they did not understand what AI changes, and you cannot sponsor what you cannot see.

But seeing is not doing. Training gives leaders the mindset. It does not do the recomposition. Left there, it just moves the real work further down the road, one more thing understood and still not driven. Putting leaders in the driving seat is the point, doing the recomposition with their teams, at the level where product and marketing have to be redesigned together. Awareness that never becomes that is just a more expensive way to stall.

Where I am probably wrong

The theory does not hold everywhere.

• Greenfield: build something genuinely new and there is nothing to inherit, which is why “look how fast that startup moved” is a trap, they had nothing to transform.

• Shadow adoption: people routing around bad processes with private AI are escaping the inheritance, not scaling it, but only until it becomes official, and then it inherits everything it was dodging.

• Substitution: now and then AI removes a step entirely and the process disappears, politics and all. Rare, but real.

The executive conversation that matters most

If any of this sounds like your company, do not ask which tools you are rolling out. Ask your leadership team, or the CEO of the company you backed, to present how AI changes their business model and operating model, and the three or four areas where they are concentrating the redesign, each with an owner, a target, and a way to know it moved.

The presentation is the diagnostic. Tools and pilots means the work is still stuck at the team level. A clear picture of where the business is going means a leadership team that has taken ownership. Most cannot do it yet. That is the first honest map of the gap.

Inside each area, hand them the sharper tool. Take the highest-impact workflow and force every step into one of four boxes.

• What stays fully human, and will not be touched.

• What is assisted: the human decides, the AI supports.

• What is automated, with a human in the loop.

• What is fully automated.

Most companies never do this, so AI inherits an allocation of judgment nobody examined and nobody agreed to. Then ask one more thing, the question nobody asks because everyone is planning for AI to fail: what happens when it works? When the team is genuinely faster, where does that output land, and is that place ready for it? That place is your real constraint, and it is almost never the technology.

The tools were never the work

The companies that win the next decade will not be the ones with the best models. Everyone will have those. They will be the ones whose leaders built the capacity to redefine how work happens inside their organisation. I am making sure mine is one of them, and helping others who want the same. Tell me what you are seeing. We are all learning as we do this.

After standing up our AI transformation teams, each team came out with a document setting out its strategic goals, the workflows we were targeting with AI, the AI capability infrastructure it would need, and what the organization looks like once AI supports part of the execution. It also named where we would need judgment, what would never be done with AI, what has to be reviewed, and what could be automated.

From there, each team started prioritizing their first few workflows.
We started evaluating around impact: biggest margin, biggest cost line. Then we our AI technology leader pointed them at the places where we could get everyone using the tools, because he did not believe a workforce that had never worked differently could be asked to transform something that mattered without seeing first how they could do it.

We achieved adoption, and once enough people were building, a different constraint appeared. We were no longer limited by whether people could use AI. We were limited by how quickly we could decide what to build next.

What surprised me was not what AI could build. It was what the organization could no longer coordinate. Execution accelerated because every team could suddenly produce far more than before. Coordination did not, because we were still deciding through the same structures we had used before any of this started.

We are now moving faster than we ever have, and we got there by starting slower than anyone wanted to. In a way the team came with us on, which is the only kind that holds.

Here is the field story, and what we learned from it.

How it showed up

People across the company were doing genuinely good work with these tools, work that was making us better. It was simply time to move to the next step in our maturity.

The signal came from two directions at once:

The first came from the builders. In a follow-up on wave one, the message was that vibe coding had gotten us this far, and we were missing something to really tackle the biggest things we knew we could do. They were willing to go further, because they had seen the capability and they felt confident moving forward.

The second came from my AI Operations leader. Everybody is coming to us, he said. There are people across the organization genuinely taking ownership. But we have not set up the governance yet, or the standing forum where the AI team sits down with the business owner. Because the things with the biggest impact need the right people at the table, with the right knowledge and the right ability to decide.

They were one problem. One team could see the ceiling and was hungry for more. The other could see why the ceiling was there.

Three stages, and you cannot skip one

Looking back, what we went through was not a compromise followed by a correction. It was a sequence, and I think it is the sequence.

1. Stage one. The human at the center, AI as the tool. People still do the work. AI assists. What you are buying is not productivity. It is belief. So the metrics that matter here are delivery of the first prioritized workflow redesign and the volume of self-driven requests coming in, not business impact. It is also why every leader who skips this stage ends up with an expensive tool nobody trusts. We did it with the direction of travel defined by each team rather than by me. They owned it, which is why it held.

2. Stage two. AI takes the execution seat. The building stops being the constraint. A prototype that used to take a quarter takes an afternoon. And the person who used to do that work moves up rather than out, into deciding what should be built and in what order. Judgment becomes the scarce input.

3. Stage three. Infrastructure scales it. This is where we are now, and it is the least glamorous stage by a wide margin. Reliable data. Governance. Operational discipline. Clear ownership.

The lesson that took me longest to accept sits between stages two and three. Vibe coding gets you much further than most people expect and nowhere near as far as you ultimately need. Great builders create extraordinary things quickly. Then those things need data they cannot reach, systems they cannot write to, and someone senior enough to decide they should exist at all.

The constraint had become organizational

Our constraint had become how quickly the organization could make good decisions about what to build, how to prioritize it, and how to redesign the work around it.

That gap does not announce itself as a failure, every project works but you end up with a portfolio of things that function individually and add up to nothing, and the honest diagnosis is not that AI underdelivered, it is that we became capable of producing more than we could coherently direct.

What that looked like on the ground

The first team to tell me something had changed was not Engineering. It was our Creative team.

They came to me frustrated. “Presentations are being generated everywhere. They’re not aligned with our brand. Nobody is coming to us anymore.”

At first it sounded like a branding problem, it was not, it was an organizational problem.

Creative teams are typically built around review. Work comes in, they check it, they protect the brand before it goes out. That model works when content is made by a relatively small number of people and arrives at a manageable pace.

AI changed that almost overnight. Suddenly people across the company could produce presentations, campaigns and content themselves. The volume increased dramatically. The review process had not changed. The queue had simply become impossible.

That is what execution outrunning coordination actually looks like: the operating model just no longer matches the pace of execution.

I expected the Creative team to ask for more control, more reviews, more approvals, more people checking more work but they asked for none of that.

They took ownership of the problem and rethought their work instead. They built AI skills that carried our tone of voice. They created presentation workflows that reflected our design standards. They moved the judgment that used to live in review meetings upstream, into the system where the work is actually made.

Nobody instructed them to do this. Nobody rewrote their job description. They saw that the old model no longer worked and they redesigned it themselves.

Looking back, I do not think I could have mandated that outcome. What I could do was build an environment where the people closest to a standard had both the authority and the tools to improve it, and then resist the temptation to solve it for them. Giving Agency.

While we are already seeing productivity improve. Quality is the one I actually care about, and that is a question of maturity rather than effort.

Because maturity is not about reviewing more work. It is about moving judgment to where the work is created. You crawl, then you walk, then you run, and the mistake is assuming you can start at a run because the tools are fast.

Review does not disappear. It moves. The people who own a standard can no longer sit at the end of the process checking outputs. They have to build themselves into the system that produces them.

That, more than anything else, is what convinced me we were not simply deploying AI anymore, we were redesigning the organization around it.

How we reset for wave two

Our cluster leads are leaders on the executive team. Because if the bottleneck is organizational judgment rather than technical delivery, the people running the clusters need to be in the room where the company decides things.

Then we separated three accountabilities that most organizations leave tangled.

• Our Data & Decision Intelligence leader is accountable for ensuring the right data exists, that learning loops are built into the system, and that the organization continuously improves from evidence.

• Our AI Operations leader, who calls his organization the AI Factory, owns the engineering platform, operational excellence, governance, and compliance that let AI scale safely.

• And the business leaders own the outcomes. They decide what their cluster works on and in what order. Not the technologists. Not me.

But the design choice that matters most is that neither the Decision Intelligence team nor the AI Factory works in isolation. Their people are embedded inside every business cluster.

Because no central AI team, however talented, understands the workflow as well as the people doing the work every day. Technology expertise has to meet domain expertise where decisions are actually made.

So ownership is shared and accountability is not. Business leaders own outcomes. Decision Intelligence owns data and learning. The AI Factory owns the platform. The embedded teams connect all three.

Why we separated three responsibilities

I did not arrive at three from a theory. I arrived at it from a mess. But having built it, I think the shape is forced rather than chosen.

Something has to own whether the organization is learning. Something has to own whether the platform can carry what gets built. Something has to own whether any of it changed a business result.

They cannot be the same person, because they optimize for different things. Business leaders optimize for outcomes. AI Operations optimizes for reliability and scale. Decision Intelligence optimizes for learning. Ask one organization to maximize all three and one of them quietly loses.

Learning goes first, almost every time, because it is the only one of the three without a deadline attached.

Those capabilities can report anywhere. Different companies will place them differently, and the titles will not match ours. They cannot disappear.

That is the difference between deploying AI and redesigning an organization to work with it.

This is starting to show elsewhere. JPMorgan reset its AI leadership this month, moving the mandate a level down as delivery shifts into the businesses and functions. That is the right direction and it is the one we took. The question, there and everywhere, is whether ownership stayed whole when the people moved.

Expertise can be distributed but accountability cannot.

Three questions for any leader who is not seeing a return on AI

If coordination is the constraint now, these are the three I would ask first.

Where does prioritization actually sit? Not who approves the roadmap. Who decides the order when two clusters want the same engineer. If the answer is a technologist, you have a coordination problem you have not named yet.

How many of your green metrics measure capacity? If adoption, usage, and enablement are healthy and the outcome numbers are flat, you are measuring the thing that stopped being scarce.

Ask one embedded person who they escalate to. If they hesitate, you have proximity without accountability, and that produces goodwill and very little else.

Pick one. Ask it this week. You will learn more in that conversation than in your next quarterly review.

What I am committing to

I no longer think AI transformation fails because organizations underestimate the technology. It fails because execution accelerates long before coordination does. AI raises what a company can produce. It does not raise what a company can decide, own, or govern.

That part you have to build, and almost nobody starts building it until they can feel the gap.

We can feel ours, and the next phase is the harder one. The problems that matter most need real engineering underneath them, not a faster prototype.

So here is what I am holding myself to. No adoption numbers in front of my leadership team this wave. The only number I bring is what changed for a customer, and how quickly it got there. Weeks, not months. I would rather be six months into that than another year into something everyone uses and nobody can point to.

If you are somewhere in the middle of this yourself, I would like to know which one moved first in your organization. The building, or the deciding.

But prompting is only the surface. With the rise of agents, a deeper skill is suddenly the one that matters: how to collaborate with AI on judgment. This is the first piece in a series about that skill, and it starts with the moment I first ran into it, before I had any name for it.

In 2016, at GlobalEnglish, we did something I was proud of. We had twenty years of digitized speaking tests, each one scored by a human teacher. Two decades of judgment, captured. So we trained a model on it and built an automated speaking assessment.

The results were better than I expected. Not just faster, fairer. The machine’s scores showed less bias than our human teachers had, across most of the variation we could measure. We had taken the best of twenty years of human judgment and smoothed out some of the unevenness that comes with tired people scoring at the end of a long day.

We tested it again. For weeks we had our own trainers checking the output. Then we put it into production, in front of our whole audience, and a different bias surfaced.

Test-takers who were not Indian were getting systematically weaker assessments. Not because the model had anything against them. Because our twenty years of recordings held a very large number of Indian speakers, and the model had learned the sound of proficiency from that history. It heard a different accent and quietly marked it down.

We paused. We rebalanced the training, deliberately increasing non-Indian samples, and corrected it.

But here is the line I have not been able to let go of in the years since. We would not have seen it until we tested it on everyone. The bias was invisible in the lab. It only appeared when the whole world spoke into it.

How AI decides: 101 for leaders

Here is where the skill begins. As you let more decisions made by AI, it asks you to understand three things: how AI decides, how people decide, and how you decide. This piece is about the first. And the surprise inside it is that AI decides like us, including the parts of us we never meant to pass on.

An important thing to understand is how AI decides, because it is not how we decide and yet it ends up in the same place. A model does not reason from principles. It finds patterns in the data it was given and optimizes toward whatever it was told to aim at. It has no concept of accent, or fairness, or proficiency, only correlations and a target.

Think of how it actually learns. You show it hundreds of thousands of examples, each one labeled with the answer a human gave, this recording scored high, this one low. The model’s only job is to find whatever in the audio predicts that score, and to get better at predicting it each time it’s wrong. Nobody tells it what to listen for. It discovers that on its own, by finding the threads that run through the highest-scored examples. If most of those examples happen to be Indian-accented English, then “sounds like the recordings that scored well” quietly becomes part of what it hears as proficiency. It was never instructed to care about accent. It found accent, because accent was there in the pattern, and it cannot tell the difference between a signal that matters and a signal that just happens to be common.

That last part is the whole problem. A model cannot distinguish a real cause from a coincidence in its data. It only knows what went together. So when my test learned that the sound of an Indian accent went with high scores, it did not choose to favor that sound. It had no other way to work. The bias was not a malfunction. It was the system doing precisely what it was built to do, faithfully, at scale, on a foundation we never examined.

Three things to remember about how AI decides:

• It learns from labeled examples. The human answers are the lesson. Whatever judgment is in your history becomes the judgment it copies.

• It finds the predictors itself. Nobody specifies what to look for, which is exactly why you can’t anticipate what it will latch onto.

• It can’t tell cause from coincidence. It only knows what went together. This is why proxies appear, zip code standing in for race, and why “just remove the biased variable” never fully works.

Mirroring is the easy case. Substitution is the hard one.

Read my speaking test the easy way and it says: AI is biased, be careful. Read it the harder way and it says something far more useful.

The first bias, the one we removed, was the human one. Twenty years of teachers, each scoring honestly, had left an unevenness in the record. The model smoothed it. That is the mirror doing its kind work: it took our collective judgment and gave back a fairer version of it. If that were the whole story, the job would be simple. Find the human bias, expect it in the output, correct for it. Most people stop there. It is not the dangerous part.

The dangerous part is the second bias. The accent one. That was not a human bias faithfully reproduced. no teacher of ours had ever systematically marked down a Brazilian or a German speaker. It was a new artifact the model manufactured out of the shape of our data: it had simply heard Indian-accented English so many thousands of times that it had quietly learned to equate that sound with proficiency itself. We did not put that bias in. The model built it, optimizing on the history it was handed, in a form none of us thought to look for. This is why it is important for us to understand how AI makes decisions before we let it decide.

This is the move leaders miss. You de-bias the obvious thing, you strip the variable you know to watch. And the model, optimizing on everything that remains, finds a proxy you didn’t think to strip. Accent stands in for ability. Zip code stands in for race. CV length stands in for class. The bias you removed reappears wearing a coat you don’t recognize.

Amazon hit the same wall from the other direction. They built a recruiting engine on roughly ten years of CVs, mostly from men, because their tech hiring had mostly been men, and it taught itself that being a man was a feature of a strong candidate. Then it went further, exactly as ours did: it began penalizing the word “women’s,” as in “women’s chess club captain,” and reportedly downgraded graduates of two all-women colleges. Nobody wrote prefer men. The system reverse-engineered the rule, building new proxies for gender out of words on a page. They caught it and scrapped it around 2017.

France named this pattern early and out loud. The Défenseur des droits, the national rights ombudsman, with the CNIL, warned that self-learning systems don’t merely carry discrimination, they massify it, and that for too long this had been, in their words, a blind spot of public debate. Not hypothetical: an employment-office algorithm in Austria effectively marked down women, older workers, and people with disabilities, while everyone treated its outputs as administratively neutral.

Neutral. That word is the whole problem. A number feels neutral. A score feels neutral. When my test went into production, the assessments looked exactly as clean for the Brazilian speaker as for the Indian one. The bias was not in how it looked. It was in what it had quietly learned to hear.

Every leader using AI now has to understand how it decides before letting it decide, and build the systems to review those decisions for what training could never anticipate.

