I came home from the last of those conversations and started writing down what I actually heard, some of which I’m still processing. Here are five things that stuck, and where I agree or disagree with the room.

The agenda is unanimous, but no one agrees on the verdict.

If you ask any such room whether AI belongs in the top three priorities this year, you will get instant agreement. Ask the same room whether it’s working and the disagreement is apparent. One widely cited industry survey has 54 percent of leaders reporting positive AI returns. Another, run the same season, has 56 percent of CEOs saying AI produced neither more revenue nor lower costs over the past year. Both numbers came from real surveys of real executives. Neither is wrong, because they’re measuring different things at different companies with different definitions of what counts as a return.

I don’t think this is an AI immaturity problem that will fix itself with time. I think “is AI working” stopped being one question a while back. It’s five or six different questions wearing the same name, and most of the executives trying to answer it haven’t agreed internally on which question they’re actually answering.

Is the CIO an AI innovator, a gatekeeper, or both?

This is the one I discussed in every room, and it attracted the most disagreement. Plenty of my peers describe their current posture as balancing innovation against discipline, playing both the person who pushes AI into the business and the person who makes sure it doesn’t get ahead of itself. I used to think that balance was a leadership skill worth admiring. I’ve come around to a less flattering explanation as I argued more and more about it.

One recent survey found that nearly a quarter of CIOs don’t know which department is actually accountable for meeting AI’s goals beyond the technology function. A separate study of more than a thousand C-suite executives found that only 2 percent of organizations have made their CFO accountable for AI’s financial value, and in the small group where they have, three out of four report the AI is delivering real value. The other 98 percent are, in effect, running the experiment without an owner, and the problem gets worse where there’s a standalone AI function sitting apart from the central technology function. The pattern isn’t just a technology problem, it’s an ownership one. A recent benchmark found fewer than four in ten large companies have even appointed a chief AI officer parallel to the CIO, and there’s almost no agreement on who that person reports to, or whether the outcome truly sits with them beyond the trophy title.

Without true ownership of the outcome, the CIO defaults into the gatekeeper role, whether they want it or not, and that burden quietly undermines the innovator instinct they’d rather be leading with. The peers I respect most skipped the balancing act and just forced an explicit owner onto the org chart, in writing, and aligned it with rest of executive team and their board.

The vendor pitch is selling confidence the buyer hasn’t built the infrastructure for.

This is where I might part ways with some of the room, and it’s worth saying directly. Three out of four CIOs in one recent survey admitted regret over at least one major AI vendor or platform decision made in the past year and a half. Six in ten said their own CEO had questioned that decision. Gartner expects more than 40 percent of agentic AI projects to be shut down by 2027 over cost and unclear value, and projects that only a small fraction of the vendors calling themselves agentic actually are, the rest having relabeled older chatbot and automation tools with a new word.

Some of my peers frame this as a vendor discipline problem, and to be fair, plenty of vendors are pitching capability beyond what an average buyer’s data and governance can actually support. The latest trend in vendor pricing ties fees to measured outcomes instead of seats or usage. Outcome pricing sounds like accountability until you notice it gives the advantage to whoever can actually measure the outcome, and in most companies that still isn’t the business stakeholder. I think that’s the same gap from the point above. You can’t hold a vendor to an outcome you haven’t defined ownership internally, or one you’ve never truly measured yourself.

Build versus buy is a question about where your data and workflow live, not a bragging point at a dinner.

I’ll keep this one short because it deserves its own piece later. The argument I hear most often, build your own AI capability versus buy it off the shelf, turns out to be three different choices converging into one. You can buy AI embedded in a platform you already run, buy the platform’s underlying AI layer and build your own tools on top of it, or connect an outside model directly to your own systems and take full responsibility for what comes out. Those three paths make different assumptions about where your data sits, how you affect workflow, and who governs it. It’s currently a very passionate strategy conversation, and no one path is better than the other. There’s more to say here about compute, vendors, and talent as a sourcing problem. I’ll save it for next time.

Talent is still the ceiling, and it’s the one thing no vendor can solve for you.

Forty percent of CIOs still name a lack of in-house talent as their top obstacle, ahead of budget and ahead of the technology itself. That tracks with what I heard this summer. Every AI capability that’s become genuinely accessible this year, and there are a lot of them, still needs someone inside the building who understands both the technology and the actual problem well enough to know when the output is wrong. We need to own our forward-deployed engineers; they must be cultivated within both the business and the technology function.

One of the pieces I read going into these conversations states: the CIO’s job used to be having the deepest technical answer in the room, and now it’s building a room where the right answer can come from anywhere, including from someone junior enough to still be willing to say something isn’t working, almost like a NASA launch command center where every voice has a critical role in the mission’s success. No vendor platform gives us that.

Heading into fall

I went into this summer expecting the usual complaints about budget and vendor noise. What I actually heard was a room full of serious people quietly renegotiating who’s responsible for what, without describing it that way. I don’t think that’s a bad place to be. It’s an earlier stage of the adoption that gets missed in the media and vendor hype.

The question I’m carrying into fall isn’t whether AI is working. It’s whether anyone in an organization could answer, without hesitating, who owns the answer to that question.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

The latest arrived in late July, when Jensen Huang posted on X for the first time in his life. Not a product launch but a policy letter titled Open Weights and American AI Leadership, arguing that open models strengthen safety and cybersecurity, accelerate innovation, and enable sovereignty. It launched with 25 signatures. By the time I read it, that number was 235.

The list of participants on that letter is worth a look. Chipmakers, cloud providers, inference companies, model labs, developer tools, venture firms. The operating companies on it are technology companies. There is no bank, no hospital system, no insurer, no utility, no retailer. Every organization on that list is telling us how enterprises should buy AI. Ironically, each one of them is in the business of selling it.

The arguments in the letter are quite sound, but it is still a lobbying document written for Washington, and it is being translated in executive meetings as the new AI manifesto. An operational and technical strategy is hard to form when all inputs are coming from suppliers, and they keep changing their positions.

Open source won the server room, but a profession had to be built first

When open source started winning the data center race, almost nobody inside a large enterprise ran the free version. They bought Red Hat, and what they paid for was a subscription, support, certification, and someone to call at two in the morning. Microsoft spent those years arguing that Windows Server cost less once you counted staffing, and for a great many enterprises that was true for a long time. The license was never the expensive part.

What matters now with AI is what had to happen first before we all go down the open weight path. Red Hat launched its Certified Engineer program in January 1999, built around a performance-based exam, and within a decade over 35,000 people held it. Linux did not become enterprise-ready because the code improved. It became enterprise-ready when enough people existed who could make it behave in production.

However, there is one difference worth noticing. You could read Linux source code. Nobody can read a set of weights.

The skill almost nobody has

Downloading a model takes minutes. Making it perform reliably against a specific business problem is a craft, and a very scarce one. The people who can do it are what my HR friends call purple squirrels.

It means deciding what to tune and what to leave alone, building evaluation sets that tell you whether a change helped or only appeared to, compressing a model to fit hardware you own without degrading its answers, and catching model drift on a workload before anyone downstream gets impacted. Very few large institutions have someone who can do all of that, and those who can are being bid on by companies that pay on a different scale than a health system or a regional insurer.

In Mozilla’s developer survey earlier this year, teams working with open models reached production about 53% of the time against 63% for closed. The gap widens as the surveyed companies get larger, with closed climbing toward 73% at the biggest organizations while open barely moves. Organizational scale usually helps in such scenarios. It does not help here, because the binding constraint is a scarce and unglamorous skill rather than money.

The impact of open models matters for the enterprise because the ROI is quite material. Frank Nagle and Daniel Yue found closed models taking roughly 80% of usage and 96% of revenue while costing about 6 times more than open models for comparable work. They concluded the adoption of open models is poor because of switching costs and information frictions, not because organizations are unaware of the alternatives. The savings are untapped because most enterprises do not have the talent to run the cheaper, open model.

What this letter actually means for you

Nothing you have already built is invalidated. What has changed is that choosing a model, or a vendor partnership, stops being a permanent decision and becomes something you architect to switch, based on workload, budget, and talent.

If you have a frontier partnership, you should keep it. That supplier is still the only party who signs anything, which is worth paying for. But the mistake would be letting a commercial relationship harden into a technical architecture.