Why this is a leadership skill and not an engineering one

You could read all of this as a technical problem, better data, better audits, better fairness metrics, and hand it to a team. That instinct is the mistake. We had smart engineers. They had tested the model and it passed. The accent bias still got through, because it was invisible until the whole world spoke into it. Strip the bias you can name, and the model finds the proxy you didn’t. Fairness is not a setting you switch on. It is a question about what your organization actually values and who it is willing to be wrong about — and that question has no engineering answer. It has a judgment answer, and judgment is yours.

This is also why “human in the loop” is not the safety net people think it is. A human reviewing the machine catches nothing if they only ever review the cases the model already handles well. Keep checking the same population, and the human just rubber-stamps the same blind spot. The loop only works if you point it at the people the model is most likely to fail.

What I added to my leadership toolbox

Three practices I now reach for by reflex, with any model, every time:

• Test it where it’s weakest, on purpose. Never trust a model that has only been seen in the conditions it was trained in. Put it in front of the part of your audience least represented in its history before you trust it anywhere. We found the accent bias because we ran the test across our whole population and watched who it failed. We would have shipped it clean otherwise.

• Cross-examine the machine like you’d cross-examine a person. Ask one question of yourself, not the system: if a person had produced this result, on whom would I want to check that they were fair? You already know how to interrogate a human decision. The skill is doing it to the machine with the same suspicion, especially when the output looks clean.

• Distrust the moment the room relaxes. The most expensive moment in an AI-native organization is when a clean-looking output makes a capable room go easy. That ease is the bias walking through unchallenged. Almost nobody is trained to catch it.

This is where the new skill starts. Not with the machine. With the discipline to treat the mirror as a mirror, to know that what it shows you is yours, and that seeing it clearly, across everyone and not just the people who look like your data, is the first thing leadership now requires.

I am building this into how my own teams work: a standing habit of testing the clean-looking output hardest against the people it is most likely to fail. Not because I distrust the tools. I have seen them beat us at fairness. Because I have also watched a tool I was proud of quietly fail a Brazilian speaker, and I never want to ship that blind again.

Next week: we go deeper into how humans decide, and why the machine turns out to be the clearest place we have ever had to see it.

The first was Accenture. They beat revenue of $18.7 billion, earnings up nine per cent, operating margin expanding to 17 per cent. By any old scorecard, a good quarter. And the market took nearly 18 per cent off the stock the same afternoon. Not because of the revenue. Because of one line buried underneath it: new bookings slipped to $19.3 billion from $19.7 billion a year ago.

The second number was this: the price of a million AI tokens has fallen roughly 280 times in two years.

I sat with those two together for a while. Because they are the same story, and it is a story I am living inside, not just reading about. I am running a business through this transition. I am trying to transform the company and read the new economics at the same time, and most of the time those two jobs feel like they pull in opposite directions. This morning they snapped into focus. The thing I keep telling other leaders, I had to say back to myself.

We are measuring the wrong things.

Revenue always lagged. When the cost of production is deflating, it misleads.

Here is what those two numbers are really telling us. Revenue was always a lagging indicator, it reports what already happened, last quarter’s decisions arriving in this quarter’s accounts. We lived with that, because in a world of roughly stable prices the trailing number still pointed the same way as the value underneath it. If revenue was climbing, you were broadly winning.

Deflation breaks that link. When the cost of producing what you sell falls 280-fold in two years, faster than any cost curve we have ever had to plan around, revenue stops tracking value at all. A company can do ten times the real work and barely move its revenue, because the price per unit fell as fast as the volume rose. The top line sits still while the business underneath it is transformed. Revenue simply cannot see it. Compute, storage and bandwidth all ran this exact play before, and the lesson is always the same: in a deflationary regime the top line does not just lag the business, it can point in the opposite direction from it. That is what is new. Not that revenue is late. That it now lies.

Accenture does not sell tokens. It sells hours toward an outcome. But the same physics apply. Watch what they actually disclosed: 104 deals over $100 million, up 13 per cent, and yet total bookings fell. Megadeals growing while the total shrinks means the commoditisable middle is being hollowed out. The mid-size body-of-work that used to be staffed with pyramids of junior consultants is the first thing AI eats. That hollow middle is what deflation looks like inside a services firm.

So it is not that AI consulting is cooling. The unit is deflating and when the unit deflates, last quarter’s revenue tells you even less than usual. The market looked straight past it, good as it was, to the forward signal: bookings. Not what Accenture had already billed, but what clients are committing to buy next. Revenue is the rear-view mirror; bookings are the road ahead. And Accenture’s had narrowed.

And be honest about the size of it: that dip was only a couple of points. An eighteen per cent fall is not a verdict on one soft quarter, it is the market asking a larger question out loud. Does AI grow consulting, by handing every client a transformation to buy, or gut it, by doing the work consultants used to bill for? Accenture is the test case, because it sells AI transformation and is exposed to AI deflation at the same time. A softening order book was just the first place that question surfaced.

That is the whole game now. And revenue can’t see it.

A few weeks ago I wrote here about Bank of America hiring close to 4,000 graduates while Citi runs with roughly 20,000 fewer people, two banks, the same AI, opposite bets, and I called it reading the economics of this transition. This is the next question, and it is the harder one. Once you start reading the economics, you eventually have to change what you count. Because the old scorecard was built for a world where the unit price held still. That world is gone.

The price was never the labour. We just priced it that way.

Let me make this personal, because it is.

I work as an advisor. And the honest truth is that I can now do, on my own with AI beside me, work that two years ago would have taken a team of ten, faster, and in some ways better.

So here is the question that should make every leader sit up. If it takes me a tenth of the people to produce the same result, what happens to the price?

If I price the way the industry always has, by the hour, by the head, by the size of the team in the room, my price collapses. I have automated my own value away. That is the deflation, and it is exactly what is eating Accenture’s hollow middle. They built a business on bodies. The bodies were never the point; they were a proxy. And when the proxy gets cheap, a price built on it falls through the floor.

But the result I deliver has not become less valuable. If anything it is worth more, it arrives faster, it costs the client less to absorb, it frees their people for something else. The value did not deflate. The cost of producing it did.

That gap, between the collapsing cost of production and the value that did not move, is the whole game. The price was never really the labour. We priced it that way because, for a century, labour was a decent stand-in for value. AI has broken the stand-in. And the businesses that stay valuable will be the ones that price on the outcome they create, not the effort it took to create it.

Activity is not value

Here is the part most leaders miss, and I missed it too. Cost-saved and pilots-launched feel like progress, but they are activity metrics, they tell you how much happened, not whether anything durable was built.

I have made this mistake myself. We took roughly 15 per cent out of the P&L, real numbers I am proud of, because the cost came after we redesigned, not before. But cost-saved is still a lagging number. It does not tell you whether the work got better, whether people keep coming back to the new way, or whether you now own a piece of the workflow someone will pay for tomorrow.

And the evidence is everywhere. BCG found 60 per cent of companies generate no material value from AI; McKinsey, only 39 per cent see any EBIT impact at all. Two trillion dollars of spend this year, up 36 per cent, and most of it showing up as activity, not value. You cannot fix that with more activity. You fix it by changing what you count.

The new scorecard, and the order matters most

So here is the question underneath all of this: when the price of what you sell is falling, what tells you that you are still building a valuable business, not just a busy one? Revenue will not answer it. It is the very thing the deflation is distorting. You need metrics that hold their meaning while the price drops.

Here is how I now think about it. There are five things worth measuring, and the sequence is not optional, each one only means something if the one before it is true. And the first thing to get straight is that not one of them is “how much AI we use.” They start with the customer and end with the economics.

First, efficacy, the outcome. This is the one we almost left out, and it may be the most important of all. Before anything else: did it actually produce a better result for the customer? Are they measurably better off? You can redesign how you work, get faster, cut your costs, and if the customer’s outcome has not moved, you have created no value at all. Only activity, more efficiently.

Measure it on the customer’s terms, not your own. Find the outcome they actually hired you for, the cost they wanted out, the cycle they wanted shorter, the decision they wanted sharper, and measure the movement in their number. Output is “we delivered the work.” Efficacy is “their number moved.” And here is why it comes first: you cannot price on the outcome, the way I argued you must, unless you can measure the outcome. Efficacy is what turns value-based pricing from a nice idea into a line you can defend on an invoice.

Second, service penetration. How deep your service now sits inside your customer’s work. What share of their need, their workflow, their spend now runs through you? Efficacy proves you helped one customer; service penetration is whether that help spreads, more of their work, and more customers like them. A result you cannot repeat and scale is a case study, not a business. This is the number that grows even while price falls, which is exactly why it is the one to watch in a deflating market.

Third, retention. Service penetration without retention is a pilot that hasn’t failed yet. Do customers come back tomorrow, and the day after — or quietly drift the moment something cheaper appears? Does your share of their work deepen, or decay? Inside your own walls the same question holds: does the redesigned workflow become the only workflow, or do people slide back to the old way the moment no one is watching? Retention is where penetration turns into a moat.

Fourth, learning velocity. This is the one almost nobody puts on a board slide, and it is the most AI-native of all. How fast does the service, and the organisation running it, get better per unit of use? Every interaction either teaches the system something or is wasted. The businesses that compound are the ones whose product improves with use and whose people improve alongside it, so that scale becomes advantage instead of just cost. Penetration gets you the usage. Retention keeps it. Learning velocity is whether all that usage is actually making you harder to catch.

Fifth, margin. Because margin is not a peer to the other four; it is what they produce when you get the first four right. And here is the trap, and it is sharper than it looks. Penetration you can never convert into better economics is just expensive activity. And do not assume the falling price rescues you. The unit price of AI has dropped roughly 280 times and total spending on it has gone up, not down, past two trillion dollars this year, because the moment a thing gets cheap, everyone uses far more of it. Cheap units, exploding volume, a larger bill. So the question is not whether AI is cheap. It is whether, as you scale it, the economics of your business actually improve, or whether you have simply handed all the productivity to your customers and competitors and kept the bill.

Efficacy, service penetration, retention, learning velocity, then margin. Get them out of order and you will optimise for the wrong thing at the wrong time. Measure margin first and you starve the penetration that creates the position. Measure your own efficiency and stop there, as almost everyone does, and you will mistake activity for value, which is the exact place most companies are stuck right now.

And this sequence is also the answer to the question every leader is quietly asking after the cloud years: how do I avoid getting locked in, and where does lock-in actually live now?

The model is not the moat. The model is the thing deflating 280 times. The moat is the embedding, the share of the workflow you occupy, the context the system accumulates, the orchestration layer that lets you swap the model underneath and keep the customer. That logic may help explain why buyers are increasingly paying up for the workflow surface rather than the model itself. When SpaceX agreed to pay a reported $60 billion for the AI coding tool Cursor this June, it was not really buying a model. It was buying the surface where code gets written.

But here is the warning I would give every one of you, because we have been here before. After the cloud, everyone swore they would never get locked in again, and most did anyway, because the embedding that creates value is the same embedding that creates the lock-in. You cannot get share-of-workflow without wiring in. Cloud lock-in, at least, you could repatriate. Pay the egress fee, move the workloads. Organisational lock-in has no egress fee. If you automate before you redesign, and let the human judgment atrophy, you are not locked into a vendor’s contract, you are locked into their way of thinking, and no migration gets that capability back. And the rise of all those AI-consulting arms inside the big firms is there for exactly that.

Sequencing is the real job now

So here is what I have come to believe is the biggest change for a CEO in this transition. It is not which model you pick. It is not how fast you deploy. It is sequencing, the order in which you build, and the order in which you measure.

Because what you measure is what your company focuses on. That is not a slogan, it is mechanics. Put a number on the wall and the whole organisation bends toward it. Put “pilots launched” on the wall and you will get a company very good at launching pilots and very bad at knowing whether any of it mattered. Put “share of the real work now running the new way, and still running it next quarter” on the wall, and you get a company building something that compounds.

I wrote in April that the organisations winning with AI are not deploying more, they are drifting less. This is how you drift less. You put the right number on the wall, and the organisation stops drifting toward the wrong one. Measurement is not the scoreboard at the end. It is the steering wheel.

My own methodology has always run in one order: capabilities, then workflows, then judgment, then the data flow and learning loops that make it improve over time. The mistake leaders make under pressure is to invert it, to automate the workflow before they have redesigned it, to measure the cost they took out before they have asked what the humans should still hold. Redesign before automate. Measure the leading thing before the lagging thing. Sequence the scorecard.

So do this. Pick the one number that, if it goes up, means you are building long-term value, not the one that makes this quarter look good. Put that one on the wall. Watch what your company does next.

What I am doing about it

I am changing my own scorecard. In the business I run, I am starting with the unglamorous part, getting our data in order so we can actually measure efficacy, service penetration and retention at all, and moving those to the top of the page, while pushing cost-saved down to where it belongs: a result, not a goal. I am going to walk into my next leadership meeting and ask one question before any other: is the customer measurably better off, and are they coming back for more? If I cannot answer that, I have no business reporting the cost line at all.

We are transforming ourselves and reading new economics at the same time. We do not get to do one and wait on the other. But we do get to choose what we count, and in this transition, that choice may be the most consequential one a leader makes.

What you measure is what you become. Choose it on purpose.

If this resonated, I write every Tuesday at karine.substack.com on building the companies of tomorrow, not automating the companies of yesterday.

Judgment before automation.

We had spent most of the day on the things you would expect, agent architectures, decision rights, where to keep humans in loops that no longer strictly needed them. Useful, structured, the kind of work that fills a deck.

Then she sat back and said, almost to herself: “I am not sure I know what I am for anymore.”

She was not in crisis. She is one of the most capable operators I work with. She was naming something I have been hearing, in different words, in almost every senior room I have walked into this year.

The work she had built her career on, gathering information faster than her peers, distilling it, distributing it, mediating between the layers of the organisation — was being absorbed into systems that did it more cheaply and, in some narrow ways, better. What was left, she sensed, was something the optimisation era had never asked her to develop, and never paid her for.

She did not have a word for it yet. I am not sure I do either. But I think it is the most important thing happening in enterprise leadership right now, and almost no one is naming it.

The optimisation era told us what to suppress

For thirty years, the operating logic of large organisations rewarded a specific kind of person. Fast. Information-rich. Calm under volume. Able to hold a coherent narrative across a hundred slides and a dozen stakeholders.

We did not call this suppression. We called it professionalism. But it was suppression. We were trained out of the slower forms of attention. Out of the long pause before answering. Out of saying “I don’t know yet” in a room that punished you for not having a view. Out of bringing the body into the work — the gut signal, the discomfort, the something is off here that arrives before the analysis catches up.

The optimisation era did not have a place for those things. It did not need them. The system carried enough redundancy that you could route around a leader’s discernment. The cost of suppressing it was lower than the cost of slowing down to listen to it.

That is the part that is changing.

What the machines are quietly doing to us

The dominant story about AI is that it replaces human work. That story is true at the surface and incomplete underneath.

What I keep watching — in the rooms I advise, and in the data on operating-model redesign coming out of IBM and BCG this year — is subtler. AI is not just removing work. It is removing the cover that work provided.

For decades, a great deal of senior labour consisted of producing the artefacts that justified a decision after the fact. The analysis. The deck. The benchmark. These were not really how the decision got made. The decision got made earlier, somewhere quieter, by someone with judgment. The artefacts were the scaffolding we built around it so the system could absorb it.

When an agent produces that artefact in twenty minutes instead of two weeks, something strange happens. The decision is still required. The judgment is still required. But the comfortable middle layer — the months of analysis that let everyone feel the decision had been earned through process — is gone.

What is left, exposed, is the leader and the call.

That is why the executive I was sitting with did not know what she was for. She had spent her career building the middle layer. The middle layer was being eaten. And no one had shown her what the work on the other side of it actually consists of.

I will tell you what I think it consists of. It is going to sound strange in a strategy piece. I have stopped apologising for that.

Five capabilities the optimisation era did not pay for

These are the things I see compounding in the leaders quietly pulling ahead in this transition. Not the loudest adopters. The ones whose organisations actually work.

Discernment. Not analysis. Not pattern-matching. The capacity to look at five plausible options — all of which an agent can now generate in under a minute — and feel which one is right for this organisation, this moment, these people. Discernment is what survives when first drafts are free. Most leaders have it. Few have been asked to use it as their primary instrument before.

Embodied attention. I am going to use that phrase even knowing what it costs me in a C-suite article. The leaders navigating this well have not entirely outsourced their nervous system to their calendar. They notice when a meeting is producing the wrong kind of energy. They notice when a number is technically correct and somehow still wrong. They trust that signal long enough to investigate it. The optimisation era called this unscientific. It is not. It is the oldest signal we have, and the agents do not have it.

Sovereignty. The capacity to stay a coherent self in an environment constantly producing plausible answers that are not quite yours. I watch leaders lose this in real time. First they ask the model. Then they prefer the model. Then they realise they have not had an unmediated thought in a week. The ones who keep their sovereignty treat AI the way a conductor treats an orchestra — they use it, they are not used by it. It is a skill. It can be lost. It can be rebuilt.

Relational intelligence. Trust at scale, in an environment where most communication is now synthetic or nearly so. Knowing how to be the human in the room when people are starting to wonder whether anything they read came from a human at all. This is becoming the most underpriced capability in senior leadership, and the gap between those who have it and those who do not is widening fast.

Wisdom about what not to automate. The meta-skill. Every organisation will automate everything it can. The leaders who look brilliant in five years will be the ones making careful, sometimes counterintuitive calls about what to protect — which decisions, which conversations, which forms of contact stay human. Not because the human version is more efficient, but because it is what the organisation is actually for.

These five do not sit outside the operating model. They are about to become it. Everything else is being commoditised underneath them.

The reframe leaders are slowly arriving at

Here is the move I keep watching senior leaders make, usually around month four or five of serious AI work, often in a moment they did not see coming.

They begin believing the question is how much can I automate.

They end up at a different question entirely: what is this organisation for, and which parts of being human do I need to operationalise to deliver it?

That second question is not soft. It is the most operational question a leader will face in the next five years. Every answer commits budget, structure, hiring, governance, and time. You cannot delegate it to a transformation office. You cannot outsource it to a vendor. The model cannot answer it for you, because the model does not know what your organisation is for — and, frankly, neither do most boards.

This is the part that surprises me. The AI transition, which everyone expected to be the most technical transformation of our generation, is turning out to be the most human one. Not because anyone planned it that way. Because the machines have stripped away enough of the scaffolding that the human questions are now structurally exposed. You can no longer hide them inside a process.

What I am committing to, and what I am asking of you

I have led four major technology transitions before this one. None of them did this. None of them sent leaders home, late on a Friday, asking what they were for.

This one does. And I have decided to stop treating that as a side effect.

I am building my work, my writing, and the next chapter of what I do around one conviction: the leaders who will matter in this era are the ones willing to develop the capabilities the optimisation era never paid them to develop. Discernment. Embodied attention. Sovereignty. Relational intelligence. Wisdom about what not to automate.

So here is my ask, if you are responsible for an organisation, a team, or even just your own working life in this transition.

Stop asking how much can I automate. Start asking what am I protecting, and why.

Then commit to something specific. A meeting you will not let an agent run. A decision you will not let a model frame for you. A conversation you will keep human even when the cheaper version becomes available. Write it down. Tell someone.

The optimisation era trained us to believe that the things that could not be automated were the things that did not matter.

The next era will be built by the people who understood it was the other way around.

AI transformation, leadership, future of work, organisational design, human capital

The first: Bank of America is hiring close to 4,000 graduates and interns this year, even as it pushes AI deeper into the work. It received around 315,000 applications for those campus seats. The bank is letting AI lift productivity and managing its headcount through natural attrition, not through a press release about cuts.

The second: Citigroup expects to run with roughly 20,000 fewer people, its CFO Mark Mason pointing to automation and AI taking over middle- and back-office work. Wells Fargo’s internal budgets already assume a smaller workforce.

Same industry. Same week. The same technology landing on both their desks. Two opposite decisions.

And here is the uncomfortable part: I don’t think either one is obviously right. Anyone telling you Bank of America is the noble one and Citi the cold one is reading the surface, not the economics.

The reset is not optional

Let me say the thing my own board would say, because it’s true.

AI changes your cost structure whether you like it or not. If your competitor can deliver the same value at a structurally lower cost and you’re still carrying yesterday’s cost base, you are not “principled.” You are priced out. The market moves to the lighter, better offer, and it does not wait for you to feel ready.

So a bank taking 20,000 roles out is not, on its own, evidence of lazy automation. It can be exactly the right call, if it’s funding a genuinely better, cheaper proposition for the customer. The failure isn’t cutting. The failure is cutting and delivering nothing better. That just buys you a leaner march to the same death.

And the other side of it: Bank of America’s hiring announcement is not pure virtue either. For a consumer brand, what you stand for is part of the value. “We still bet on people” is positioning. It’s talent magnetism. It’s a reason a 22-year-old with options picks you. The press release is doing strategic work, not just moral work.

Both moves are strategy. The question is what each one is in service of.

The Reset — a grille de lecture

Here’s how I now think about this. Stop asking “cut or keep.” Ask two different questions, and you get a far more honest picture.

One: are you actually resetting your cost structure? Did real cost come out, or are you carrying the same base you had before AI walked in?

Two: are you raising the value you stand for? Is the thing you uniquely deliver getting sharper, the product, the trust, the brand, the judgment, or is it flat?

Put those on two axes and you get four companies, not two.

1. Reset cost and raised value — you’re building the company of tomorrow. The only quadrant that compounds.

2. Reset cost, flat value — efficient decline. You got lighter and you’re still nothing special. This is the Citi risk if the 20,000 doesn’t fund a better customer.

3. Carried cost, raised value — priced out. Lovely product, structurally too expensive, undercut within a year. This is the warning my chairman would give.

4. Carried cost, flat value — you’re already gone, you just haven’t been told yet.

Notice what the binary hides. “Lighter cost structure or human values” feels like a choice because the two sit on different clocks. A lighter cost base pays out this quarter. Standing for something pays out over years and shows up nowhere on this quarter’s P&L. So leaders experience it as discipline versus softness. It isn’t. On a long enough horizon, the human values are the economics.

The cost that doesn’t show up this quarter

There’s a second axis underneath the first, and it’s the one most commentators will never reach. How you cut decides your future.

If you take your cost out by removing the first rungs, the entry-level seats, the apprentice work, the roles that look most “automatable”, you book a clean saving today. And you quietly write off the mechanism that grows senior judgment. Cut the bottom of the ladder and in ten years you have no one at the top who climbed it. You will have automated yourself out of the very judgment AI cannot supply.

That’s what Bank of America’s 4,000 graduates actually buy. Not goodwill. A pipeline. The next decade’s judgment, in training.

I wrote in March that your org chart is the bottleneck, that AI exposes every implicit decision your organisation never made explicit, every judgment call that lived in people rather than process. Here’s the sequel to that argument. Making judgment explicit only matters if you are still producing humans who have judgment to make explicit. Automate the apprenticeship and you’ve drained the reservoir you were trying to map.

What I actually did with it

I’m not writing this from theory. When operating at my company I’ve taken cost out of the P&L, while shipping innovation close to ten times faster. The reset is real. I lived it.

But the question I kept asking my team was not “how much can we cut.” It was “what are we funding with it?” We chose to put the freed capacity into building, into becoming, whether we wanted it or not, a learning institution. The cost came out so the value could go up. That order matters.

If you’re sitting with this decision right now, here’s what I’d tell you. Audit which costs you’re cutting before you celebrate the number. Are you removing duplication and toil, or are you removing your own future bench? Decide what better thing the saving funds before you take it, not after. And if your only answer to “what does AI give us” is “a smaller team,” you haven’t finished the work. You’ve done the easy half.

So, which one?

Back to the question I opened with. A lighter cost structure, or standing in human values, which builds the most durable company?

Neither. Alone.

The lighter structure with nothing behind it is a slow death. The human values with no cost discipline is a fast one. The companies still standing in ten years are the ones who refused the choice, who used AI to get genuinely lighter and to deepen what they uniquely stand for, and who were careful never to cut away the people their future judgment depends on.

The leaders who think they have to pick one are the ones who won’t be here to argue about it.

I know which company I’m trying to build. Build the companies of tomorrow, not automate the companies of yesterday.

#HumanXAI #FutureOfWork #AIStrategy #Leadership #Judgment

Not because I have to. I have a team. I could hand it to someone. I have chosen not to and the more I watch other leadership teams, the more I am convinced that this choice is the only reason it is working.

Here is the scene I keep seeing in the rooms where I advise. A board meeting. The most senior leader in the room points at AI and assigns the work to someone else, a CTO, a newly appointed Chief AI Officer, sometimes a transformation lead. The intent is real. The delegation is real. And within a quarter, that workstream quietly changes shape. It stops being about what the function could become. It starts being about how the function can be made smaller. Reinvention turns into a cost exercise. And nobody notices the substitution, because the headline number — efficiency — still goes up.

That substitution is what I want to write about today. And the data has finally caught up with what I have been watching.

The substitution

On 4 May, IBM’s Institute for Business Value published a study of more than 2,000 organisations. The headline: 76% of companies have now appointed a Chief AI Officer, up from 26% a year ago. An enormous shift in a single year.

But the finding that actually carries weight sat lower in the report. The organisations genuinely delivering on their AI objectives, the ones converting investment into outcome, had done something the others had not. They had redesigned five core areas of the business: technology, finance, HR, operations, and the cross-functional work between them. That group was four times more likely to have delivered on its business objectives than the group that had only made the appointment.

Read that again. The appointment is not what produces the value. The redesign is. And most of that 76% have not done the redesign. The same study found only 25% of the workforce uses AI regularly, even though 86% of CEOs believe their people have the skills to. The gap is not capability. It is design.

Here is what I think is happening. The CAIO appointment has become the most common substitute for the design work itself. It looks like a decision. It feels like a decision. It can be announced on a board call. But the decisions AI actually requires, what stays human, what becomes agentic, where the handoffs sit, which functions get rebuilt rather than augmented — are not decisions a new title can make from outside the operating committee. They are architectural choices about how the company makes decisions. And those belong to the person whose job is to design the company.

What the Cloudflare story was actually about

You saw the Cloudflare announcement. So did I. Matthew Prince and Michelle Zatlyn cut 1,100 people, 20% of the workforce, on the same day they reported Q1 revenue up 34%. The post was titled Building for the Future. The market took more than 20% off the stock by close.

Cisco cut 4,000 the same week, also during record revenue, and the market barely moved.

The difference is not the size of the cut. It is whether investors could see the new company behind the smaller one. Prince offered productivity multipliers — two times, ten times, a hundred times more productive per person. He offered a 600% rise in internal AI usage. What he did not offer was the redesigned operating model. He offered the consequence of one, not the architecture of one. And the market priced the gap.

That is the same gap I see in the rooms I sit in. The leader has not designed the new company. They have delegated the question of what the new company should be. So when the productivity gains arrive, the only thing the operating model knows how to do with them is shrink, because nothing else has been designed to absorb them.

This is what I mean when I say recomposition, not reduction. Cutting is what an undesigned system does with an efficiency gain. Designing is the work that turns the same gain into something that compounds.

The CEO is the systems architect

Here is how I now think about this. The CEO is no longer the chief allocator of AI investment. The CEO is the systems architect of the AI-era company.

That is a different job. It is not about choosing models or vendors, or whether to centralise the data lake. It is about answering five questions no one else in the company has the authority to answer.

One. What decisions will humans continue to make? Not what tasks humans will do, what decisions. This is the question most leaders skip. Decision rights, not task lists.

Two. Where will agents execute end-to-end, and what does a wrong answer cost? Agentic work compounds. A wrong agent decision propagates faster than a wrong human one. The architecture has to know which calls are reversible and which are not.

Three. Which functions get reinvented, not augmented? Augmenting an obsolete function is the most expensive thing a company can do with AI. Some functions need to be rebuilt from the operating model up.

Four. Where does accountability sit when the work is mixed? When an agent and a human co-produce a decision, who carries the consequence? If you cannot answer that in plain language, the EU AI Act’s Article 14 will answer it for you in August.

Five. What does the company sound like to a customer in two years? The easiest question to skip, and the most important. If you cannot answer it, you do not have an AI strategy. You have an automation programme.

These are not technology questions. They are design questions about what the business is. A CAIO who joined three months ago cannot answer them. The person whose job is to design the company has to.

What this reveals

The danger here is not loud. It rarely is. AI failures arrive quietly, fewer questions asked, less friction, dashboards staying green while judgment leaves the system. Delegated design fails the same way. The productivity numbers climb. The board is satisfied. And two years on, the company is smaller, faster, and no more valuable than it was, because no one ever designed what it was becoming.

What to do this week

Audit last quarter’s AI workstreams. For each one, ask whether the outcome looked like reinvention or like reduction. If the answer is always reduction, you are running a cost programme wearing agentic vocabulary.

Look at where your CAIO reports. If the answer is technology leadership, ask whether you have hired a senior engineer with a new title, or a senior designer of the company. If it is the former, the CAIO is not the problem. The design vacuum above them is.

Then sit down, alone, for two hours, and answer the five questions for one function. Not the whole company. One. Whichever one you think about most. If you cannot answer them after two hours, you have just found your next strategic priority.

The commitment

I am driving the AI redesign of the business I operate myself, this year. Function by function, decision right by decision right. Not because I enjoy the discomfort, but because I believe the leaders who try to hand this off will spend the next two years explaining to their boards why productivity went up and value did not.

I am asking my fellow C-level peers to do the same. Drive it yourself. The design of the company of the AI era is not a workstream. It is the job. It is your job.

I will write again in a few weeks about what I am learning doing it. In the meantime, if you are sitting on a CAIO appointment and a quiet sense that something is not landing, call it what it is. The appointment is not the strategy. You are.

AI strategy, leadership, CEO, organisational design, Human × AI, agentic AI, future of work

Sources: IBM IBV Study, May 2026 · BCG, “AI Has Made Work Reinvention a CEO Mandate” · Cloudflare Q1 earnings (CNBC) · Davenport & Bean, MIT Sloan

The headlines have been clean, fast, and almost entirely wrong about what Dorsey is actually claiming.

Because Dorsey is not Andy Jassy. And Dorsey is not Mark Zuckerberg. All three cut management. All three said “flatter is faster.” But Dorsey is making a fundamentally different argument than the other two — and most leaders reading his essay as a template for their own restructuring are about to find out the hard way.

The Three Cuts, Properly Read

Let me lay out what these three CEOs have actually said, in their own words, because the conflation is everywhere.

Andy Jassy at Amazon. Jassy’s argument is cultural. “Bureaucracy is really anathema to startups and to entrepreneurial organizations.” He’s not replacing managers with anything. He’s removing layers so that individual contributors have more ownership and the company moves faster. He created a “Bureaucracy Mailbox” and invited employees to flag excessive process. This is a speed-and-culture intervention. AI is barely in the frame.

Mark Zuckerberg at Meta. Zuckerberg’s 2023 “Year of Efficiency” memo said it plainly: “Every layer of a hierarchy adds latency and risk aversion in information flow and decision-making.” He asked managers to become individual contributors. His frame is latency reduction — flatter so information moves faster. AI came into Meta’s org design much later, with the CEO Agent and the digital clone. It was a consequence of the flattening, not the cause.

Jack Dorsey at Block. Read his essay carefully. Dorsey is not arguing for speed. He is not arguing for culture. He is making a much more audacious claim: hierarchy exists to perform three specific functions — route information, pre-compute decisions, and maintain alignment across a complex organization. And AI, specifically, can now perform those three functions better than humans can. Therefore the hierarchy is not inefficient. It is obsolete.

That is a categorically different argument.

Jassy and Zuckerberg flattened the org chart. Dorsey is claiming he has replaced what the org chart was for. One is a structural efficiency move. The other is a philosophical claim about what a company is.

The Misreading That Is About to Get Expensive

Here is the pattern I am already watching form in the leadership teams I advise.