Capture the price advantage with open models where it is safe, typically through managed endpoints rather than hardware you must find super talent for. Avoid the temptation to build your own GPU fleet unless something other than the cost justifies it.

Then start hiring, slowly and deliberately, for the skill above. It needs to be a small team, wrapped around your core data and platform teams.

Steady hands

Underneath all of this is the key question: how does any organization commit to a multi-year direction when its suppliers change the product, the price, the support model, and the value proposition every few weeks?

The leader at the front of these changes should have the discipline of separating supplier movement from the organization’s own constraints, whether those are technology debt, budget, talent, or simply stakeholder readiness. June 12th showed a constraint nobody had imagined before. A government directive reached an AI lab at 5:21pm and a frontier model went dark for every customer that same day. It stayed dark for 19 days. Imagine your agentic workforce disappearing for that long. More of this is ahead of us: innovation, supplier movement, government intervention, and the dizziness they cause. A company that rewrites its strategy on every announcement does not have a strategy.

You are not being asked to bet the organization on which model wins. That contest belongs to your suppliers and nothing you do will influence it. You are being asked whether your organization could absorb a change of models if one arrived. That is a bet you control, and the things that pay off are identical in every version of the future: people who can operate models, the ability to move a workload without rebuilding it, and clarity about which work must never leave your walls. Fund those three and you can let the model wars play out without you.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

I’ve asked a version of the same question from the buyer’s side, sitting across from a vendor renewal and wondering how much of the invoice was paying for expertise I no longer needed to rent. That instinct is showing up across the market as a genuine insourcing conversation, not just a cost-cutting one. KPMG’s most recent pulse survey found the share of enterprises with agents already integrated into their workflows nearly quadrupled in about a year, and a majority now say they favor building agent capability in-house alongside whatever they still buy. (KPMG)

That instinct deserves a check before you follow it too far. Cheaper to produce is not the same thing as cheaper to buy. Research on automated pricing shows that faster execution doesn’t reliably lower what the buyer pays; it can support a firm price just as easily as a falling one, depending on how much competition sits around the transaction. (NBER) Most of us paid intermediaries for years without pricing out what we were actually buying, because nobody had to. AI is starting to force that itemizing, and what an intermediary is worth now depends on what was inside the bundle.

An intermediary is a bundle of functions, not one business

Set the industry language aside and an intermediary is an organization that sits between two parties because dealing directly would be slower, riskier, or harder to govern than dealing through a trusted third party, the way an insurance broker exists because a business can’t easily tell which insurer will actually pay a claim, or price its own risk as well as someone who prices it across thousands of similar businesses. That’s the shape of almost every intermediary business.

Treating “the intermediary” as one category hides the fact that each of these businesses is really several distinct jobs stacked together: finding and matching the two parties, translating between formats and systems, executing the routine steps, verifying that everything is legitimate, absorbing the risk if something goes wrong, and owning the relationship when a customer needs to be made whole.

AI does not touch these jobs evenly. It’s very good at the first three, far less capable of the fourth and fifth, and it can’t touch the sixth at all, because owning a relationship and accepting responsibility for a bad outcome aren’t computational problems. It’s also worth resisting the habit of calling all of this “friction” and assuming AI should remove it on sight. Some of it is waste, like duplicate data entry. Some of it is the opposite of waste, like authorization and fraud screening. AI eliminates the first kind well. It may make the second kind more necessary once agents start acting on their own.

That distinction is easiest to see with a simple equation. The cost of a transaction isn’t really one number:

Cost per successful outcome = (execution + exceptions + fraud losses + compliance + capital held + liability) ÷ successful, compliant outcomes

AI is shrinking the execution term, sometimes sharply. It’s doing far less to the terms after it, and it can grow them, since cheaper execution tends to produce more volume, and more volume produces more exceptions and more fraud to catch. An intermediary whose value sat entirely in execution is genuinely exposed. One whose fee was always covering the rest of that equation, whether it knew it or not, has more room than the excitement suggests.

Travel already ran this experiment once

Travel is the clearest proof this isn’t theoretical, because everyone has lived through it personally. Twenty-some years ago the internet was supposed to eliminate the travel agent, and it largely did. What it didn’t eliminate was the toll booth: the fee still collected, somewhere in the transaction, by whoever the traveler had to trust to hold the booking, the payment, and the promise of what happens if something goes wrong. Online travel agencies absorbed the search and booking function and became that new toll booth, not by beating the airlines on price, but by owning a single trusted place where a traveler’s loyalty history, payment details, and cancellation rights all lived.

Travel is now running the same experiment a second time, in real time. Google and OpenAI have both rolled out agentic booking flows this year that let a traveler plan and compare a trip entirely inside a chat window. The interesting detail is what happens at the moment of paying: in both cases, the booking still completes on the travel platform’s own systems, not inside the chat, because that’s where the payment relationship, the loyalty account, and the promise of what happens if the flight is cancelled still live. Expedia and Booking.com are betting, explicitly, that travelers will trust an AI to plan a trip but will still want a familiar name holding the reservation when something goes wrong. Whether that bet holds is genuinely unresolved: the Skift reporting on it describes a scramble to win the trust of AI agents, not a victory lap. But the shape of the fight tells you where the value moved the first time, and it wasn’t to the interface.

Payments show the same instinct, one layer down

I look to the finance industry more than any other to see how a new technology actually gets made useful to an organization’s goals before anyone else has worked it out, and payments are showing the same instinct as travel, just a few years behind. Card networks were supposed to be among the most exposed intermediaries of all once AI agents started transacting on people’s behalf. Instead, Visa’s Intelligent Commerce Connect and Mastercard’s Agent Pay for Machines, both launched this year, aren’t built to make the rails cheaper. They’re built to credential which agents can spend money, enforce spending limits, and guarantee settlement across cards, accounts, and even stablecoins. That’s not a company defending its processing fee. That’s a company repositioning itself as the entity that decides whether an autonomous purchase was authorized at all, and who eats the loss if it wasn’t. The toll booth didn’t disappear. It moved from the checkout page to the authorization layer, and it may be worth more there than it ever was at the point of sale.

The honest counterexample, and the tension underneath it

None of this should read as an argument that intermediaries are safe. Call centers and clerical back-office processing are the honest counterexample, and the industry is naming what’s happening plainly. Vendors and buyers are openly discussing what’s being called Business Process Insourcing: bringing offshore, execution-only work back inside the company and running it through an agent instead of a vendor’s headcount. A 2026 industry panel at SSON spent an entire session debating whether traditional BPO even fits the agentic era, and the honest answer from several providers on stage was that it has to become something else to survive.

The reason this domain breaks the other way is structural. Much of the value a call center or clerical BPO firm charged for sat almost entirely in labor arbitrage on routine execution, exactly the function AI is best at replacing, with no license, balance sheet, or liability underneath it to retreat to once the labor got cheap.

But it would be too neat to call this a clean story about insourcing winning. Two different costs are falling at once, and it isn’t obvious which is falling faster: the cost of managing an external vendor, and the cost of coordinating that same work internally, since the skill a company once needed a specialist vendor to supply is increasingly something any competent internal team can operate directly. Economists have long argued the boundary of a firm sits wherever those two costs cross, and the evidence on which side is moving faster is genuinely split. One 2026 study found firms sharply cutting managerial hiring after adopting generative AI; an earlier study of the predictive-AI era found the opposite. (Academy of Management; Strategic Management Journal) This is being fought out function by function inside real companies right now, not settled by a single trend line.

Which layers will matter, beyond these three

Healthcare claims is a good fourth check, because it shows the pattern splitting inside a single value chain rather than across separate industries. Regulators are standardizing the transmission layer: new CMS rules require faster, more transparent prior-authorization decisions and open APIs starting in 2027. (CMS) That should make transmitting a claim close to commoditized. But the capital moving into the industry is going the other way: Waystar processed more than $2.4 trillion in gross claims last year and says every transaction sharpens its proprietary denial-prediction models, and Carlyle’s 2026 acquisition combining Knack RCM and EqualizeRCM was priced around scale and complex financial risk, not headcount reduction. (Waystar; Carlyle) The transport layer is commoditizing. The layer that decides whether a claim actually gets paid is consolidating.

That’s the rule worth carrying into any other industry. If a fee was tied to moving, matching, or translating something, treat that revenue as exposed. If it was tied to guaranteeing something, absorbing a loss, or holding a license or relationship nobody else can easily replicate, that revenue has real room to hold, and possibly to grow.

What this means on both sides of the table

For an intermediary, the task is simple to state and hard to do: know which part of your fee pays for something scarce, a license, a balance sheet, years of exception-handling experience, a relationship customers trust, and which part just pays for labor. Once you know the difference, say it plainly. Stop selling speed and start selling the guarantee: tell customers exactly what you stand behind when something goes wrong, and price for that, not for the task.

For the enterprise on the other side, the make-or-buy question is genuinely open right now, not a one-way march toward insourcing everything cheap AI can touch. The useful discipline is checking that, function by function, on a regular basis, instead of waiting for the next contract renewal to force it: does this work need real institutional standing, or is it just labor that happens to run through a vendor today. Get it wrong, and you either keep paying for something you could do yourself, or take on a risk you aren’t set up to carry.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

Without anyone naming it, that company has taken on two dependencies it has no function to manage. Its operations increasingly run on a purchased input whose price it does not control and has not hedged. And the thing that makes those agents good at its business, the accumulated judgment of its people, is quietly becoming raw material for the vendors it buys from. Neither looks like a supply chain to a company that thinks it is buying software. Unfortunately, in the AI-enabled world, they both are.

I wrote earlier this year that AI is a capital decision, not a budget line, and that returns are downstream of the digital foundation you build it on. The important question is what that capital actually buys. Look closely, and the enterprise’s exposure runs in two directions at once.

The commodity you are buying at spot price

Start with the input everyone can see. Something shifted this year in the way serious capital talks about compute. At the Milken Institute this spring, Larry Fink suggested that computing power could become a tradable asset class of its own, with futures contracts priced the way markets already price energy and grain, because the world is short on chips, memory, and power. A few weeks later, OpenAI offered an entire Y Combinator cohort millions of dollars each, not in cash but in tokens: compute for equity. Then Nvidia’s CFO, Colette Kress, described the company’s new arrangement to backstop GPU financing for cloud partners in exchange for a recurring cut of the revenue those chips generate, calling it a “usage-linked earnings stream.”

Read individually, these are three unrelated headlines. Read together, they carry one message: the people closest to the economics of compute have stopped treating it as a service you rent and started treating it as a commodity you finance, hedge, and lock in. Some of this activity may be a bubble; Sequoia’s David Cahn has argued the revenue needed to justify the buildout runs to hundreds of billions a year, and the circular financing between chipmakers and labs mimics dot-com vendor finance. But the signal is true either way, because you do not build a futures market for something abundant. When the world’s largest asset manager, the leading model lab, and the dominant chipmaker all move the same direction inside two months, the market is pricing a raw material.

This matters on a P&L. Epoch AI estimates that GPU-hour prices rose roughly 30% this year as the frontier labs consolidated supply. Any operator who has bought a volatile input on the open market knows what follows: show up with no contract, no forecast, and no hedge, and you pay the premium, and the premium lands in your margin. It is what jet fuel does to an airline, or a thin formulary to a hospital pharmacy. The sourcing discipline capital-intensive industries built for those inputs exists for one reason: a business whose operations depend on a commodity cannot afford to buy it by accident.

Enterprises are already behaving like commodity buyers without admitting it. Over the past year, US companies routed close to half of their API traffic to cheaper open-weight models from firms like Alibaba and DeepSeek, purely to protect cost. That is the oldest move in any commodity market, but it is happening one engineering team at a time rather than as a strategy anyone owns. A discipline is forming around this: the share of FinOps practitioners managing AI spend jumped from 31% to 98% in two years. The question for a non-tech company is no longer whether compute is a commodity it depends on. It is whether anyone in the building is sourcing it like one, or whether the premium is simply leaking out of the margin, every quarter, as the agent fleet grows.

The supply chain running the other way

Now turn the picture around, because the more consequential exposure flows out of the enterprise rather than into it.

When Alex Karp told CNBC that enterprises are “paying for tokens that create no value” while the labs are “stealing the weights and alpha” of their business, it was easy to dismiss as a competitor talking. Satya Nadella made a cooler version of the same point at Stanford, asking how a company that is only a consumer of a foundation model can retain enterprise value, let alone create it. Strip away the theater and there is a real mechanism underneath, and Everett Randle of Benchmark has given it a name: the task economy.

Here is the distinction that most executives have not yet absorbed. Tokens measure how much AI gets used. Tasks measure how AI gets better. For years the labs improved their models by training on the public internet, and they have now largely exhausted it. The next increment of capability does not come from more text scraped from the web. It comes from watching an expert do real work and grading the result: give the model an actual problem, let a professional judge whether it got the answer right, and feed that judgment back into training. A task, in this sense, is a unit of instruction. It is how a model learns the things the internet never wrote down.

And the things the internet never wrote down are precisely what make your business your business. How a claims adjuster weighs an ambiguous injury report. How a plant manager reads a vibration signature and decides whether to shut a line down. How a merchandiser prices a slow-moving SKU three weeks before a holiday. This tacit, hard-won operating judgment is exactly what frontier labs now compete to capture, because generic intelligence is abundant and domain judgment is scarce. An entire industry has grown up to broker it. Mercor, the leading platform, went from one million to two billion dollars in annualized revenue in twenty-four months by paying tens of thousands of experts to convert their professional judgment into graded training data. When Randle argues that data will be the next trillion-dollar category, this is the raw material he means, and much of it lives inside your walls.

This is where an abstract IP concern becomes a concrete governance failure, and it is the part that should hold the attention of executives and boards. This spring Mercor disclosed a breach, and reporting indicates the exposure touched not only contractor data but material connected to the companies whose work those contractors were training models on. The finer point, setting aside the real labor questions, is that the moment your experts’ judgment begins flowing into a model, whether through a vendor, a shadow-AI experiment some team started quietly, or the ordinary use of an agent trained on your workflows, you have acquired a data-governance exposure that most organizations have never inventoried, priced, or assigned to anyone. Karp’s complaint, stripped of its salesmanship, describes something that has already produced a security incident. You do not have to take his side to take the mechanism seriously.

What this does and does not solve

Let me try to make this argument clearly. Neither of these disciplines will fix AI’s return problem on its own. Most of the value failure enterprises are seeing is organizational, not a sourcing or governance gap: roughly 95% of generative-AI pilots deliver no measurable P&L impact, and a company with a perfect hedging strategy on a fragile digital foundation will still be disappointed. So these two supply chains are not the master key to AI value. They are two real, measurable, currently ungoverned exposures that grow as agentic workflows move from pilot to production, and in most non-tech enterprises no one has been asked to own either one.

The reassuring part is that neither requires a new theory of the firm. Every capital-intensive enterprise already knows how to manage a critical input by forecasting demand, contracting ahead, and diversifying suppliers so none can hold it hostage; every serious enterprise already knows how to govern what leaves its walls. What is new is only that compute and expertise now belong in scope for both.

If you sit on a board, run a company, or hold the capital in a privately held one, the two questions to put on the table are simple, and neither has “IT will handle it” as an acceptable answer. Who is the named, accountable owner of our compute sourcing strategy? And who owns the policy on what happens to our people’s expertise the moment an agent, or a vendor training one, begins to learn from it? The uncomfortable answer in most companies is that no one owns either, because both fall in the seams between functions. Sourcing sits between technology, finance, and procurement. Protecting expertise sits between operations, legal, and HR. So the work goes undone, quarter after quarter, while operations grow more dependent on a raw material no single leader has been asked to secure.

Notice what kind of person those questions call for. Not a model expert, and not a procurement clerk. Someone who can sit in a capital committee and argue for a multi-year sourcing commitment, then walk into an operating review and explain which of the company’s judgment is worth protecting and which is safe to expose. That blend of capital fluency and operating instinct is rare, because for most of the last decade it did not need to exist. But it does now. The companies that come out of this era ahead will not be the ones that talked the most about AI. They will be the ones that treated compute and expertise as the raw materials they have quietly become, and put someone credible in charge of both before the cost of not doing so showed up in the numbers.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

In June, Meta committed $115 million to America’s Workforce Academy, which trains electricians, welders, and fiber technicians for free and guarantees each graduate a job. BlackRock has pledged $100 million to train fifty thousand tradespeople over five years, with Larry Fink arguing that capital alone won’t build the country’s future. A venture called Skill Factory is opening its first campus in Texas. It reads as a statement of values — a rediscovery of the dignity of working with your hands — but I don’t think that’s what it is. Meta has committed more than $600 billion to AI infrastructure by 2028, and none of that concrete pours itself. You can sign machine-learning researchers to nine-figure packages, but you cannot summon the electricians who wire the buildings they’ll work in, and there aren’t enough of them. So Meta is building its own supply — paying to clear a bottleneck it can see coming, on the one input its capital can’t buy fast enough.