Boards read the Dorsey essay. They see the 40% cut. They see the word “mini-AGI.” And they conclude: if Block can do this, we should too. Cut the middle layer. Buy the AI tools. Move faster.

But what they are copying is the outcome, not the precondition. And Dorsey is explicit about the precondition, even if the press isn’t.

Block has two things most companies don’t have. First, it is remote-first by design. Every decision, every discussion, every plan exists as a recorded artifact in a machine-readable system. Second — and this one is the bigger structural advantage, the one almost nobody is naming — Block operates in a domain where reality itself is machine-legible. Cash App and Square do not produce a customer satisfaction score. They produce money movement. Transactions clear or they don’t. Merchants take the loan or they don’t. Consumers abandon the cart or they complete it. The feedback loop is transactional, high-frequency, and has almost no ambiguity about the outcome.

Most enterprises do not operate in that kind of reality.

They operate in negotiated reality — customer satisfaction surveys, brand perception studies, employee engagement scores, lagging financial indicators, quarterly reviews. Learning outcomes, if you work in education, which I have spent much of my career doing. Every one of these signals is contested, interpreted, and mediated by judgment before it ever becomes something a system can reason over. You cannot build a world model on top of negotiated reality without first doing the heavy interpretive work of deciding what counts as truth. That work is, itself, judgment.

In Dorsey’s own words: “In a remote-first company where work is already machine-readable, AI can build and maintain that picture continuously. What’s being built, what’s blocked, where resources are allocated, what’s working and what isn’t. That’s the information the hierarchy used to carry. The company world model carries it instead.”

Read that sentence twice. Dorsey is saying that his model works because the artifacts already exist, and because the signal feeding them is ground truth rather than interpretation. The “world model” is not built from scratch. It is harvested from a company that has been operating as if every action were telemetry for fifteen years, in a domain where reality is legible enough to be computed at all.

Your company is not Block. Most organizations I work with are hybrid at best. Decisions happen in meetings that are never transcribed. Context lives in Slack DMs. Judgment flows through conversations that never touch a document. The reason the middle layer exists is not because leadership is lazy. It is because the middle layer is the only part of the organization that has the context required to coordinate anything.

Remove that layer without first building what Dorsey built, and you are not implementing an AI-native operating model. You are deleting your coordination function and hoping something fills the gap.

The Distinction Nobody Is Making

I want to name the thing I keep seeing conflated. Dorsey removed the coordination layer. He did not remove the judgment layer. Those are different.

Coordination is routing information, tracking what’s in flight, maintaining shared context across a large system. AI is genuinely good at this — especially when the underlying artifacts are already structured. This is the work Dorsey’s world model actually replaces.

Judgment is deciding what matters, making calls under ambiguity, taking responsibility for outcomes when the model cannot. AI does not do this. It frames options. Humans judge. Roelof Botha — Dorsey’s own board member and co-author — made this point directly on the Sequoia podcast, correcting Dorsey live: “I don’t think the AI makes most of the decisions. The AI helps with communicating the alignment, and the management team or the inner core helps set the framework.”

The board member walked back the claim in real time. That should tell you something.

If you read the Dorsey essay carefully, judgment has not disappeared from Block. It has been concentrated. Leadership compresses into three irreducible functions: define outcomes, allocate attention, decide under uncertainty. Those functions now live with a small core at the top, supported by the intelligence layer — not distributed through a middle layer that is gone.

That is a very different outcome from “AI replaces managers.” It means the top of the company carries more judgment weight, not less. The AI carries the coordination. The humans carry the decisions. And if the humans at the top are not themselves AI-native — if they cannot reason inside the model the intelligence layer produces — then the whole structure collapses back onto them as over-extension.

I’ve written about this before. I am living it. And I suspect many CEOs who copy the Dorsey model without the Dorsey preconditions will too.

The Contrarian Test: “But AI-Native Companies Are Different”

I can already hear the counter-argument. “This is the template for the AI-native future. Legacy companies just need to catch up.”

Look closer at what Dorsey actually built.

Block did not eliminate its management layer by deploying better tools. Block eliminated its management layer by spending years turning every action inside the company into structured, machine-readable artifacts — and by owning a direct, transactional customer signal that most companies do not have and cannot manufacture. The model was the output of a decade of infrastructure work. It is not a quarter of restructuring.

The companies I see rushing to copy this model are skipping that decade. They have beautiful off-site decks describing their new “intelligence layer.” They have a team of twelve working on an internal agent. And they have firing lists for middle management that they plan to execute before the infrastructure is actually in place.

That is not transformation. That is an airplane deleting its landing gear while still in the air.

What Dorsey Actually Changed And What Most Leaders Are Missing

I do think the Dorsey manifesto is important. I think the part that is important is not the part that is being discussed.

Here is the real shift. For twenty years, operating models have treated alignment as a communication problem. You align people by cascading goals, running all-hands, holding QBRs, building OKR trees. Alignment was the output of human effort — meetings, documents, rituals.

Dorsey is arguing that alignment can now be a system property. The intelligence layer holds the company’s state, updates it continuously, and surfaces priorities as they emerge. Alignment stops being something leaders enforce and becomes something the system enables.

This is a bigger idea than “flatten the org chart.” It is a claim about where alignment lives. And it has a dangerous implication that almost nobody is discussing.

If alignment becomes system-driven, misalignment scales at machine speed. A flawed assumption in the model propagates instantly. A biased signal amplifies everywhere. A local optimization cascades across every function before any human notices.

This is the risk I think boards are most underweighting, and I want to be precise about why.

In a human-coordinated system, misalignment is localized, slow, and visible. Someone notices. Someone pushes back. Someone walks into the office of the person two floors up and says the thing that everyone knows but nobody has said out loud. It is friction, and it is inefficient, and it is also — unglamorously, unceremoniously — the reason bad decisions do not compound as fast as they could.

In an AI-coordinated system, misalignment is global, fast, and invisible. The middle layer was not only routing information. It was error damping. It caught anomalies. It applied contextual override when the rule said one thing and the situation called for another. It created friction against bad propagation. Remove that layer, and you have not only lost coordination. You have lost the system’s ability to notice that it is wrong.

This is a phase change in how organizational failure happens. Not a degree, a kind. And almost nobody implementing the Dorsey model is designing for it.

The quality of alignment, in Dorsey’s world, becomes the quality of system design. That is a much higher bar than most organizations are currently clearing.

The Missing Word: Design

If you strip the Davos misdirection on AI, you find the same word at the center that I argued for three months ago: judgment.

If you strip the Dorsey misdirection, you find a different word at the center: design.

Jassy is running a cultural program. Zuckerberg ran a latency program. Dorsey is running a design program. And almost everyone I see adopting Dorsey’s vocabulary is running a headcount program in disguise.

Design means: what decisions does this system need to make? What context does it need to hold? What does the customer signal have to look like? Where does human judgment sit, explicitly and by choice? What scales through the system, and what must never scale through the system because it is the thing the system is for?

These are not questions a restructuring plan answers. They are questions an operating model answers. And the gap between a restructuring plan and an operating model is where most “AI transformations” will quietly fail over the next twelve months.

What I Am Saying to the Leadership Teams I Advise

If your board has sent you the Dorsey essay and suggested you “do something like this,” here is what I would tell you.

Read past the 40%. Read what Dorsey actually says about the preconditions — remote-first, artifact-generating, direct customer signal. Ask yourself honestly whether you have them. If you don’t, building them is the transformation. Not the layoff. Not the agent. The infrastructure underneath.

Then ask the harder question. What does your middle layer actually do? Not what it is supposed to do on the org chart. What judgment does it exercise, what context does it hold, what coordination does it perform? If you cannot answer that in writing, you cannot replace it — because you have not yet named what you are replacing.

Dorsey can cut because Dorsey knows. Most of your organizations do not yet know. That is not a moral failing. It is the actual work.

The Real Shift

Jassy and Zuckerberg removed layers. Dorsey removed the reason layers exist.

That is the distinction. And the corollary is the part that should keep boards awake: if you remove the layer without removing the reason the layer was there, the system will recreate it — poorly, and under stress. Usually on top of whoever at the company is AI-native enough to hold the gap. Often the CEO. Sometimes no one, in which case the coordination simply stops happening and nobody sees it for a quarter.

Jassy flattened to move faster. Zuckerberg flattened to reduce latency. Dorsey redesigned what a company is.

Only one of those is an AI-native move. The other two are speed plays dressed up in AI language — and the speed plays will look correct for eighteen months until the judgment layer quietly erodes underneath them.

If you want to take Dorsey seriously, take him seriously all the way down. Build the world model. Harvest the customer signal. Design the judgment architecture before you delete the human one. And accept that the hardest part of this is not the cutting. It is the naming — making explicit, in writing, what your organization has been running on implicitly for years.

Because here is the thing about the Dorsey manifesto that almost nobody is saying out loud.

He did not prove that AI replaces management. He proved that Block, specifically, had built the substrate that made its management layer redundant. The essay is an invoice for fifteen years of infrastructure work — and a warning to anyone who thinks they can send the invoice without doing the work.

In my last article, I wrote about the seven traits of AI-native leadership, the behaviours I’ve watched separate leaders who transform their organisations from those who produce excellent documents that nobody acts on. The trait that mattered most was the simplest: doing the work, not just planning it. This article is about what that looks like in practice.

I was running a cluster workshop with the leadership team of an organisation I advise. We were three hours in. The room was deep in process mapping, workflows, decision trees, data requirements. Good work. Thorough work. The kind of work that makes everyone feel productive.

Then I stopped the room.

“This is not a planning exercise,” I said. “Your engineering team already transformed how they work with AI. They didn’t do a full plan. They learned by doing. They learned with iterations.”

It landed hard. Because it was true. Over the previous six months, the engineering function at that same organisation had quietly, radically changed how they operated —with AI embedded throughout their workflows. No transformation workshops. No formal roadmap. No steering committee. They had simply started experimenting. Testing tools in their actual work. Iterating on what worked and discarding what didn’t. And six months later, they were operating completely differently.

Nobody had asked them to wait for permission. So they didn’t.

The Planning Trap

Here is the pattern I see in nearly every organisation I advise: leadership recognises that AI is strategic. They assemble a cross-functional team. They commission a landscape review, a capability assessment, a gap analysis. They build a transformation roadmap. They align stakeholders. They present to the board.

And then six months have passed. And nothing has actually changed.

I’m not saying planning is worthless. I’m saying that planning without experimentation is fiction. You are writing a story about a future that doesn’t exist yet, based on assumptions you haven’t tested, about capabilities you haven’t tried. The document feels like progress. It isn’t.

The engineering team I mentioned didn’t succeed because they were smarter or more technical. They succeeded because they collapsed the gap between thinking and doing. They didn’t theorise about where AI could add value, they tried it, watched what happened, and adjusted. Their transformation was messy and uneven and sometimes wrong. It was also real.

What Experiments Actually Surface

Let me tell you about a customer success lead at one of the organisations I advise. She didn’t wait for the transformation plan. She took a real client account and decided to test the methodology herself, building an AI-generated value realisation report from scratch.

The AI produced beautiful output. Structured, clean, impressive. Activation rates, engagement metrics, cost per unit. The kind of thing that looks great in a slide deck.

Then she hit the wall.

To actually prove customer value, she needed data from three different platforms that weren’t connected. She needed context from the CRM that was incomplete. She needed to define what “value” meant for this specific client, something the organisation had never formally decided. And when the AI pushed back and essentially said “if you want to prove ROI, you need to know the business impact on the client’s side,” she had to stop. Not because the experiment failed. Because it succeeded.

That one experiment surfaced more about what was actually missing, the data connections, the undefined value metrics, the gaps between systems, than weeks of planning would have produced. No document review could have found it. You had to try the thing to discover what the thing required.

This is what I mean when I say experimentation is the strategy. Not because planning doesn’t matter, but because the most important information only becomes visible when you act.

Build First, Optimise Later

One of my advisors said something in a workshop that I’ve since adopted as a principle: “We are not optimising for that now. We first need to build. And then optimise.”

This runs against every instinct in enterprise culture. We are trained to design before we build, to anticipate before we execute, to get alignment before we move. And in stable environments, that sequence works. But AI transformation is not a stable environment. The tools are changing quarterly. The capabilities are expanding faster than any roadmap can track. The workflows that make sense today may be obsolete in six months.

In that context, the cost of over-planning is higher than the cost of an imperfect experiment. A plan that takes three months to build is already outdated by the time it’s approved. An experiment that takes three days to run gives you real data you can act on immediately.

I push this hard in every workshop I facilitate. “I don’t want to be in analysis paralysis,” I tell the room. “It’s for us to get into a mindset to kick start with an open view and then we iterate and we’ll learn. So we’re not going to codify everything before we do.”

That statement makes some people uncomfortable. Good. Discomfort is the leading indicator that the organisation is about to learn something.

The Operating Principles Behind This

I’ve codified this into how I work and how I advise. Two principles in particular.

Decide with enough information, not all information. Start, review, adjust. Iterating on something real is faster than deliberating on something theoretical. This doesn’t mean being reckless. It means recognising that the last 30% of information you’re waiting for will only reveal itself through action. You cannot research your way to certainty in an environment that hasn’t stabilised yet.

Scope to prove, then expand. Validate with the smallest viable audience before broadening. The customer success lead didn’t try to redesign value reporting across the entire organisation. She took one client. One report. One test. And what she learned from that single experiment reshaped how the entire team thought about value measurement. Small scope isn’t a limitation, it’s a discipline.

These principles work together. You make a decision with enough information. You scope it to something provable. You run the experiment. You learn. You adjust. You expand. That loop, not the plan, is the transformation.

What This Looks Like In Practice

If you’re leading an AI transformation right now, here’s what I’d tell you.

Stop waiting for the complete picture. You will never have it. The organisations I’ve watched stall are the ones that kept adding requirements to the plan, one more data source, one more stakeholder sign-off, one more risk assessment. The organisations that moved kept the scope small enough to start.

Find your engineering team. Every organisation has someone, a team, a function, an individual, who is already experimenting with AI in their actual work. Find them. Learn from them. They have more practical knowledge about what works and what doesn’t than any consultant report will give you. The engineering team I described didn’t need permission. They needed to be noticed.

Run one real experiment this month. Not a pilot programme. Not a proof of concept with a vendor. Take an actual workflow, an actual client, an actual problem and test whether AI changes the outcome. It will be imperfect. That’s the point. The imperfections are the learning.

Treat what breaks as the roadmap. The customer success lead’s experiment didn’t produce a perfect value report. It produced a clear map of everything the organisation hadn’t built yet, the data connections, the definitions, the cross-system integrations. That map was more valuable than any planned architecture review because it came from reality, not theory.

The Real Risk

I want to be direct about something. The biggest risk in AI transformation is not moving too fast. It is not a failed experiment. It is not an imperfect first attempt.

The biggest risk is spending so long planning that the organisation never develops the muscle of experimentation. Because that muscle, the ability to scope something small, run it, learn from it, adjust, and go again, is the actual capability that separates organisations that transform from organisations that talk about transforming.

Planning feels like control. Experimentation feels like exposure. And most leadership teams, honestly, choose the feeling of control.

I’m asking you to choose the exposure instead.

The engineering team at that organisation didn’t have a better plan than everyone else. They had a bias to action. They had the discipline to scope small and iterate fast. And they had something that no amount of planning can produce: six months of real learning baked into how they work every single day.

You can have that too. But you have to start. Not after the roadmap is finished. Not after the data architecture is complete. Not after everyone is aligned. Now. With whatever you have. On something small enough to finish this week.

Start. Review. Adjust. That’s it. That’s the whole method.

The Moment That Started This

I was midway through a market engine workshop with a leadership team I advise. We’d spent the morning mapping workflows, what the system needs to know, what it needs to decide, where intelligence should sit. Standard territory for these sessions.

Then one of the product leaders did something I wasn’t expecting. Instead of mapping her cluster around headcount and handoffs, which is what most leaders default to, she redesigned the entire thing around what she called a “collapse loop.” Signal to testable output in hours, not weeks. She wasn’t optimising existing roles. She was reimagining the unit of work itself.

And she wasn’t describing a vision. She had built a working prototype. On her laptop. During the break.

I wrote down what she’d done. Not because it was impressive, though it was, but because I’d seen that exact combination of behaviours before. In different people. At different organisations. In different workshops. And I’d never named it.

That evening I went back through my notes from every workshop I’d facilitated over the past year. I wasn’t looking for a framework. I was looking for repetition. What do the leaders who are actually moving their organisations forward with AI do differently from those who talk about it convincingly but don’t shift anything?

Seven patterns kept showing up.

What I Actually Observed

I want to be clear about the methodology here, because it matters. I did not sit down and ask “what should AI-native leadership look like?” That question produces aspirational lists. Aspirational lists produce posters. Posters produce nothing.

Instead, I reviewed specific moments from real workshops. People making real decisions about real workflows. And I asked: what separates the leaders whose sessions produce genuine transformation from the ones whose sessions produce excellent documents that nobody acts on?

Here’s what I found.

Thinks in systems, not tasks. The strongest leaders I observed don’t start with “what can we automate?” They start with “how does this system work end to end and what does it need to know and decide?” In the market engine workshop I mentioned, the team discovered that the right first question wasn’t “which AI tool should we buy?” It was “what does the workflow need to know and decide?” The unit of design is the system, not the activity. This sounds obvious when you write it down. It is remarkably rare when you watch people work.

Treats data as organisational infrastructure. At one organisation, a data lead refused to let different clusters build competing definitions of truth. This was not a popular position. It slowed things down. It created friction with teams that wanted to move fast with their own metrics. But she held the line, positioning data as shared organisational infrastructure rather than just reporting. She understood something most leaders miss: clean, connected data is a precondition for every intelligence layer you want to build. You can’t have five versions of “customer health” and expect an AI system to make sense of any of them.

Separates role architecture from capability development. This one is uncomfortable. The leaders who move fastest design the role for the AI-first system first, what does this function need to look like when intelligence is embedded throughout? and then ask whether current people fit. They don’t retrofit people into legacy structures. That doesn’t mean they’re ruthless about people. It means they’re honest about the sequence. Design the system, then develop the capabilities to match. Not the other way around.

Tolerates ambiguity in the build phase. I watched a senior leader name, openly and in front of her entire team, that she didn’t know the answer to a sequencing question. “We’re going to hold a directional vision without false certainty,” she said. “Iteration is the method, not a fallback.” This is harder than it sounds. Most organisations reward the leader who has the plan. These leaders have the direction and the discipline to resist pretending they have more than that.

Uses AI demonstrably in their own work. This is the one that separates everything. The product leader with the collapse loop prototype didn’t just describe what AI could do. She built it. She ran live experiments. She brought her team into real workflows. Demonstration is required stated intent is not enough. I have sat in rooms where senior leaders give eloquent speeches about AI transformation and have never once used an AI tool to do their own work. You can feel the gap. So can their teams.

Maintains a strong human judgement layer. Counter-intuitively, the most AI-forward leaders I observed are clearer about what humans must own, not less. They make the boundary between automated and human decisions an explicit leadership call. Not a default. Not an accident. A deliberate choice about where human judgement is irreplaceable and where it’s just habit.

Makes the implicit explicit for the organisation. The final trait is about scale. These leaders don’t just transform their own teams. They create shared language and structures that make the whole system smarter. They codify what works so it travels beyond their own function. This is the difference between a great leader and a great leadership culture.

The Cascade Problem

If you’re reading this and thinking “great, we’ll distribute these seven traits and build a development programme” slow down. Because I’ve watched that exact move fail, repeatedly.

The most common failure in leadership-led transformation is rushing the cascade. Organisations announce the standard, distribute the framework, and discover that the middle layer, who never saw senior leadership genuinely living it, has no real reference point. The framework becomes wallpaper.

Here’s how I now think about the cascade, and it’s changed how I advise every organisation I work with on this.

The executive team goes first. Not symbolically, substantively. They redesign their own functions through an AI-first lens before asking anyone else to. They run their own cluster workshops. They build their own prototypes. They sit with the discomfort of not knowing the answer. They demonstrate the traits before they name the traits.

Only then do you cascade. And when you do, the middle layer has something they almost never have in transformation programmes: a living reference point. Not a slide deck. Not a competency model. An actual example of their senior leaders doing the work.

Making It Stick

The seven traits on their own are observation, not infrastructure. What makes them operational is connecting them to something the organisation already uses.

In the organisations I advise, we took these seven patterns and mapped them onto the organisation’s existing performance framework, the behavioural model they already use for development conversations, performance reviews, and promotion decisions. This was deliberate. If you create a parallel framework, you create a parallel universe that everyone ignores. If you embed the traits into what already exists, they become observable and verifiable. Not aspirational.

So instead of “this leader is visionary about AI”, which means nothing, you get “this leader redesigned their function’s workflow around system-level outcomes before assigning tools or people.” That’s something you can see. Something you can coach toward. Something you can hold someone accountable for.

What This Means If You’re Leading Right Now

I want to speak directly to you for a moment, if you’re a senior leader in the middle of this.

You do not need to be technical. You do not need to build your own AI tools, though it helps more than you think. What you need is to stop outsourcing transformation to a programme and start demonstrating it in your own work.

Pick one of the seven traits. The one that feels most uncomfortable. And do it this week. Not as a workshop exercise. In your actual work, with your actual team, on an actual problem.

If you think in tasks, try mapping the system. If you’ve been retrofitting people into old structures, try designing the role first. If you’ve been talking about AI without using it, open a tool and build something. It doesn’t have to be good. It has to be real.

Because here’s what I’ve learned from watching dozens of leaders across these workshops: the gap between the ones who transform their organisations and the ones who don’t isn’t knowledge. It isn’t budget. It isn’t even vision. It’s whether their teams have ever seen them do the thing they’re asking everyone else to do.

That’s it. That’s the whole thing.

The seven traits aren’t a framework I invented. They’re a pattern I recognised. And the leaders who embody them didn’t learn them from a model. They learned them by doing the work, messily, imperfectly, in front of their teams, and letting the organisation watch.

In a future article, I’m going to write about one leader I’ve watched embody all seven, not as a profile in perfection, but as a case study in what it actually looks like when someone leads this way, including the parts that were uncomfortable and the moments they got it wrong. Because leadership development for the AI era needs fewer aspirational lists and more honest mirrors.

The Room

I was standing at a whiteboard with a product lead who had just walked me through one of the best functional documents I’d seen in months. Workflows mapped end to end. KPIs attached. Decision points identified. It was thorough. It was clear. It was also incomplete, and neither of us knew it yet.

The next morning, I brought engineering, data, and commercial into the same room and asked them to co-create the same document together. Within an hour, the original version was unrecognisable. Not because it was wrong. Because no single function could see the full system. Engineering surfaced dependencies the product team had never mapped. The data team pointed out that two of the KPIs were actually measuring the same behaviour from different angles. The commercial lead asked a question about customer impact that reframed an entire workflow.

That session is when I stopped thinking of AI Capability Recomposition as a framework and started thinking of it as a practice. Something you do with people in a room, not something you present on a slide. A key element of AI transformation and the best learning investment.

The Question

In my last article, I promised to share the single question I now open every workshop with, the one that consistently surfaces what the organisation has never codified. Here it is:

“What does success look like for this cluster?”

That’s it. Not “what do you do?” Not “what can we automate?” Not “what’s your AI roadmap?” Just: what is this group of people here to achieve?

It sounds simple. It is not. One of the first teams I worked with came in with a beautifully detailed funnel view — every stage mapped, every handoff documented, every role assigned. A perfect execution view. But when I asked that question, there was a pause. They had mapped how the work flows. They hadn’t defined what the work was for.

I pushed them to answer that question first. What emerged were two outcomes nobody had named explicitly: a frictionless customer experience, and a cost-to-serve ratio that made the business model sustainable. Once those were on the whiteboard, the funnel view reorganised itself. Some activities turned out to be essential. Others turned out to be motion. You could feel the energy shift in the room.

Step two: build the KPI tree. This one surprised me. In every workshop, we hit the same wall: “everyone contributes to LTV.” Or retention. Or engagement. Pick your north-star metric, everyone can make a case that they influence it. Which means nobody truly owns it.

So we had to go deeper. I started building what I now call KPI trees, decomposing the top-level outcome metric into the specific leading indicators that each cluster can actually own and move. Not “we all contribute to retention” but “this cluster owns time-to-value for new customers, measured by activation within 14 days.” Specific. Measurable. Owned.

The KPI tree concept wasn’t in my original methodology. It emerged from the workshops because the teams needed it. That matters. It tells me the methodology is alive, not fixed.

Step three: map the real workflows cross-functionally. Not the documented version. The version that actually happens. Every function in the room, mapping together. This is where the invisible architecture surfaces — the handoffs that happen over Slack, the decisions that live in someone’s head, the ownership gaps that humans have been quietly filling for years. I covered this in my first article, so I won’t repeat it here. But I will say: it hits differently when you’re standing in the room watching people discover it in real time.

What Surprised Me

Three things I did not expect.

The overlap problem is not a problem, yet. One of my advisors, someone I trust deeply on organisational design, said something that stuck with me: “We might think we are overlapping, but because the context is different, we don’t know yet. We are not optimising for that now. We first need to build. And then optimise.”

This ran against my instinct. I wanted clean boundaries. Clear ownership. No duplication. But he was right. When you’re redesigning around clusters for the first time, trying to eliminate overlap too early kills the process. Teams need permission to be messy. They need to build first, see what emerges, and then refine. Premature optimisation in organisational design is just as dangerous as it is in engineering.

The engineering team was already there. While I was carefully designing workshop agendas and facilitation flows, the engineering team at one organisation had already transformed their ways of working with AI. No formal programme. No workshops. No methodology slides. They just started using AI tools, learned what worked, iterated, and evolved their practices organically.

When I found out, my first reaction was mild panic, had my workshops been unnecessary? But then I realised: this was actually proof that the methodology works. Not because the engineers followed my process, but because the core principle held. Iterative, cross-functional learning beats top-down planning. They had done intuitively what the clusters are designed to do deliberately. The workshop structure isn’t the point. The behaviour change is.

People don’t resist the reframe, they’re relieved by it. I expected pushback when I told teams to set aside their functional plans and think in clusters. I expected territorial behaviour, turf wars, the usual organisational immune response. Instead, what I mostly got was relief. People had been feeling the limits of functional thinking for years. They knew the silos weren’t working. They just didn’t have language for what to do instead, or permission to try.

The Mindset That Makes It Work

Here’s the thing I keep coming back to. I don’t want to be in analysis paralysis. It’s for us to get into a mindset to kick start with an open view and then we iterate and we’ll learn.

That sentence isn’t polished. I said it mid-workshop, off the cuff, to a team that was stuck trying to make their cluster design perfect before they’d even tested it. But it captures something I now believe is foundational to this work.

AI Capability Recomposition is not a one-time reorganisation exercise. It is an operating rhythm. You define the outcomes. You build the clusters. You map the workflows. You deploy AI where it makes sense. And then you learn. The clusters evolve. The KPI trees get refined. The workflows get redesigned as you discover what the AI can actually do versus what you assumed it could do.

If you’re reading this and you’re leading an AI transformation or advising someone who is, here’s what I’d tell you directly. Stop trying to design the perfect AI operating model before you start. You will not get it right the first time. Nobody does. The organisations that are actually moving are the ones that gave themselves permission to begin with an imperfect structure and iterate their way to something better.

The Method, Summarised

For those of you who want the practical takeaway, here’s how I now run this.

Start with outcomes. Not tasks, not workflows, not org charts. What is the strategic result this group of people exists to deliver? Name it. Write it on the wall.

Build the KPI tree. Decompose the outcome into the metrics each cluster can actually own. Go deep enough that ownership is unambiguous. If everyone can claim credit, you haven’t gone deep enough.

Co-create cross-functionally. Never let one function design the workflow alone. The product lead’s document was excellent. It was also incomplete. The system only becomes visible when every function that touches it is in the room.

Tolerate overlap early. Don’t optimise for clean boundaries on day one. Build first. See what emerges. Optimise later.

Iterate relentlessly. The engineering team that transformed without workshops taught me something I needed to learn: the method matters less than the mindset. Build, test, learn, adjust. Repeat.

What I’m Committing To

In a future article, I’m going to publish a before-and-after — the original functional document from one of these workshops alongside the version that emerged after cross-functional co-creation. Anonymised, but real. Because the difference between the two is the clearest illustration I have of why no single function can see the full system.

There are parts of this I’m still figuring out. How to handle the transition when clusters cross business units. How to measure whether a cluster is actually performing better than the functional model it replaced. But if you are trying to transform your organisation with AI and you haven’t put every function that touches a workflow in the same room to co-design it around outcomes — you are building on sand.

Start there. Start messy. Start now.

I have been running AI capability workshops with executive leadership teams I advise, and in a mid-size tech company redesigning how their customer lifecycle works. The head of the function had put together a strong document. Clear workflows, KPIs mapped, roles identified. I remember thinking: this team is further ahead than most.

Then I asked one question: who owns the handoff between marketing and sales?

Silence.

Not confused silence. Recognition silence. The kind where everyone in the room suddenly sees something that was always there. A human had been sitting in that gap for years, reading the room, making a judgment call, passing the baton based on instinct and experience and a dozen signals that were never written down. It worked. It had always worked. Nobody had ever needed to answer the question because the question had never been forced.

Until we tried to design an AI workflow around it. And then the gap became visible. Not as a failure, as a mirror.

What the Workshops Keep Revealing

That moment has now repeated itself in every workshop I’ve facilitated. Different organisations, different functions, different levels of AI maturity. The same discovery.

At one organisation, we were mapping the customer lifecycle end to end. The team had a beautifully detailed funnel view — every stage, every handoff, every role. When I asked “what does success look like for this cluster?”, there was a pause. They had mapped how the work flows. They hadn’t defined what the work was for. Two outcomes emerged that nobody had named: a frictionless customer experience, and a cost-to-serve ratio that made the business model sustainable. Neither was in the original document. Once those were on the whiteboard, the team could see immediately which activities drove value and which were just motion. I watched a senior ops lead lean back in her chair and say, “We’ve been optimising the wrong things.”

At another organisation, I asked the customer success team how they assess account health. The answer was different from every CSM in the room. Not slightly different. Fundamentally different. Different data inputs. Different thresholds. Different definitions of what “at risk” even means. It worked, because humans are adaptive, and each CSM had built a mental model that functioned well enough in their own territory. But hand that to an AI agent? It would fail spectacularly. Not because the AI isn’t capable. Because the organisation had never decided what “healthy” actually means.

This is the pattern. And the data backs it up. 56% of companies are not seeing ROI from their AI investments. When you look at where those investments went, the pattern is consistent: they started with tasks. Automate this report. Speed up this process. Reduce this headcount. What they did not start with was the outcome.

The Invisible Architecture

Here’s what I’ve come to understand and it took watching enough workshops to see it clearly.

A huge amount of what makes an organisation actually work was never codified. It lives in people’s heads. In the way an ops lead knows which data to trust and which to ignore. In the way a customer success manager reads a client’s tone on a call and decides whether to escalate. In the handshake between teams that happens in a Slack message, not in a process document.

I watched this surface in real time during a product development workshop. The product lead had put together a genuinely excellent document — workflows mapped, KPIs identified, decision loops designed. It was thorough. It was clear. Then we brought in engineering, data, and the commercial team and asked them to co-create the same document. Within an hour, it had changed fundamentally. Engineering surfaced dependencies the product team had never mapped. The data team pointed out that two KPIs were measuring the same behaviour from different angles. The commercial lead asked a question about customer impact that reframed an entire workflow.

None of this was visible from inside any single function. The system only became visible when every function that touches it was in the same room.

And then something happened that I didn’t plan. One of the customer success leads decided to test the framework herself. She took a real client account and tried to build an AI-generated value realisation report. The AI produced beautiful output — activation rates, engagement metrics, cost per unit. Impressive on the surface.