What the Money Skips

The strange part isn’t where this capital is going; it’s what it steps around. In the same months Meta was guaranteeing jobs to welders, the bottom rung of white-collar work was being pulled up behind the people already on it. Dario Amodei has argued that AI could erase as much as half of entry-level office roles within a few years. He’s since softened that, and it was always more forecast than fact — but the direction shows up in the hiring data, where new-graduate roles in tech have fallen well below pre-pandemic levels. The work that goes first is the work that used to train people: the first-pass analysis, the document review, the model built by hand to check whether the numbers hold. I’ve watched it inside several tech shops — the tasks once given to a first-year to learn on are increasingly the ones handed to a tool.

That’s the asymmetry that should bother anyone running an institution. Physical work is getting a rebuilt way in — paid for, with a job at the end — because a hyperscaler has a concrete reason to fund it. Knowledge work is losing its way in at the same moment, and no one is rebuilding it, because no single employer has an obvious private reason to. A contractor can let Meta train the welders and hire them. A law firm, a bank, a health system cannot ask anyone else to raise its analysts.

I wrote a while back that the first rung of the professional ladder was moving. This is what came after: we’re rebuilding it on one side of the economy and letting it rot on the other, and the line between them is capital structure, not merit.

The Employer Was Always the University

None of this is new, which is the part that keeps getting lost — an industry built the answer a generation ago, and I came up inside it. When I started out in India, the IT services firms faced the same gap the AI economy is now opening: the country was turning out engineers by the hundreds of thousands, and a degree told you someone was capable of the work without having taught them to do it. Nobody waited for the universities to fix themselves; the employers built their own. Infosys put up a campus in Mysore that is still among the largest corporate training centers anywhere — some twenty thousand new hires a year move through a residential program of nineteen to twenty-three weeks, at roughly eight thousand dollars a head. TCS, Wipro, and several others ran their own similar programs. The assumption underneath was very clear: readiness is something you build, not something you assume the diploma delivered.

The American economy already has pieces of this, though it treats them as perks rather than infrastructure — Deloitte University, KPMG’s Lakehouse, the onboarding academies the big consulting firms run. Meta’s trades program is the same instinct aimed at a different worker. When the outside pipeline stops sending people who are ready, the work of making them ready moves back inside the company, and the employer becomes the university again. I won’t oversell it — it’s expensive, some of the people you train will leave, and the Mysore math leans on scale most American firms can’t reproduce. The example doesn’t transfer as a balance sheet; it transfers as a conviction about whose job it is to make people employable.

Where Judgment Comes From

There’s a stronger objection that maybe the knowledge-work rung doesn’t need rebuilding, because AI lifts it rather than removing it. The starting point becomes directing the models instead of doing the rote work, and people pick it up on the tools as they go. The firms already hiring experienced people to orchestrate AI can look like proof that the rung moved up, not away.

That’s right about the tasks and too quick about the people. Directing the work well still runs on judgment, and judgment gets built by doing the rote work first. The analyst who can glance at a model and feel that a number is off earned that instinct building a few hundred models by hand. Automate the workflow without putting right harnesses in place, you can have people who prompt fluently and can’t tell when the answer is wrong. The scarce thing here was never intelligence, which gets cheaper by the month; it’s the earned judgment that used to accumulate on the way up, and the climb that produced it is exactly what we’re automating. So it has to be built deliberately now, and the one with the reason and the vantage is usually the employer.

The Bench You Aren’t Funding

For anyone running an enterprise, this turns a cost decision into a succession one. Cutting the junior class flatters this year’s margin and quietly thins the bench you’ll draw leaders from a decade out, because you can’t make a senior operator who was never a junior one. It’s a slow, compounding risk that loses every budget argument to the quarter in front of it — and it loses quietly, which is what makes it easy to miss. Holding a ten-year horizon against a twelve-month one is precisely what a board is for.

Where I’d push anyone who asks is against treating this as a reason to stand up a workforce training department and call it handled. The Meta academy works because it’s built around the work Meta needs done. The version inside a bank or a hospital won’t be a campus with a logo on it; it will be a deliberate, employer-owned way of forming people, wired into how the work itself is changing as AI absorbs its routine parts. If the machine takes the first draft, the first pass, the first model, then the apprenticeship people used to serve by grinding through those has to be rebuilt around what’s left — the judgment, the exceptions, the calls only a person can make. Most organizations will have to become, in some unglamorous sense, their own university: not because it’s fashionable, but because the outside world stopped supplying the graduates they need, and the work changed underneath them while they watched the quarter.

Meta has decided to fund its own bench. Every other institution will face the same decision, and most haven’t yet noticed it’s already on the table.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

The enterprise version is less cinematic, but the impact is real. “Agents are coming” should not lead executives to ask only which AI assistant they have deployed or whether someone has stood up an MCP server. Those questions sit too low.

The harder question is whether the organization can be understood, trusted, and acted upon by machines. I recently wrote about AI becoming the new front door of the business: the website gave companies presence, the app gave them proximity, and AI gives agency. This goes deeper. If a customer’s agent arrives at your business, what does it actually encounter?

Tool rollout is not agent readiness

Many organizations will confuse access with readiness. They will roll out enterprise ChatGPT or Claude, enable copilots, connect a few tools, and report that the organization is preparing for agents. Some of that work will create value, but it does not make the institution agent-ready.

Agent readiness requires the organization to make its operating model readable. An agent can only act well if it can answer basic questions. What does this company know? Which source is authoritative? What policy applies? Who can approve an exception? Which data is permitted? Where does recommendation end and decision begin?

Most enterprises do not have clean answers. They have policy documents, wikis, dashboards, email threads, shared drives, data catalogs, contracts, and a few people who know how things really work. Humans often navigate that mess through judgment and relationships. An agent will not have that advantage. It will either stop, guess, or act confidently inside the wrong boundary.

That is why the knowledge-layer capability will matter. Google’s Open Knowledge Format is an early signal: a vendor-neutral, human- and agent-readable way to represent curated knowledge using directories of markdown files with metadata (which are simple text documents not code). The important idea is not markdown. It is that institutional context needs to become explicit, maintained, portable, and usable beyond the application that created it.

The enterprise needs rails

The serious preparation is not a single platform decision. It is the construction of rails. A knowledge rail tells an agent where authoritative context lives. A trust rail verifies whether the agent is legitimate and whom it represents. A permission rail defines what it can see, change, initiate, or recommend. A payment rail allows machines to transact safely. A provenance rail preserves source and lineage. An escalation rail defines when judgment returns to a human.

These rails are operating controls. A customer’s agent comparing products needs structured attributes, current inventory, price, terms, and trust signals. An internal agent preparing a decision memo needs to know which metric definition is current and which source wins when systems disagree. An agent initiating a payment needs evidence of intent, limits, authorization, and auditability.

The market is already moving in this direction. Cloudflare has introduced signed agents, where bots controlled by end users can be verified through cryptographic signatures. Visa’s Trusted Agent Protocol is designed to help merchants distinguish legitimate AI agents from malicious bots, and Google has moved its Agent Payments Protocol to the FIDO Alliance with an explicit emphasis on platform-agnostic, community-led standards for secure agentic payments. These are early examples, not final answers, but the pattern is clear: agents need identity, authorization, intent, and accountability.

This is why MCP is useful but insufficient. The Model Context Protocol helps standardize how AI systems connect to tools and data sources. That plumbing is important but plumbing does not decide what is true, sensitive, recently changed, or when the machine should stop. A company can expose tools through MCP and still lack a coherent theory of agent behavior.

Machine traffic changes the operating assumption

The web was built around human attention. A person searched, clicked, read, compared, and bought. The digital enterprise (like all e-commerce) was designed around that behavior: pages, forms, carts, funnels, conversion rates, and session analytics.