Then she hit the wall. To actually prove customer value, she needed data from three different platforms that weren’t connected. She needed context from the CRM that was incomplete. And she needed to define what “value” meant for this specific client, something the organisation had never formally decided. When the AI pushed back, essentially saying “if you want to prove ROI, you need to know the business impact on the client’s side”, she had to stop.

Not because the experiment failed. Because it succeeded. That single test surfaced more about what was actually missing than weeks of planning would have. The data connections, the undefined metrics, the gaps between systems, all of it became visible because someone tried to do the real thing instead of just mapping it on a whiteboard.

Why Codification Is the Transformation

What I tell every leadership team I work with now is this: do not automate before you codify.

But here’s the thing I didn’t expect when I started saying that. The codification itself is transformative. Even before you add any AI.

I’ve watched teams come out of these workshops operating differently, not because of any technology, but because the act of making explicit what was implicit changed how they saw their own system. People who had been filling gaps for years finally saw those gaps acknowledged. Teams that had been working around broken handoffs finally got to redesign them. A data lead at one organisation refused to let different clusters build competing definitions of truth. It was not a popular position. It slowed things down. But she held the line and six months later, every intelligence layer they built worked because there was a single version of reality underneath it.

That data lead understood something most leaders miss: you cannot build intelligence on top of ambiguity. Clean, connected, agreed-upon data is not a nice-to-have. It is the infrastructure.

What I’d Tell You Directly

If your AI strategy starts with “where can we use AI?”, stop. Start with “what have we never codified?”

Define the outcome before you touch the workflow. The 56% of companies not seeing AI ROI automated activities without first defining what success looks like at the system level. Before you touch a single workflow, put every function that touches it in the same room and ask: what is this cluster for? What does success actually look like? What KPIs tell us whether we’re getting there? If you cannot answer those questions clearly and cross-functionally, you are not ready to automate.

Map what’s in people’s heads. Audit the decisions humans are making that nobody documented. The CSMs with five different definitions of account health. The ops lead who knows which data to trust. The handoffs that happen over Slack because there’s no formal process. Map the actual system, not the documented one. The gap between those two versions is where your AI deployment will break.

Treat the workshop as the first ROI. Don’t wait for the AI deployment to see results. The customer success lead who tested the framework on a real client learned more in one afternoon than the planning team learned in a month. The product team whose document was transformed by cross-functional co-creation started operating differently the next week. The codification itself improves how the organisation works, before a single AI agent is deployed.

In my next article, I’m going to share the single question I now open every workshop with, the one that consistently surfaces what the organisation has never codified. It’s not complicated. But the conversations it triggers have changed how every team I’ve worked with sees their own system.

The question is not whether your organisation is ready for AI. It is whether your organisation has ever been fully designed or whether it’s been running on human improvisation the whole time.

If it’s the latter, AI will find out. Fast.

We were talking about AI oversight — how to keep humans meaningfully involved as more decisions move into automated workflows. And one of my colleagues looked up and said: “But who actually wants to do oversight? Just watching a system run all day — that sounds incredibly boring.”

He was right.

And the moment he said it, I realised we had been thinking about this entire problem backwards.

The signal we are missing

Knowledge is leaving people.

Not because people are becoming less capable. But because AI systems are absorbing knowledge at a pace and scale that no individual, team, or training programme can match.

The person who had been in the room for the last ten budget cycles — who knew things a new hire could not learn from a handbook — that person’s institutional memory is now being encoded into models, workflows, and decision layers that run continuously, without fatigue, without the friction of human availability.

This is not a future state. It is already true in every function that has seriously adopted AI: marketing, sales, finance, operations, product.

And most leadership teams are responding to it the wrong way.

The misunderstanding

The instinct is governance. Review processes. Approval layers. Human-in-the-loop checkpoints.

The question organisations keep asking is: how do we make sure humans stay involved?

That is the wrong question.

It assumes the challenge is containment — keeping humans relevant by inserting them into AI workflows at regular intervals. But not all human involvement is equal. A human reviewing an AI output because the process requires a review is not the same as a human making a judgment the system genuinely cannot make.

One is theatre. The other is leverage.

And here is the part that most leadership teams are not naming: oversight is not a junior role. It is not a transitional role. It is not something you hand to someone while you figure out what else to do with them.

Real oversight — the kind that catches what the system misses, that knows when to intervene and when to let it run, that can read an anomaly and understand what it means — requires experience. Deep experience. And it is genuinely hard to find.

If you design oversight roles as if they are passive, you will attract people who treat them passively.

And then your AI drifts.

What drift actually looks like

Drift is not a dramatic failure. That is what makes it dangerous.

What I keep hearing across conversations with leadership teams is a version of the same problem. The definitions were never properly codified — what a qualified lead actually looks like, what customer acquisition criteria genuinely mean, what the guardrails are and who owns them. The system runs on assumptions that felt clear enough at the start. And then the market moves, the business evolves, and nobody has gone back to revise what the system is optimising for.

I have seen this with something as fundamental as ICP definition — ideal customer profile — or what qualifies as a sales-accepted lead. These are not exotic edge cases. They are the core logic of a growth function. And if that logic is not codified precisely, tested regularly, and revised deliberately, the system does not fail loudly. It just quietly drifts away from what you actually need it to do.

Until it gets out of your hands.

That is the real cost of under-designed oversight. Not a single dramatic error. A slow accumulation of small misalignments that compound — until the gap between what the system is doing and what the business needs is too wide to close cheaply.

Three layers and the one nobody is building

Here is how I now think about this, across the work I do with leadership teams on AI transformation.

Every enterprise operating with AI at scale is running on three layers, whether it knows it or not.

The embedded knowledge layer. Where knowledge actually lives — in models, agents, workflows, semantic data layers, system memory. This layer operates whether or not humans are paying attention. It accumulates. It drifts. It becomes the de facto source of organisational truth.

The extraction layer. How knowledge is accessed — through interfaces, queries, dashboards, natural-language analytics. The capability that matters here is not technical literacy. It is the ability to ask the right question of the system at the right moment. This is the new form of organisational intelligence.

The intervention layer. When humans step in — at defined decision thresholds, escalation triggers, judgment calls, correction loops. This is where human involvement is designed, not assumed. Not defaulted to. Intentionally placed where it creates the highest value.

Most organisations are building Layer 1. Some are beginning to think about Layer 2.

Almost none have seriously designed Layer 3.

That gap is where enterprises fail quietly.

What intervention design actually looks like

This is not governance. It is engineering.

The difference matters.

Governance asks: are humans in the loop? Intervention design asks: are humans in the right place in the loop, doing the right kind of work?

Here is what I ask every leadership team to define before they automate anything.

Who owns the final decision — and at what seniority. Not a team. A person. With a title and accountability to match.

What triggers escalation. Not “when something goes wrong.” A specific, measurable signal that everyone agrees on before the system goes live — not after the first incident.

What is never automated. Every organisation has judgment calls that belong to humans permanently. Name them explicitly. If you do not, the system will absorb them by default.

How human corrections feed back into the system. Every override is a signal. If it is logged, tagged, and fed back, it compounds into organisational intelligence over time. If it disappears into an exception report nobody reads, you will be making the same correction in six months.

How long the system runs before human review is required. Autonomy without a review clock is how drift starts.

This is not a checklist. It is a design conversation. And most organisations are having it too late — after deployment, when the gaps are already showing.

In practice, what surfaces almost universally is that organisations have been conflating three things that are not the same.

Execution tasks — high-volume, repeatable, rule-bound. AI is the engine.

Oversight tasks — validation, anomaly detection, quality review. AI assists, humans govern. And the humans doing this need to be senior enough, experienced enough, and engaged enough to actually govern — not just sign off.

Architecture tasks — defining the logic, setting the thresholds, interpreting the trade-offs. Humans must own this, entirely.

The structural mistake is leaving execution-level people in oversight roles, and leaving oversight-level processes in place of architectural thinking. When this happens, AI does not fail loudly.

It drifts.

The economics of it

Every human intervention has a cost and a value.

The cost is real: time, cognitive load, attention, decision fatigue. In an AI-native workflow, human involvement is no longer free. It is a resource with a market rate.

The value is also real — but unevenly distributed.

Human judgment at the wrong point creates friction without adding value. Human judgment at the right point — interpreting an anomaly the system flagged but cannot explain, deciding whether a legal redline crosses a line the model cannot read, navigating a relationship dynamic that no system can model — creates value the system cannot replicate.

But here is what I keep coming back to after that conversation with my team: the right point also has to be an interesting point.

If you design oversight as passive monitoring, you will not attract the people who can actually do it well. You will attract people who are, understandably, just waiting for something to go wrong.

The goal is not to maximise human involvement. The goal is to optimise intervention ROI — to place human judgment precisely where it creates more value than system autonomy would.

And to make that a role worth doing.

Three things that matter most right now

Audit your intervention points before you automate. Map where human judgment currently lives — not where it is supposed to live, but where it actually lives. What decisions are made by instinct? What escalations are informal? What knowledge exists only in one person’s head? Make it explicit before the system absorbs it.

Redesign roles around the three layers, not around legacy job titles. Which roles need to shift from execution to oversight — and which people have the experience, the judgment, and the appetite to do oversight well? Not everyone does. That is not a criticism. It is a design reality.

Build correction loops as infrastructure, not as process. Every human override of an AI decision is a signal. If it is logged, tagged, and fed back into the system, it compounds into organisational intelligence over time. If it is handled as an exception and forgotten, the system never learns and the human never stops correcting. The difference between those two outcomes is not technology. It is design intent.

The enterprises that navigate this well will not be the ones that adopt AI fastest.

They will be the ones that treat experienced human judgment as the scarce, valuable resource it actually is — and build roles worthy of the people they need to fill them.

I am personally committed to pushing on this in every leadership conversation I have. Because my colleague was right. Nobody wants to do boring oversight.

And that is precisely the problem we need to design our way out of.

• Signal: The enterprise shift from AI-as-intelligence to AI-as-agency is no longer speculative. Gartner, McKinsey, and the WEF are pointing at the same structural change and most organisations are not designed for it.

• Misunderstanding: Treating the agency layer as an automation problem misses its defining characteristic: the proliferation of consequential actions that outpace organisational oversight. The bottleneck was never capacity. It was always judgment.

• Reframe: The right frame is humans as the coherence layer, the organisational function responsible for ensuring distributed, autonomous systems remain aligned, interpretable, and purposeful at scale.

• What leaders should do: Define who owns coherence. Redesign performance to surface judgment quality. Build sensing mechanisms for what systems are actually doing, not just what they are producing.

• The thesis: Agents multiply action. Only humans multiply meaning. The question is not how many agents to deploy, it is whether the coherence layer exists to make them worth deploying.

The Conversation That Keeps Happening

I had a version of the same conversation three times last month with a CHRO at a consumer brand, a COO at a financial services firm, and a Chief Digital Officer at a global logistics company.

Each of them is deploying AI seriously. Each of them has a governance framework on paper. And each of them, when I asked the same question, went quiet for a beat too long.

The question was simple: Who in your organisation is accountable when your AI systems start drifting out of alignment with each other?

Not “who owns the AI strategy.” Not “who approves the models.” Who is accountable for coherence, the ongoing alignment of distributed autonomous systems with organisational intent, as those systems scale?

In all three conversations, no one held that role. And all three leaders already knew it. They just didn’t have a frame for it yet.

That gap is what I want to name today. Because it is not a technology gap. It is a structural one. And it is already operational.

The Signal

If you have followed the AI conversation in the enterprise this year, you have heard one story: agents. Every major technology platform has announced an agentic product. Every consulting firm has published a framework. Gartner projects that 33% of enterprise software will embed AI agents by 2028, and 40% of enterprise applications will carry agent capabilities by 2026. Three-quarters of organisations are already in active experimentation.

The consensus is clear: move faster, deploy more, automate everything.

But while the industry races to answer how do we act more? McKinsey’s research on multi-agent deployment quietly identified a different and more dangerous problem that almost no governance board is currently tracking.

Enterprises are not replacing legacy systems with AI. They are layering agents across existing stacks, creating fragmented ecosystems where redundancy, misalignment, and invisible drift become the dominant operational risks.

Why the silence on this? Perhaps because deploying an agent is a procurement decision. Building coherence is a leadership one. It is easier to buy capability than to design accountability.

But the coherence crisis is already operational and it changes the fundamental question of where human value sits in organisations that now act at machine speed.

The Misunderstanding: Output vs. Orchestration

The prevailing response to the agency layer treats it as an automation problem. More tasks delegated. More efficiency unlocked. More cost removed. This frame is not wrong but it is structurally insufficient.

Automation addresses capacity. The agency layer introduces something categorically different: the proliferation of consequential actions taken by systems that operate faster than human oversight can track, across more dimensions than any governance structure was designed to hold.

In cybersecurity, AI agents now automate a third of threat detection workflows and generate thousands of reports monthly. They are producing at scale. But production at scale without coherence at scale is not efficiency, it is managed entropy.

The bottleneck was never capacity. It was always judgment. And AI’s arrival does not reduce the need for judgment. It concentrates it.

The Reframe: Humans as the Coherence Layer

The right frame is not humans versus AI, nor even humans alongside AI.

It is humans as the coherence layer, the organisational function responsible for ensuring that distributed, autonomous systems remain aligned, interpretable, and purposeful as they scale.

The World Economic Forum’s research on future workforce competencies points to the same conclusion: the skills gaining value are not technical fluency but systems thinking, cross-functional orchestration, and the capacity to design and govern agent-driven workflows. The most valuable contributors are becoming designers of systems, not performers within them.

Connect the dots: McKinsey identifies fragmented agent ecosystems as the central operational risk. The WEF identifies systems thinking as the central human competency. They are pointing at the same gap, the absence of a coherence function in most organisations deploying AI at scale.

This repositioning involves four structural shifts that I see playing out in every serious AI transformation I work on:

From doing to framing. When systems can execute, value moves upstream to the articulation of goals, the definition of constraints, and the deliberate navigation of trade-offs. The most effective leaders will not be the fastest. They will be the clearest.

From executing to orchestrating. Individual output becomes less strategically relevant than systemic throughput. The question shifts from what a person produces to how effectively they direct and align what distributed systems produce on the organisation’s behalf.

From visibility to interpretability. As decisions distribute across AI layers, the ability to trace causality, identify contradictions, and understand what a system is actually doing, as distinct from what it was instructed to do, becomes a core organisational competency. This is not a UX problem. It is an organisational awareness problem.

From performance to judgment. When AI dramatically increases the volume of available actions, human judgment about what matters, what to stop, and when to intervene becomes the genuine scarce resource. Judgment cannot be automated. It can only be cultivated, and it must be structurally valued.

The Counterargument Worth Engaging

There is an obvious objection. AI-native companies — unicorns built entirely on model-driven workflows — appear to have minimal human-in-the-loop. They are moving faster than incumbents. Does the coherence argument hold for them?

Look closer. The AI-native companies that endure are not the most automated. They are the ones with the most disciplined human feedback loops.

Midjourney is not simply a model. It is a community of millions of humans whose aesthetic judgments continuously train the system. That judgment is the moat. Remove the human, and you have a commodity inference engine awaiting obsolescence by the next model generation.

The pattern also holds in the negative. The wave of “wrapper” startups launched in 2023 — lightweight interfaces over GPT-4, with no differentiated human layer — were largely eliminated when OpenAI released successive updates. They had no sovereignty because they had no human judgment embedded in their systems. They were renting coherence, not building it.

The rule holds at every scale: more agents without human coherence produces more drift, not more value.

Three Questions Every Leadership Team Needs to Answer

I use these three questions in almost every leadership conversation I have now. Most teams cannot yet answer all three. That gap is where the work starts.

1. Who owns coherence?

Not AI performance. Not output volume. Who is accountable when autonomous agents begin producing outputs that no longer align with organisational strategy? In most enterprises, no one holds this role explicitly. That is not a technology gap. It is a structural one. Name the role before you deploy the next agent.

2. Are you measuring judgment or just output?

Current performance frameworks reward task completion and efficiency. The most critical human contributions in an agentic environment — catching silent drift, resolving contradictions between systems, maintaining strategic alignment under scale — are invisible to these frameworks. What organisations do not measure, they do not protect. And what they do not protect, they will lose.