That assumption is now under pressure. Cloudflare’s 2025 Radar analysis showed that non-AI bots began the year responsible for roughly half of HTML page requests, while AI bots accounted for a smaller but growing share. The scope matters: this is HTML request traffic, not all internet activity. Still, machines are becoming a material audience for the enterprise.

A human shopping for a camera may visit five sites. An agent can compare thousands. A human may tolerate vague copy, inconsistent terminology, or a few extra clicks. An agent will look for structured facts, trusted signals, current inventory, price, policy, and a transaction path. It will not care about brand language if the business cannot be parsed.

That does not mean humans disappear. It means companies must increasingly serve two audiences at once: the person and the person’s agent.

Platform neutrality is not optional

One mistake right now is to hardwire the organization too deeply into today’s AI platform winners. OpenAI, Anthropic, Google, Microsoft, and others will continue to move quickly. Models will leapfrog, agent harnesses will change, and some protocols will disappear.

For that reason, the durable work is not picking the perfect agent platform. It is making the enterprise portable across platforms. A company should be able to move its institutional context, definitions, policies, permissions, service logic, and transaction rules across models and agent environments. If the operating model only works inside one vendor’s surface, the enterprise has not become agent-ready; it has become vendor-dependent.

The company should preserve the ability to adopt better models, negotiate better economics, meet regulatory requirements, and maintain control over its own knowledge and customer relationships.

The CIO’s strategic moment

This is where the CIO or CTO role becomes more important, not less. The easy version of AI leadership is procurement-led: buy the tool, enable access, count usage, collect anecdotes, and report momentum. That may be necessary, but it is not sufficient. It gets the organization using AI. It does not make the organization ready for agents.

The harder work is institutional. Someone has to help the enterprise decide what knowledge must be memorialized, what policies must become computable, what trust boundaries must be explicit, which workflows can safely become agentic, and where human judgment must remain non-negotiable. That work cannot sit inside technology alone. It touches legal, finance, risk, compliance, cybersecurity, operations, customer experience, and the front line.

It will also expose organizational debt. Conflicting definitions, unclear decision rights, undocumented exceptions, weak data ownership, and informal approvals become visible when a machine is asked to act. A well functioning agent will force the organization to encode the institutional friction.

For investors, this is an important part of the opportunity. The value will not only be in the AI platforms themselves. A large pool of value will be created in helping organizations build the rails around those platforms: knowledge layers, trust infrastructure, permission models, provenance, workflow controls, and safe transaction paths. This market is still early and will change quickly, but the implementation need is real. Most enterprises will not become agent-ready by buying access to a model. They will need help making their operating models legible to machines.

That is the executive takeaway. Stop asking which AI tools have been deployed. Ask where an agent would fail inside the business today. Where would it lack context? Where would it find two definitions for the same metric? Where would it be unable to verify authority? Where would consent, payment, or escalation break down? Where would it act too slowly because the enterprise is fragmented, or too quickly because the controls are weak?

That map is better than another demo. The companies that are ready for agents will not be the ones with the most pilots. They will be the ones whose knowledge, trust boundaries, permissions, and transaction logic are explicit enough for machines to use and governed enough for people to trust.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

Most of it is framed around loss: which jobs will disappear, which functions will shrink, which tasks will become automated. That is a necessary conversation, but it is not a sufficient one. It treats technology mostly as a substitute for labor, when history suggests something more complicated.

A major technology shift does not simply remove work from the economy. It changes where the work is needed.

When the automobile replaced the horse, it did not merely eliminate stable hands, carriage makers, and blacksmiths. It created demand for roads, mechanics, gas stations, traffic systems, auto insurance, vehicle financing, motels, oil refining, rubber, steel, logistics networks, and eventually suburbs. The first-order story was substitution. The second-order story was an entirely new economic system built around mobility.

AI deserves to be understood in the same way. The important question is not only what AI will replace. It is what becomes scarce once certain kinds of intelligence become abundant.

Productivity does not end scarcity. It moves it.

Economists have studied this pattern for decades. David Autor’s essay, “Why Are There Still So Many Jobs?”, remains useful because it challenges the recurring assumption that automation simply subtracts labor. His point is that technology often substitutes for some tasks while increasing demand for complementary human capabilities.

Daron Acemoglu and Pascual Restrepo describe a related idea in their work on automation and new tasks. Automation can displace labor when machines take over existing tasks, but labor demand can recover or expand when new human tasks are created around the new production system. That is the key question for AI: not whether old tasks are automated, but whether new work is created around the constraints AI exposes.

AI is a productivity shock to cognitive work. It makes drafting, summarizing, coding, searching, translating, classifying, modeling, and first-pass analysis cheaper. But enterprises do not run on cognition alone. They run on trust, governance, integration, accountability, adoption, capital allocation, and human judgment.

That is where the constraint is moving.

Inside large organizations, the bottleneck is already less about whether AI can produce something. It is whether the organization can safely absorb, govern, deploy, and act on what AI produces. Teams can generate more analysis than leaders can use. More code than technology organizations can release. More content than brands can trust. More recommendations than operating teams can execute. More automation ideas than governance models can responsibly support.

This is not a small distinction. It changes how executives should think about productivity.

If AI makes work faster but the enterprise cannot make decisions faster, the constraint is decision architecture.

If AI writes code faster but the enterprise cannot deploy faster, the constraint is engineering operations, cybersecurity, testing, and reliability.

If AI improves customer service but customers do not trust the answer, the constraint is escalation, policy, empathy, and accountability.

If AI increases clinical documentation speed but access remains constrained, the value shifts to care delivery, care coordination, and patient navigation.

The pattern is consistent: AI removes friction in one part of the system and exposes friction in another.

The new jobs may not be called AI jobs

A common mistake is to assume the AI economy will mostly create roles with AI in the title. Some of those roles will matter. But the more important job creation may happen in the surrounding economy, where new constraints become commercially important.

Here are the kinds of roles and services that are likely to grow.

AI deployment architect
Not someone who experiments with tools, but someone who redesigns workflows, data flows, controls, and adoption models so AI can operate inside a real institution.

Model risk and trust officer
As decisions become more automated, organizations will need people who can assess reliability, bias, auditability, explainability, and accountability across regulated processes.

Human escalation designer
In call centers, banking, insurance, healthcare, and government services, the routine interaction may be automated. The harder work will be designing when, how, and why a human intervenes.

Clinical AI workflow lead
Healthcare will need people who understand both care delivery and AI-enabled operations: documentation, triage, patient messaging, revenue cycle, quality, and safety.

AI-enabled customer recovery specialist
As digital service becomes more automated, the moments that reach a human will be higher-stakes: angry customers, complex exceptions, trust failures, and policy ambiguity.

Enterprise code release and reliability lead
If code generation accelerates, deployment becomes the constraint. Testing, security, observability, change control, and production reliability become more valuable, not less.

Data provenance and governance specialist
As organizations generate and consume more machine-produced content and analysis, they will need sharper controls around data lineage, consent, quality, ownership, and permissible use.

Memory, infrastructure, and inference capacity planner
AI will create new physical and technical bottlenecks: compute, memory bandwidth, storage, power, latency, and cost of inference. The “memory wall” is not only a hardware issue; it becomes a capital allocation issue.

AI adoption and workforce redesign leader
Productivity does not arrive because a tool is purchased. Jobs, incentives, roles, training, and management systems need to be redesigned around the new capability.

Domain expert evaluator
In law, medicine, finance, education, engineering, and public administration, AI output still needs review by people with judgment, context, and professional accountability.

These are not speculative fantasies. They are logical extensions of the bottlenecks already appearing in enterprises. Erik Brynjolfsson’s research on generative AI at work, for example, shows meaningful productivity gains in customer support, especially for less experienced workers. But the study also points to a deeper managerial question: if AI helps people produce faster, what new systems are needed to supervise quality, judgment, learning, and escalation?

That is where the work shifts.

The operator’s reality

From the outside, AI can look like a clean productivity story. From inside a large institution, it is more complicated.

A model can summarize a patient record, but someone still has to decide whether the summary is clinically safe to use.

A system can draft a customer response, but someone must own the policy, tone, risk, and exception path.

A developer can generate code faster, but the enterprise still has to test it, secure it, integrate it, monitor it, and support it in production.

A finance team can generate analysis quickly, but a leadership team still has to decide which trade-offs it is willing to make.

This is why the CIO role is changing. The job is no longer simply to provide systems, manage vendors, or operate infrastructure. The role is increasingly to understand how technology changes institutional capacity, where constraints move, and how the enterprise should redesign itself around those constraints.

For CEOs and boards, this is the more useful lens. AI should not be treated only as an efficiency program. It should be treated as a constraint-discovery mechanism. Wherever AI creates speed, leaders should ask what now becomes slow. Wherever AI creates abundance, leaders should ask what now becomes scarce.

The same applies to investors. In every technology cycle, the obvious winners are not always the only durable winners. When automobiles scaled, value accrued not only to car manufacturers but to fuel, roads, repair, finance, insurance, logistics, and real estate. In AI, value may accrue not only to model builders but to companies that solve deployment, trust, data quality, compliance, infrastructure, workflow integration, and sector-specific adoption.

The alpha may sit less in the visible technology and more in the hidden constraint it creates.

Preparing for the other side of AI

The practical takeaway is not that everyone should become technical. Nor is it that everyone should become an AI specialist. That advice is too narrow.

The better career question is: what human constraint can you help solve when AI makes more output possible?

For early-career professionals, the risk is relying on work that AI can produce cheaply: first drafts, basic analysis, routine research, simple code, generic content. The opportunity is to move closer to context, customers, operations, judgment, and implementation.

For mid-career professionals, the challenge is more serious. The people who use AI to redesign their own work will begin to separate from those who merely use AI to do the same work faster. The pivot is from task execution to workflow ownership.

For senior executives, the mandate is different. AI will test whether leaders can reimagine the functions they oversee. Productivity gains alone will not define relevance. The ability to redesign the operating model around those gains will.

AI will make many things easier to produce. That does not mean the economy will need fewer humans. It means the economy will need humans in different places.

The work ahead is to find those places before they become obvious.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

I hear versions of the question constantly. Parents ask what their children should study. Graduates wonder whether the jobs they prepared for will still exist in recognizable form. Mid-career professionals quietly ask whether the experience they spent twenty years building is becoming less valuable. Even senior executives are trying to determine what parts of their organizations — and what parts of their own leadership models — are durable.

The honest answer is that no one fully knows yet. But I also think the loudest narratives are missing something important.

The labor market is not giving us a clean story. The New York Fed reports that recent college graduates faced 5.7% unemployment and 41.5% underemployment in the first quarter of 2026, which means many graduates are working, but not necessarily in roles that fully use their degrees. At the same time, NACE reports that more than one-third of entry-level jobs now require AI skills, nearly triple the share from fall 2025, and nearly 60% of employers are assigning interns projects that use AI tools. The signal is not collapse. The signal is a rising bar.

The future is probably neither the collapse scenario that dominates social media nor the frictionless productivity utopia often presented on conference stages. What we are seeing instead is something more subtle and, in some ways, more consequential: the gradual movement of the first rung of the professional ladder.

For decades, entry-level knowledge work followed a relatively stable pattern. Young professionals entered organizations by doing work that was repetitive, process-heavy, and often not especially glamorous. Junior analysts built spreadsheets. Associates prepared presentations. New engineers handled smaller tickets and routine debugging. Coordinators managed workflows and documentation. Call center agents learned customer behavior one interaction at a time.

None of this work existed purely because it was economically efficient. It existed because institutions needed a way to develop judgment. Organizations were not simply producing output; they were producing future operators, managers, and leaders.

That is the layer AI is beginning to reshape.

The first draft of a document, the synthesis of research, the summary of a meeting, routine coding tasks, document comparison, administrative coordination — these are precisely the kinds of activities that AI systems can increasingly perform or accelerate. In many industries, the entry-level employee is no longer competing only against another graduate. They are competing against software that can compress hours of routine work into minutes.

This does not mean young professionals are becoming obsolete. But it does mean the economic logic of apprenticeship is changing.

Inside large enterprises, this dynamic already feels visible. In healthcare, AI can assist with documentation, coding workflows, patient communication, and administrative processes. In consumer businesses, it can generate marketing variations, summarize customer sentiment, and support service operations. In technology organizations, it can write meaningful amounts of code and documentation that previously would have been assigned to junior developers.

And yet, despite all of that, the hardest problems inside institutions still look remarkably human. Someone still has to make decisions when the data is incomplete. Someone still has to navigate regulation, trust, accountability, organizational politics, and operational trade-offs. Someone still has to exercise judgment when the answer is not obvious.

That distinction matters because it changes what will likely become valuable over the next decade.

I often tell students not to think too narrowly about “safe majors” or “AI-proof careers.” I do not think the future divides cleanly between technical and nontechnical disciplines. Some highly technical work is becoming more automated. At the same time, many traditionally human-centered skills are becoming more valuable precisely because AI makes routine production easier.

The people who will likely do well are those who combine fluency with judgment. They understand technology well enough to use it effectively, but they also understand context, communication, decision-making, and institutions. They know how to ask good questions, evaluate outputs critically, and apply knowledge responsibly.

That advice applies just as much to mid-career professionals as it does to graduates.

The risk for experienced workers is usually not that AI suddenly replaces decades of expertise. The greater risk is that someone else learns to combine comparable expertise with faster tools and new operating models. Experience still matters enormously, but experience by itself is becoming less sufficient. The professionals who remain relevant will be the ones who learn how to extend their judgment through technology rather than compete against technology on speed alone.

I also think executives need to think carefully about the second-order effects of all this. There is a temptation to view AI primarily as an efficiency tool, especially in large organizations under constant cost pressure. Some of those efficiencies will be real and necessary. But organizations should be careful not to optimize away the very experiences that develop future leadership capacity.

If younger employees no longer spend years learning how businesses actually function from the ground up, how will institutions produce experienced operators ten years from now? Every organization depends on some form of apprenticeship, whether it formally acknowledges it or not. If AI compresses the work traditionally assigned to junior professionals, companies will need to become far more intentional about how they develop talent.

That is why I remain more optimistic than pessimistic, even while acknowledging the disruption ahead.

Economic transitions are rarely smooth, and this one will almost certainly create real dislocation for some people and some professions. But history also suggests that societies eventually adapt around technologies that initially seem destabilizing. The nature of work changes. Expectations change. The definition of valuable expertise changes.

What matters now is not trying to predict every job that will disappear or emerge. It is developing capabilities that compound across technological shifts: judgment, adaptability, communication, domain depth, and the ability to learn continuously.

So to graduates entering the workforce, I would simply say this: your career may not unfold in the same way it did for the generation before you, but that does not mean there is no place for you in the future economy. There is.

The ladder is still there. The first rung is just moving.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

The hottest job in technology right now is not the model researcher or the product architect. It is the Forward-Deployed Engineer (FDE).

That should give enterprise leaders pause. When the market rewards a role whose primary function is to make technology usable, it is usually a sign that the technology itself is not yet ready to stand on its own.

The title is new. The underlying need is not.

Structural Insight

Every major technology cycle creates this role.

Mainframes needed customer engineers. Client-server systems needed integrators. ERP needed implementation consultants. Cloud needed solution architects and site-reliability engineers. Machine learning needed MLOps.

Now generative AI has created the Forward-Deployed Engineer.

The pattern is consistent. When technology is powerful but incomplete, the market inserts a human layer between product and reality.

A simple analogy from racing: the car may be engineered in the factory, but performance is determined on the race day. The race engineer sits at that intersection—reading conditions, interpreting signals, and making real-time adjustments. The forward-deployed engineer plays a similar role, translating environment into outcome.

Operator’s Reality

I have done a version of this job before the title existed.

Early in my career, I worked on building an EMR product modeled on an Oracle ERP stack. Later, we took that system into hospitals and deployed it. That meant sitting with clinicians, administrators, billing teams, and IT — watching how care was actually delivered, where workflows broke, and why technically correct systems failed in practice. Then we rebuilt the next version based on that reality.

It was one of the most formative roles I have had. Not because of the technology, but because it forced proximity to the business.

That is the essence of this role.

Generative AI is now running into the same constraint. The issue is not model capability. It is the gap between capability and operational value.

McKinsey’s recent work on AI makes this clear: most organizations have moved beyond experimentation, but only a small fraction are achieving meaningful, scaled economic impact. The constraint is not access to models. It is integration into workflows, governance, data readiness, and adoption.

This is where the Forward-Deployed Engineer sits — not as a novelty, but as a bridge.