3. Can you see what your system is actually doing?

The greatest risk of agentic AI is not dramatic failure, which is visible and correctable. It is gradual, silent misalignment — systems optimising for proxies, decisions compounding in unintended directions, outcomes degrading before anyone notices. The organisations that will lead are those that build deliberate sensing mechanisms for understanding what their systems are doing, not just what they are producing.

What I’m Doing Differently

I’ll be honest about the change I’ve made in my own work.

Every engagement I take on now starts with those three questions before we touch technology. Not as a checklist, as a diagnostic. Because I’ve learned, sometimes the hard way, that deploying more capability into an organisation that hasn’t defined its coherence layer doesn’t accelerate transformation. It accelerates drift.

The question I’d ask you to sit with before your next AI investment: Does your organisation have a coherence layer and does the person responsible for it know that’s their job?

If the answer is no, that’s your starting point. Not the next model. Not the next pilot.

Write it in the comments. I’d genuinely like to know where your organisation is on this.

The prevalent anxiety about agentic AI focuses on control, systems acting in ways humans cannot govern. That concern is legitimate. But the more immediate risk is subtler. Not loss of control. Loss of coherence. Systems that drift. Decisions that conflict. Outcomes that degrade in silence while the metrics still look clean.

Human × AI is not, ultimately, a question about technology. It is a question about where judgment, coherence, and intent reside in systems that now act.

Agents multiply action. Only humans multiply meaning.

The question is not: how many agents can we deploy? It is: are we building the coherence layer that makes them worth deploying?

A few weeks ago I was on a call with the leadership team of a mid-size infrastructure and software company. Smart people. A technically sophisticated CTO. A board that had already mandated AI as a strategic priority.

When I asked how the AI journey was going, the CEO paused.

“Mixed. Some groups have done great work. Others are more skittish. We’re scattered.”

They weren’t short on tools. They weren’t short on talent. They had teams with two decades of machine learning experience. They’d run workshops. They had an AI policy.

What they didn’t have was a coherent answer to a simpler question: who decides what, and when?

That gap — not the technology — was the bottleneck.

And in every organization I work with, it shows up the same way. Different industry, different maturity level, different leadership team. Same fracture point.

The Misunderstanding

The instinct, when AI deployment stalls, is to look at the technology.

Better models. More sophisticated orchestration. A more integrated platform.

That instinct is wrong and acting on it is expensive.

Organizations aren’t stalling because their systems don’t talk to each other. They’re stalling because their decisions don’t.

For decades, companies scaled through functions. Work moved step by step — marketing to sales to operations to finance. Each team owned its part. And humans absorbed everything in between: the ambiguity, the trade-offs, the gaps that existed between handoffs but belonged to no one explicitly.

That model worked, because human intelligence was the scarce resource. Humans were hired, in part, to navigate what couldn’t be codified.

Now the constraint has shifted.

Execution is increasingly automated. The cost of intelligence is collapsing. And the functional model that served organizations for decades begins to fracture — not because AI underdelivers, but because it illuminates what was never explicitly designed.

AI doesn’t fill gaps. It executes whatever you give it.

Which means every ambiguity that humans quietly absorbed, every implicit trade-off, every undocumented judgment call, is now exposed. Not gradually. All at once.

The Reframe

The conversation inside most organizations is about how to deploy AI across functions.

That is the wrong level of analysis.

The deeper issue is that functions were never the right unit of organization for what AI makes possible, or for what it demands.

When AI operates across systems simultaneously, work stops behaving like a chain of ownership. It starts behaving like a flow. Customer onboarding is no longer contained within one team. Revenue is no longer owned by one function. Product experience no longer sits neatly inside product.

These become continuous systems, moving across humans and machines, across departments and time zones, without natural pause points where a human can absorb the ambiguity.

And that surfaces the real constraint. Not intelligence. Not execution capacity. Decision clarity.

Who decides what? When does AI act autonomously? When does a human intervene, and which human? How are trade-offs resolved when no single function owns the outcome?

Most leadership teams cannot answer these questions cleanly. Not because they are avoiding them. Because until now, they never had to. The answers lived in people, in institutional knowledge, in informal authority, in the judgment of experienced operators who knew when to escalate and when to absorb.

On that same call, one of the technical leaders said something that crystallized this precisely:

“We’ve stopped debugging when something breaks. We just wait for the next model iteration and run it again.”

That is not a technical decision. That is an organization that has quietly outsourced its judgment to the model, because the decision loop was never made explicit in the first place.

The consequences are visible across organizations attempting cross-functional AI deployment: conflicting ownership, duplicated logic, inconsistent outputs, unclear escalation paths. Everything functions in isolation. Nothing functions together.

This is why AI initiatives stall after pilots. Not because the technology fails. Because the organizational system was never designed to absorb it.

What This Requires From Leadership

The response to this is not a new governance framework or an AI steering committee. It is a fundamental redesign of how the organization operates. Three shifts define that redesign.

Start by mapping decisions and expect the map to be incomplete.

Most organizations have documented their processes. Very few have mapped their decisions, who owns them, at what level of authority, under what conditions AI acts, and when a human must intervene.

The instinct is to build a comprehensive decision map before deploying further. That instinct is right, but only partially. Map as many decisions as you can. Then deploy , because AI will surface the ones you missed.

This is the insight most leaders are not prepared for: it is impossible to codify everything in advance, because you cannot always know what needs to be codified until the system shows you. Implicit decisions, the judgment calls that lived in people rather than processes, only become visible when AI reaches them and has no instruction to follow. That moment of friction is not a failure. It is diagnostic information. Treat it as such.

The discipline is to build the organizational habit of capturing those gaps in real time: when AI stalls or produces inconsistent output, ask not what is wrong with the model, but what decision was never made explicit. That is where the work is.

Redesign around flows, not functions.

Functions don’t disappear in this model, but they stop being the primary organizing unit. The real unit becomes the flow: customer lifecycle, onboarding, revenue generation, retention. That is where value is created, and where it breaks when ownership is unclear.

This reorientation has structural consequences. Integration is no longer infrastructure. It is the system itself. The organizations that accelerate past their competitors will not necessarily have more sophisticated models. They will have the operational discipline to connect systems, align data, and coordinate humans and AI into coherent, end-to-end flows.

Leaders who resist this — who hold onto functional purity because it is legible and comfortable — will slow execution without understanding why. The friction will not present itself as an organizational design problem. It will present as AI underdelivering.

Change the scorecard and accept what that reveals.

Governance in the context of AI-driven flows is not a control mechanism. It is a clarity mechanism.

The question can no longer be: how is each function performing? It must become: how is the whole flow performing? Those are different instruments, and the gap between them is where organizational dysfunction hides most successfully.

Most leadership teams are running the first scorecard while believing they are running the second. Changing that is harder than it sounds — not technically, but politically. Local metrics that look healthy can coexist with a system that is quietly breaking between them. Surfacing that reality requires a willingness to measure outcomes that cross functional boundaries and attribute them accordingly.

The companies that build this capability will not just deploy AI faster. They will be the ones where AI compounds — each decision loop informing the next, across functions and over time, in ways that are genuinely difficult for competitors to replicate.

The Bottom Line

AI is not simply changing how work is done. It is changing how work must be structured.

The shift is from functions to flows, from execution to orchestration, from the management of intelligence to the design of decisions.

In that world, the scarce resource is not model capability. It is organizational clarity, the ability to define, in advance and in real time, how decisions happen across humans and AI at every point in the flow.

The companies that lead this transition will not be distinguished by the sophistication of their technology. They will be distinguished by their willingness to look at what the technology exposes, the implicit decisions, the undocumented trade-offs, the judgment calls that were always there but never named, and do the hard work of making them explicit.

That work is not glamorous. It does not generate headlines.

But it is the only foundation on which AI-driven organizations are actually built.

The Conversation That Stopped Cold

I was recently speaking with a senior executive at one of the world’s most recognised brands.

They hadn’t started their AI transformation. Not for lack of budget. Not for lack of ambition.

When I walked through what AI actually requires from an organisation — explicit workflows, clear decision rights, structured data — it was the decision question that stopped the conversation cold.

“We’re a very matrixed organisation,” he said. “That’s the problem.”

He already knew. He just didn’t have a frame for it.

Why Most AI Initiatives Stall

This is not an unusual conversation. It is the norm.

Most companies are still thinking about AI as something you add to the organisation. A tool to deploy. A capability to acquire. A productivity lever to pull.

That framing looks right on paper. That’s exactly why it’s dangerous.

AI agents don’t fill organisational gaps. They expose them.

Human workers are extraordinarily good at navigating ambiguity. They escalate when something feels wrong. They carry context that was never written down because it never had to be. They fill in what isn’t defined.

AI agents cannot do any of that. They require explicit workflows, clear decision rights, structured data, and defined escalation paths. When those don’t exist, you don’t get automation. You get a mirror — precise and fast — reflecting back the absence of architecture you didn’t know you had.

This is why most AI initiatives stall. Not because of the models. Because the organisation itself was never designed to run with AI.

The Shift Nobody Is Making Fast Enough

AI is not entering the enterprise the way software did.

It is entering the way electricity did. Invisible. Pervasive. Foundational.

You don’t “use” electricity. You build everything on top of it.

The organisations that understand this have stopped asking where to use AI. They’ve started redesigning themselves to run on it. That shift — from adoption to architecture — changes everything about how you design teams, assign accountability, and measure performance.

When AI becomes infrastructure, three things change.

Intelligence disappears into the product. Users don’t “use AI” — they experience outcomes. The workflow matters more than the interface.

Work stops being owned end-to-end by any one person. Tasks are orchestrated across humans and agents. The hand-off points, the escalation paths, the edge cases — all of it has to be designed, not assumed.

Performance stops being about individual productivity. It becomes about how well the system operates. That requires a different measurement lens. And a different leadership mindset.

The Three Layers Most Organisations Are Getting Wrong

To operate in this new reality, every company needs to redesign three things. Most are skipping the hardest one.

The Data Layer is the foundation. Not just whether data exists, but whether it’s structured, accessible, and clean enough for AI to act on reliably. Without this, AI doesn’t fail dramatically. It fails quietly. Outputs that look right but aren’t. Decisions made on bad inputs nobody caught. Drift that compounds before anyone notices.

The Decision Layer is the one that stopped that executive cold. Every organisation must define which decisions are automated, which are augmented, and which remain human. This is not a technical question. It is a leadership one. And in most matrixed organisations, nobody has answered it, because nobody owned it enough to force the answer.

The Orchestration Layer is where value is actually created. This is where humans and agents work together as a system. Where the workflow is designed, not just described. Most companies try to build this layer before they’ve built the other two, which is precisely why their pilots look promising and their transformations stall.

What Leaders Now Have to Do Differently

The unit of work is no longer the employee. It is the human–AI workflow.

That changes what leadership actually means. Leaders now have to design systems, not just teams. Define decision rights across humans and AI. Manage performance at the system level. Build accountability into workflows, not just roles.

This is a fundamentally different job. And most leadership teams haven’t caught up to it yet.

Three places to start:

Run an organisational readiness audit before your next AI investment. Do you have explicit workflows? Defined decision rights? Structured data? Most teams discover they’ve been trying to run AI on infrastructure that was never designed for it. That’s not an AI problem. It’s a design problem — and it’s fixable, but only if you name it first.

Make the Decision Layer a leadership conversation, not an IT one. What AI decides, what it recommends, and what stays human shapes accountability at the system level. It belongs on the leadership agenda, not buried in a vendor evaluation.

Shift your performance lens from individual to system. The right ROI question for AI isn’t “how much time does each person save?” It’s “how much better does the whole system operate?” Leaders measuring at the individual level will consistently underinvest in orchestration, which is exactly where the compounding returns are built.

The organisations that win won’t have the best models, the most pilots, or the largest AI budgets.

They’ll be the ones that did the harder work first, governing the work before they deployed the agents, designing the system before they measured its outputs.

The question is no longer: “How do we use AI?”

It is: “Are we building an organisation AI can run on or one it will quietly break?”

The Assumption Nobody Audited

Every AI deployment I’ve seen starts from the same unstated belief: that the organisation already knows how work happens.

It doesn’t.

Most enterprises run on informal coordination.

Most teams won’t describe it this way. But they feel it every day. Processes that exist in someone’s head. Decision rights that evolved through habit, not design. Responsibilities that overlap because no one ever drew the boundary.

For years, this was manageable. Humans are extraordinarily good at navigating ambiguity. They fill gaps without being asked. They escalate when something feels wrong. They carry institutional knowledge that was never written down because it never had to be.

AI agents cannot do any of that.

And so when you deploy an agent into a workflow that was never explicitly designed, you don’t get automation. You get a mirror. A very precise, very fast mirror reflecting back the absence of architecture you didn’t know you had.

This Is Not a Technology Problem

The instinct is to treat it like one.

Build better permissions frameworks. Add monitoring dashboards. Deploy policy engines. All of that is necessary. None of it is the solution.

The enterprise AI stack is forming across three layers: an intelligence layer (the models), an execution layer (the environments where agents act), and a governance layer (the systems that manage accountability and permissions).

Most organisations are racing to build the first two. The third is arriving as an emergency, because they’re discovering something uncomfortable.

Agents are not generating content. They’re acting inside systems. Accessing sensitive information. Triggering workflows. Influencing decisions.

Without governance, agents don’t scale productivity. They scale risk.

But here’s the deeper problem: governance systems need something to anchor to. You can’t build a permissions framework for work that hasn’t been defined. You can’t set escalation paths when no one has agreed where decisions live.

Technology governance assumes organisational clarity. Most organisations don’t have it.

The Unit of Work Has Changed

For decades, organisations designed work around job descriptions.

That model is structurally insufficient now.

Consider a customer onboarding process today. An AI agent collects documentation. A human reviews the edge cases. Another agent performs risk checks. A human approves the final decision.

Who owns that process? Where does judgment live? When does a human have to intervene, and who decides that threshold?

The unit of work is no longer the employee. It is the human–AI workflow.

That sounds like a subtle shift. It isn’t. It changes everything about how you assign accountability, where you place oversight, and what it means for something to go wrong.

When work was entirely human, ownership could be implicit. When agents participate, it has to be explicit. Because agents don’t have the judgment to recognise when something is wrong. They have the speed to make it worse very quickly.

AI Doesn’t Fail Loudly

This is the part leaders consistently underestimate.

AI failure almost never shows up in the metrics first. It shows up in the erosion of something harder to measure: trust, judgment, accountability.

Small ambiguities in ownership compound. A workflow runs faster. Outputs look clean. Adoption numbers go up. And underneath, something structural is thinning.

By the time the financials reflect the damage, the architecture is already hollow.

The organisations that get into trouble aren’t the ones that deployed too slowly. They’re the ones that deployed quickly into an organisational structure that was never designed to support it.

The Question That Was Always Being Avoided

AI is forcing a question most organisations have never explicitly answered:

Who actually owns the work?

Not who is responsible in theory. Not whose name is on the org chart. Who owns each step, where decisions actually occur, and which tasks can be delegated — to anyone, human or otherwise.

For many companies, that question has been deferred for decades. Work evolved organically. Coordination happened through relationships and habits. Nobody needed to formalise it because humans compensated.

Agents don’t compensate. They execute.

Once AI participates in a workflow, ambiguity becomes a structural risk. Drift under acceleration compounds quickly. And the organisations that succeed won’t be the ones with the best models.

They’ll be the ones that finally answered the question.

Where to Start

Not with the technology.

Start with the work.

Map what actually exists. Document who owns each step. Define where judgment has to sit. Be explicit about which decisions can be delegated to AI and which must stay human, not as a policy statement, but as a designed constraint built into the workflow itself.

Then build your governance layer on top of that.

The organisations doing this well don’t look like AI adopters. They look like operating systems. Legible, intentional, capable of learning.

Because the real constraint in the AI era was never the technology.

It was always the clarity of the organisation running it. AI didn’t create that constraint. It just made it visible.