The Enterprise Framework

For enterprise leaders, the question is not whether to hire for this title. The question is whether the capability exists.

The work required is specific:

First, connecting AI to real systems — not demos, but identity, data, APIs, controls, and audit.

Second, redesigning workflows — deciding where AI actually changes how work gets done, not just how it is assisted.

Third, defining acceptance — what “good” looks like, where human oversight sits, and how failure is handled.

Fourth, driving adoption — ensuring that the system is used, trusted, and measured.

These are not model problems but institutional change management.

The organizations that will extract value from AI will not be those with the best models. They will be those that can repeatedly convert capability into operating change.

That requires a different approach to talent.

The starting point is not hiring externally at scale. It is reconfiguring existing talent:

• Engineers who understand systems need to learn AI behavior and evaluation.

• Architects and SREs need to extend into AI governance, observability, and control.

• Business technologists need to become fluent in how AI changes decision-making and workflow design.

Then these capabilities need to be embedded — not centralized in labs, but deployed into functions: revenue cycle, contact centers, supply chain, finance, clinical operations.

This is where the role becomes real. Not as a job title, but as a working model.

Second-Order Implications

The rise of this role exposes something more structural.

For technology companies, it reveals where the product ends and the customer begins. A heavy reliance on forward-deployed engineers can signal closeness to the customer — or it can signal that the product is not yet complete.

The distinction is whether deployment work becomes reusable. If every implementation is bespoke, the economics resemble services. If field learning converts into product capabilities, the model compounds.

For enterprises, the implication is more direct - you cannot outsource this layer.

Vendors can accelerate initial deployments, but they do not own your workflows, your data, your controls, or your adoption curve. The institution does.

This is why many AI initiatives stall. The missing piece is not technology. It is ownership of the middle layer — where systems meet work.

For individuals, the signal is equally clear.

The market does not need more people who can describe AI. It needs people who can make it work inside constraints.

That means understanding systems, data, and operations. It means being comfortable with ambiguity and accountability. It means learning how decisions are actually made inside organizations.

The Bottom Line

The Forward-Deployed Engineer is being framed as the hottest job in AI.

That framing misses the point.

There is nothing fundamentally new about the role. It is the latest version of a function that has always existed: translating technology into institutional performance.

The hype is that this is a new category of work. The reality is that it is the same work, under more complex conditions.

The real takeaway is not about the role itself; it is about what it reveals.

AI is not constrained by intelligence. It is constrained by the enterprise’s ability to absorb it.

Organizations that recognize this will invest in the capability to bridge that gap — through talent, operating model, and governance.

Those that do not will continue to run pilots.

The title will fade but the need will not.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

Cybersecurity in the age of AI is usually framed as a threat story.

The focus is understandable: more persuasive phishing, faster fraud, better reconnaissance, and more scalable attacks. AI is clearly changing the economics of cyber offense. But that is only half of what matters.

The larger issue for enterprise leaders is that AI creates two cyber problems at once. The first is how to defend the organization against AI-driven threats. The second is how to build the cyber posture required to deploy AI safely across the enterprise.

That is the dual agenda.

AI is not arriving as a single system. It is showing up through copilots, workflow agents, embedded models, automation layers, and third-party tools woven into core processes. In healthcare, that may mean clinical documentation, patient engagement, coding, or device-adjacent workflows. In finance, it may be fraud operations, surveillance, underwriting support, or service automation. In each case, what begins as a productivity tool quickly becomes something more consequential. It touches data, decisions, controls, accountability, and resilience.

This is why the cyber question has become more central as AI adoption grows.1

The Real Shift

It helps to name the two frontiers clearly.

The first is AI for cyber. Here, AI strengthens defense but also improves the efficiency of offense. Advanced models are compressing the time required to do expensive cyber work: identifying weaknesses, refining exploit paths, turning threat intelligence into detections, and helping defenders patch more quickly. The most important shift is not science fiction. It is speed.

The second is cyber for AI. This is the discipline required to deploy AI without quietly creating a new attack surface. Once an AI system can retrieve sensitive data, browse, call tools, act across workflows, or influence decisions, the control problem changes. Prompt injection becomes more than a model quirk. Excessive permissions become business risk. Weak logging becomes an assurance gap. Third-party model dependence becomes an operational dependency.

In simple terms, AI is changing both the threat environment outside the enterprise and the control environment inside it.

That is why this is no longer only a CISO issue. It is also a CIO, CFO, board, and CEO issue.

Operator’s Reality

For operators, the implications are practical and immediate.

Threat and vulnerability management changes first. Many programs were built around a familiar rhythm: find vulnerabilities, rank them by severity, and work through remediation queues. That logic becomes less reliable when AI accelerates vulnerability discovery and exploit-path analysis. A medium-severity issue that can be chained with others may matter more than a critical issue considered in isolation. Time to mitigation starts to matter more than volume of findings.

GRC changes next. The instinctive response to AI is often policy. Policy matters, but it is not enough. The more important shift is from AI policy to AI assurance2. Leaders need to know which AI systems exist, what data they touch, what permissions they hold, which models they rely on, and who is accountable when they fail.

This is where many executive teams have awareness, but not always enough connection to operational reality. They understand that AI is moving quickly and that cyber risk is rising. What is often less developed is the link between those risks and the institution’s day-to-day design choices: identity architecture, engineering discipline, vendor concentration, workflow controls, and resilience planning.

A hospital does not experience AI risk as an abstract model issue. It experiences it through patient data, clinician workflows, operational dependencies, and trust. A bank experiences it through fraud controls, third-party risk, and regulatory accountability. A call center experiences it through service actions, verification, and reputational exposure. The context changes. The management challenge does not.

The Enterprise Framework

The right response is not to create a small AI security function on the side. It is to adapt cyber posture as part of enterprise architecture.

That starts with a few basic moves.

First, treat AI systems as enterprise assets. Inventory them. Classify them by data sensitivity and action privilege. Know which systems can access regulated data, influence decisions, or trigger real-world actions.

Second, redesign controls around agency, not just access. It is no longer enough to ask who can see the data. The more important question is what the AI-enabled system can do with it, what tools it can call, and where meaningful human review is still required.

Third, move cyber investment closer to AI scale-up decisions. Identity, logging, observability, model lineage, third-party diligence, and incident readiness are not secondary costs. They are part of the cost of deploying AI responsibly.

Fourth, integrate cyber into delivery. Prompts, connectors, permissions, and model versions should be managed with the same discipline applied to other production assets.

To borrow from Formula 1: speed alone does not win the race. Telemetry, control, and disciplined adjustments do. AI increases the speed of the machine. Cyber helps keep it on the track.

Second-Order Implications

The second-order effects are becoming just as important as the direct ones.

One is regulatory direction. There may never be a single clean AI-cyber rulebook. Instead, leaders are facing a layered stack of obligations: AI governance, cyber requirements, operational resilience, sector regulation, disclosure expectations, and product liability. The legal architecture may still be evolving, but the management implication is already clear. Institutions will be expected to show evidence of oversight, controls, and accountability.3

Another is market structure. The recent focus on Anthropic’s Mythos and similar model developments matters not simply because one model is more capable than another, but because it signals a reset in how cyber capability is evolving.4 At the same time, state actors do not need fully autonomous tools to benefit. Faster reconnaissance, cheaper experimentation, and shorter learning curves are enough to narrow the defender’s window to respond.

There is also a quieter investor implication here. As operators work through this transition, value will not sit only in traditional security products. It will also be created in the operating layers that help institutions absorb AI safely at scale: assurance, identity, observability, third-party control, resilience testing, compliance operations, and workflow redesign. This is not separate from AI adoption. It is increasingly part of the infrastructure that makes AI adoption viable.

What Matters Now

Three things matter most.

First, organizations need to treat AI cyber risk as a dual agenda: using AI to strengthen defense, while also making AI deployment secure, governed, and resilient.

Second, the core weakness will not usually be lack of awareness. It will be the gap between awareness and operational readiness, especially across identity, data controls, third-party dependencies, and remediation speed.

Third, this is becoming a leadership and operating-model issue, not just a security issue. The institutions that do well will be the ones that act early, not after regulation, market pressure, or when a threat activity forces the change.

That is where the real work now sits.