Across boardrooms, the same directive is circulating:

“AI should reduce our cost base.”

That mandate flows downward. CEOs set margin targets. Leadership teams inherit headcount numbers. Managers are told the size of their organization before anyone has asked what work actually needs to happen.

It gets called optimization. Most of the time, it is compression.

And compression is not transformation.

The Signal

A recent study from Anthropic introduced a useful metric: observed exposure. It measures the gap between what AI could theoretically automate and what people are actually using it for at work.

The conclusion is subtle but important. AI capabilities are already high across many knowledge tasks. But real adoption remains far below what the technology can theoretically do.

At first glance, this sounds reassuring. AI is not destroying jobs yet.

But this framing can still lead leaders to look in the wrong place because the transformation is not happening at the level of jobs. It is happening inside roles and workflows.

We are no longer in a knowledge economy. We are entering a capability economy. And most workforce transformation efforts have not caught up.

The Misunderstanding

Most readers interpret studies like this through a labor-market lens. Which jobs will disappear? Which jobs are safe?

But organizations do not operate at the level of occupations. They operate through workflows and capabilities. And that difference matters.

I recently observed this dynamic in a leadership discussion. Someone shared the Anthropic research as evidence that we should focus on which tasks AI can replace versus what humans should continue doing. At first glance, that sounds reasonable. But it quietly shifts the conversation in the wrong direction.

The discussion becomes: What tasks can AI do? What tasks should humans keep?

Instead of the more important question: What capability are we trying to build in the first place?

Because organizations do not exist to perform tasks. They exist to produce outcomes. A sales team is not hired to write emails. A product team is not hired to summarize documents. A consultant is not hired to produce slides. They are hired to exercise judgment in complex situations.

The moment the conversation moves from outcomes to tasks, the transformation has already gone off track.

The Real Shift

For twenty years, organizations were designed around knowledge — who has it, who produces it, who synthesizes it. Knowledge workers became the atomic unit of value creation.

AI has not eliminated that system. It has destabilized it.

Information retrieval, summarization, drafting, and structured analysis are becoming abundant. When the constraint changes, the architecture must change with it.

The knowledge economy optimized for information access. The capability economy optimizes for governed judgment.

The Wrong First Question

Under margin pressure, the instinct is understandable. Leaders ask: “How many roles can we automate?”

But cost correction is not AI strategy. It is accounting with better PR.

The right question is harder: What does this role look like once AI is embedded in the workflow?

That shift sounds subtle. It is not. It moves the conversation from headcount to work architecture, from elimination to recomposition, from expense reduction to capability design.

Three Structural Realities Leaders Are Misreading

1. Automation without workflow mapping creates fragility

You can automate first-pass insurance claims triage. You can accelerate legal research through retrieval and summarization. But risk framing still requires context. Edge cases still require interpretation. Accountability still requires a human who can be held to it.

When organizations eliminate roles before mapping workflows, they do not remove work. They destabilize it.

2. When AI handles production, humans must handle judgment

Once AI drafts, summarizes, and retrieves, the human role changes. Less recall. More evaluation. Less drafting. More interpretation.

The scarce asset is no longer information. It is governed judgment.

The people who thrive are those who can detect when outputs are strategically wrong, navigate real ambiguity, escalate decisions appropriately, and hold accountability under uncertainty. That is not a skill you train in a workshop. It is a capability you build through role design.

3. AI often raises skill thresholds rather than lowering them

This is where many organizations miscalculate. They assume AI lowers skill requirements. In many workflows, it does the opposite.

When junior employees no longer build pattern recognition through repetition, the apprenticeship layer compresses. When AI produces the first draft, the reviewer must possess stronger judgment, not weaker.

If organizations cut foundational roles without redesigning learning pathways, they do not gain efficiency. They create capability debt.

Capability debt is worth naming carefully: it is invisible until it is catastrophic. Faster outputs, clean margins, early efficiency gains, and underneath, shallower expertise, compressed learning cycles, eroding judgment. By the time financial performance reflects the damage, the architecture has already been hollowed out.

The Framework Leaders Actually Need

When I work with leadership teams on AI workforce transformation, we anchor on three questions. Not about tools. About roles.

1. Workflow Exposure — How much of this role’s workflow can be automated or augmented?

2. Capability Shift — What human capabilities become more valuable after augmentation?

3. Decision Authority — Who owns the judgment when AI participates in decisions?

AI does not remove roles. It reveals whether those roles were designed for execution or judgment.

What This Means for Leadership Teams

Once AI moves from experimentation into workflow integration, workforce transformation stops being an HR project. It becomes a leadership architecture question.

AI does not redistribute work randomly. It redistributes responsibility across the leadership system.

CEO / GM — Direction

The CEO’s role is not to pick tools. It is to define what the organization is optimizing for. Satya Nadella did not frame Microsoft’s AI investment as a productivity add-on — he framed it as embedding intelligence into the flow of work itself, a platform shift that redefines how customers create value. That directional clarity determines how every function below it responds.

COO — Workflow Architecture

The COO becomes the architect of work — mapping where AI compresses workflows, where judgment must remain, and where escalation happens. ServiceNow embedded AI inside existing incident management and HR workflows rather than deploying it as a standalone tool, creating a suggestion-action-review-learning loop that compounds over time. Without that deliberate redesign, AI does not create clarity. It accelerates friction.

CHRO — Capability Design

The CHRO’s challenge is structural: when AI handles the repetitive tasks that junior employees once performed, the apprenticeship layer compresses and expertise stops forming the way it used to. IBM responded by building explicit pathways into AI governance, model oversight, and workflow orchestration roles as execution tasks automated away. The question is not what AI can do now — it is how human expertise develops in an environment where AI does more of the work that used to build it.

CIO / CTO — System Reliability and Learning

Their responsibility is no longer just deploying AI tools — it is building systems that learn. Databricks’ acquisition of MosaicML was fundamentally about enabling organizations to integrate learning loops into their own platforms, making organizational judgment capturable rather than just organizational output faster. Infrastructure now has a second job: not just running models, but compounding the judgment humans apply when those models are wrong.

CFO — Value Discipline

The CFO’s role is to prevent AI from becoming productivity theater — activity that looks efficient but does not compound. When Salesforce introduced AI agents into sales workflows, reported gains were framed around pipeline generation and customer outcomes, not just cost efficiency. That accountability structure is the right one. The question is not whether the organization is moving faster. It is whether it is building capability that builds on itself.

The Leadership System Behind AI

AI transformation is often described as a technology transition. In practice, it is a leadership system transition.

The CEO defines direction. The COO redesigns workflows. The CHRO redesigns how capability forms. The CIO and CTO build systems that learn. The CFO enforces value discipline.

When those five functions are aligned, AI compounds. When they are not, each function optimizes locally and the system produces activity without architecture.

That is the real implementation risk. Not the technology. The leadership coherence around it.

The Leadership Question

Organizations that redesign roles will compound. Organizations that only cut will eventually discover something uncomfortable.

The question is no longer whether AI will reshape your workforce. The question is whether you are building for what’s compounding, or managing what’s declining.

That divergence is already underway.

The Directive That’s Driving the Wrong Decisions

Across boardrooms, the same directive is circulating: “AI should reduce our cost base.”

That mandate flows downward. CEOs set margin targets. Leadership teams inherit headcount numbers. Managers are handed an org size before anyone has asked what work actually needs to happen.

It gets called optimisation. Most of the time, it’s compression, and compression is not transformation.

We Are No Longer in a Knowledge Economy

For twenty years, organisations were built around knowledge, who has it, who produces it, who synthesises it. Knowledge workers were the atomic unit of value creation.

AI hasn’t eliminated that. It’s made it abundant.

Information retrieval, summarisation, structured analysis, once scarce, now cheap. When the constraint changes, the architecture has to change with it.

We are no longer in a knowledge economy. We are entering a capability economy.

Most transformation efforts haven’t caught up.

Cost-Cutting Is Not an AI Workforce Strategy

The instinct under margin pressure is understandable: how many roles can we automate? But cost correction is not AI strategy. It’s accounting with better PR.

The right question is harder:

What does this role actually look like once AI is embedded in the workflow?

That shift changes everything, not just headcount, but how work is structured, where decisions live, what you’re even hiring for.

Three Structural Realities Most Organisations Are Misreading

1. Automating Without Workflow Mapping Creates Fragility

You can automate first-pass claims triage. You can accelerate legal research. But risk framing still requires context. Edge cases still require interpretation. Accountability still requires someone who can be held to it.

When you eliminate before you map, you don’t remove work. You destabilise it.

2. When AI Handles Production, Humans Must Handle Judgment.

Less recall. More evaluation. Less drafting. More interpretation. The scarce asset is no longer information, it’s governed judgment.

The people who thrive aren’t the ones who know the most facts. They’re the ones who catch when an output is strategically wrong, hold their ground in ambiguous situations, and stay accountable when things get complicated.

That’s not something you teach in a workshop. It’s something you build through how you design roles.

3. AI Often Raises Skill Thresholds, It Doesn’t Lower Them.

This is where most organisations get it badly wrong. They assume AI lowers the bar.

When junior employees no longer build pattern recognition through repetition, the apprenticeship layer quietly disappears. When AI generates the first draft, the person reviewing it needs stronger judgment, not weaker. Cut foundational roles without redesigning the learning pathway and you don’t gain efficiency.

You accumulate capability debt. And it won’t show up on any dashboard until it’s too late to reverse.

How AI Workforce Failures Actually Happen

AI rarely fails loudly.

It fails through structural thinning. Faster outputs. Clean metrics. Early margin lift. And underneath: shallower expertise, compounding drift, judgment that’s slowly eroded. By the time the financials reflect the damage, the architecture is already hollow.

Automation scales execution. Judgment scales advantage.

The AI Role Readiness Lens: Three Questions That Change Everything

When I work with leadership teams on workforce transformation, we don’t start with tools. We don’t start with headcount targets.

We start with three questions about roles.

Question 1: Workflow Exposure

How much of this role’s workflow can be automated or augmented?

Some roles are execution-heavy and rules-based. Others live in ambiguity and run on judgment. The distinction matters a lot.

Insurance claims triage can automate first-pass categorisation, but edge cases still need a human who can read context. Legal research can move much faster, but the strategy still requires someone who understands what the client actually needs, not just what the documents say.

Mapping exposure before you make decisions about roles isn’t optional. It’s the foundation. Without it, you’re not transforming. You’re guessing.

Question 2 : Capability Shift

After augmentation, what capabilities matter more?

When AI takes over summarisation, drafting, retrieval, and structured analysis, the human role changes:

• Less emphasis on recall → more emphasis on evaluation

• Less production → more interpretation

• Less speed → more calibration

The people who do well in this environment are the ones who notice when an output is subtly wrong, who can work through situations the model wasn’t trained for, who bring organisational context a system can’t access, and who know when to escalate.

That’s capability. Not knowledge. And you can’t shortcut it with a reskilling programme.

Question 3: Seniority Calibration

Do you now need a different level of judgment to manage AI?

Most organisations skip this one. It’s the most consequential.

AI compresses the apprenticeship model. When junior analysts don’t build pattern recognition through repetition, you lose a learning layer that took years to develop. When AI handles the first draft, the reviewer needs to bring more, not less. The supervision layer gets thinner. The margin for error narrows.

That changes your talent architecture. Not just your headcount.

A Useful Parallel: What Mainframes Taught Us

When enterprises adopted mainframes, autonomy didn’t mean open access. It meant defined permissions, clear escalation paths, and governance controls built into the architecture from the start, not bolted on afterward.

Agentic AI is the same kind of inflection point.

As the autonomy of AI systems increases, the need for explicit authority structures, boundary definition, and human judgment calibration increases with it. The technology moves fast. But the constraint isn’t the technology.

It’s the architecture around it.

Building for the Economy That’s Emerging

The organisations getting stronger right now are not the ones cutting deepest. They’re the ones recomposing, figuring out where judgment has to sit, what seniority actually means when execution is automated, and how to rebuild the learning pathways they’re at risk of losing.

The knowledge economy rewarded access to information. The capability economy will reward the people and organisations that can govern what AI cannot.

The ones that only cut will find out, too late, that they optimised themselves out of their own future.

The question isn’t whether AI changes your workforce. It’s whether you’re asking the right questions before it does.

Tags: AI Workforce Transformation · Organisational Design · Future of Work · AI Role Redesign · Capability Economy · Leadership · AI Strategy · Agentic AI

The adults have entered the room.

If you look away from the consumer product launches and focus on the recent wave of M&A and strategic partnerships, you will see a refreshing pattern emerge. The market is stopping its obsession with “Magic” and starting to value “Maturity.”

We are finally moving past the phase of “AI as a novelty” and entering the phase of “AI as an operating system.”

The Return to Sanity

For the last two years, companies tried to buy “AI” as if it were a standalone ingredient, a sprinkle of magic dust to add to their software.

The good news is that this approach is dying. The recent deals we see signal a sophisticated convergence. Buyers are no longer purchasing isolated tools; they are building integrated systems.

They are realizing that three layers, previously treated as separate, must be designed together:

1. IT Infrastructure (The Plumbing)

2. Business Processes (The Rules)

3. AI Capabilities (The Engine)

When these three merge, you don’t get a chatbot. You get a Sovereign System.

3 Signals of Sanity: The M&A Evidence

We are seeing clear archetypes of “Good AI” emerging in the deal flow. These aren’t theoretical; they are the new standard for how value is created.

Signal 1: The “Context” Buy (e.g., Thomson Reuters + Casetext) Look at how the smartest incumbents are moving. Thomson Reuters didn’t just buy a “legal LLM.” They bought Casetext—a company that had spent years embedding legal judgment into its CoCounsel product. They weren’t buying the GPU power; they were buying the Tacit Knowledge of the law. The Signal: They realized that in high-stakes industries, you don’t buy the model; you buy the governance layer that wraps around it.

Signal 2: The “Workflow” Buy (e.g., SAP + WalkMe / ServiceNow) The enterprise giants are done with “bolt-on” AI.

• SAP’s acquisition of WalkMe wasn’t about digital adoption; it was about capturing the user context—understanding exactly how humans move through a process so the AI can navigate it correctly.

• ServiceNow has aggressively acquired talent and tech not to build a generic chatbot, but to create “Now Assist”—a system where the AI is fully constrained by the IT workflow. The Signal: These deals prove that AI is useless without the “Business Process” layer to direct it.

Signal 3: The “Factory” Buy (e.g., Databricks + MosaicML) This is the ultimate sovereignty play. Databricks didn’t buy a model API; they bought MosaicML—the infrastructure to build your own models. This allows enterprises to train models on their own proprietary data inside their own secure walls. The Signal: The market is moving away from “Renting Intelligence” (OpenAI wrappers) toward “Owning the Factory” (Sovereign AI).

The “Sovereignty Loop”: How We Win

The most encouraging shift is that investors and leaders are finally valuing what matters.

For a long time, value was measured by Speed (How fast can it write?). Now, the smartest players are measuring Sovereignty (How well does it learn from us?). They get their speed through alliances and M&A, but they build their value through governance.

We are seeing the rise of “The Sovereignty Loop.” This is the mechanism that separates a toy from an asset:

• The Old Way: AI makes a mistake → We ignore it → The value leaks.

• The New Way: AI makes a mistake → A human corrects it → The correction updates the weights → The firm gets smarter.

This is a massive win for human capital. It means that Human Judgment is not being automated away; it is becoming the most premium data source in the enterprise.

Why This Is The Right Direction

This shift should give every leader confidence.

• It means we don’t have to chase every new model release.

• It means our legacy knowledge (”Tacit Knowledge”) is actually our greatest competitive advantage.

• It means that the hard work of governance—defining decision rights, auditing outcomes, setting standards—is exactly what builds value.

“AI Growing Up” means we can stop panicking about the technology and start leading the system.

For investors, the signal is clear: Don’t look for the company with the flashiest demo. Look for the company that has integrated the Loop. For leaders, the mandate is simple: Stop trying to be an “AI Company.” Start being a “Learning Company.”

The hype is over. The work begins. And that is the best news we’ve heard in years.