I am grateful to the cybersecurity professionals I work with closely, whose insight and practical judgment continue to deepen my understanding of a rapidly changing field.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

For the last twenty years, businesses have been training people where to knock.

First it was the website. Then it was the app. Each shift was described as a new channel. In practice, each became a new front door. Each time the front door changed, the business changed with it.

It took companies two decades to learn that the website was not a brochure but part of the operating model. It took less than ten years to learn the same lesson about the app. At first, businesses pushed customers there to lower cost and widen reach. Eventually, customers demanded it because convenience had become part of the product.

AI is the next turn of that wheel. The website gave the business digital presence. The app gave it digital proximity. AI will give it something more consequential: a front door that can interpret intent, hold context, and increasingly act.

Popular culture saw this before business did. JARVIS in Iron Man. The computer in Star Trek. Same instinct. The interface is no longer just a screen.

That is why AI should not be understood as another feature inside software. It is becoming the new interaction layer for software.

The Operator’s Reality

The website was mostly about reach. The app was mostly about convenience. AI will be about interpretation.

Most friction in business does not come from lack of access. It comes from lack of understanding. A patient is not really asking for a portal. The patient is asking: should I be worried, what does this result mean, and what do I do next? A retail customer is not asking for better navigation. They are asking which option is right for them. A banking or insurance customer is not asking for another menu. They are asking what happened, what it means, and what comes next.

Legacy interfaces were built to publish information and route transactions. They were far weaker when the interaction involved ambiguity, anxiety, or judgment. AI changes that because it can meet people in natural language and remain engaged across steps.

In healthcare, the next patient interface may not be your website or even your portal. It may be the patient’s most trusted AI. In retail, AI becomes the personal shopper. In banking and wealth management, it becomes the first interpreter of tradeoff and choice. In hospitality, it becomes concierge and recovery layer.

The interface no longer just presents the business. It starts to represent it.

The Strategic Test

The temptation for leadership teams will be to treat this as a channel decision.

The real question is where the business will actually be encountered when conversation becomes the primary mode of interaction. In consumer markets, that means the first moment of need, not the moment of log-in. In B2B, it means the buying journey may begin inside a procurement or finance copilot long before a salesperson enters the room.

Once that happens, the value of the interface rises because whoever owns the interaction owns the context. The burden of trust rises because advice without accountability will not hold in healthcare, banking, or insurance. And the standard for enterprise readiness rises because it is not enough to be accessible through AI. The business has to be designed for it.

That means exposing the institution safely enough that AI can do more than answer questions. It has to interpret intent, execute workflows, preserve controls, and hand off cleanly when judgment belongs with a human.

The Investor’s Read

For years, the digital maturity question was straightforward: does the business have modern channels and digitized workflows?

That was useful. It is no longer enough.

The new question is whether the business can support AI at the front without breaking trust, margin, or operating coherence.

That is a much harder standard.

A software company with AI features is not automatically advantaged. A labor-heavy services business is not automatically threatened. In many cases, the better asset will be the company whose workflows are structured enough, whose data is usable enough, and whose service model is strong enough that better interaction improves both efficiency and retention.

This is especially relevant in software and tech-enabled services. The next value creation story may not come from replacing labor outright. It may come from redesigning the interaction model so the business handles more complexity at lower cost, with better continuity and better responsiveness.

Investors should watch for a simple distinction: is AI sitting on top of the business, or is it being woven into the way the business is actually delivered?

Only one of those compounds.

The Employee Interface

The external side of this shift gets more attention because it is easier to visualize. The internal side may matter just as much.

For employees, AI will become the new membrane between people and institutional complexity.

For the last twenty years, organizations kept asking employees to learn the system. The next ten years will be about systems learning how work actually gets done.

Employees will increasingly begin with AI to retrieve context, summarize prior decisions, draft materials, compare options, identify exceptions, and prepare actions across systems. In that sense, AI becomes the employee interface, not just a productivity layer. (Microsoft 2025 Work Trend Index)

Instead of opening five applications, reading a long email thread, and setting up two meetings just to get oriented, the employee starts with an intelligent layer that compresses the enterprise into something usable.

That is the real internal shift and will define modern workforce technology.

The Closing View

Every major technology shift changes the surface of the business. The important ones change the structure beneath it.

The website gave businesses presence. The app gave them proximity. AI will give them a front door that listens, interprets, guides, and increasingly acts.

That is why this is not a tooling story. It is a control story, a trust story, and an operating model story.

For customers, AI will increasingly become the first place they ask, compare, decide, and return. For employees, it will increasingly become the first layer through which they access knowledge, process, and coordination.

The Formula 1 analogy is simple. Spectators see the driver. The race is won by the telemetry, the pit wall, and the discipline of the system behind the car. AI may look like a front-end shift. The winners will be determined by the operating model behind it.

So the strategic question is no longer whether AI will matter.

It is whether your business is being designed for a world in which the front door “thinks”.

The companies that win will not be the ones that bolt AI onto the enterprise. They will be the ones that redesign the enterprise so AI can serve as a thinking front door, for the customer, for the employee, and eventually for much of the business itself.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

There is a quiet contradiction playing out between buyers and sellers—and their investors—on the future of the services business.

Enterprises are asking their service providers to use AI to lower cost while maintaining—or improving—service levels. At the same time, most services businesses are still built on pricing models tied to labor, utilization, and scale.

The more effective AI becomes, the more it erodes the economic foundation of the services model itself.

This is not a negotiation tension. It is a structural one. And it is accelerating.

Structural Shift: From Labor to Systems

From the outside, this moment is often framed as “AI adoption in services.” That framing is too narrow.

What is actually underway is a shift from services delivered by people to services executed by systems.

Or more plainly: services as software.

Agentic systems are now capable of performing parts of research, analysis, workflow orchestration, and decision support that were previously human-intensive. In healthcare revenue cycle, tasks like coding validation, denial prediction, and documentation review are increasingly system-led. In call centers, agent assist and autonomous resolution are already reducing handle time and manual effort.

Recent industry research reflects this shift. A large majority of enterprises already report using AI within outsourced services, while also acknowledging that governance and contracting models are lagging behind the technology.

The implication is straightforward.

The historical moat in services—labor arbitrage, scale, and incremental technology enablement—is weakening.

A new moat is emerging: ownership of the workflow, the data, and the system that executes the work.

In that model, human expertise does not disappear. It moves outward—wrapping, supervising, and improving systems rather than executing tasks directly.

This is the inversion.

The Operator’s Reality

From where I sit, this shift is already visible.

As a buyer of managed services and business process management (BPM) at scale, the expectation is no longer subtle. Productivity gains driven by AI are expected to show up in the contract. Not as a future benefit, but as a current adjustment.

At the same time, most RFPs are still structured around FTEs, SLAs, and rate cards—and sales and go-to-market teams are finding it increasingly difficult to clearly articulate this pivot to customers while still protecting existing revenue models.

That misalignment creates friction.

As an advisor to investors and services businesses, I see a parallel issue. Many of these assets are still underwritten on assumptions that have held for decades: steady headcount growth, utilization discipline, offshore leverage, and annuity-like revenue streams.

Those assumptions are now under pressure.

In practical terms, we are beginning to see two very different pricing curves emerge.

Incumbent providers, layering AI onto existing delivery models, are typically able to drive incremental efficiency—often in the range of 5–10% on a given cost base, though this varies by workflow and maturity.

At the same time, AI-native entrants—re-architecting workflows from first principles—are positioning around far more aggressive reductions, in some cases targeting 50% or more. These numbers are often aspirational, but they are shaping buyer expectations nonetheless.

The gap between those two curves is where the market is moving.

And that gap is what forces the pivot.

The Services Reset

For investors and operators, this is no longer a question of whether to adopt AI. It is a question of whether the business can survive its own productivity gains.

In practice, I see four decisions defining that reset.

The first is cannibalization. Are you willing to reduce revenue per unit of work in order to remain relevant? Or do you protect the current model and risk disintermediation over time? There is no stable middle ground.

The second is control of the workflow layer. In the next model, value accrues to whoever owns the system that executes and orchestrates the work. If that layer sits outside the services provider, the provider becomes interchangeable.

The third is capital deployment. Do you build these capabilities internally, acquire them, or partner your way into them? Each path has different implications for speed, control, and long-term value capture.

The fourth is operating model and talent. The traditional pyramid—large execution layers supporting a thin management band—does not translate cleanly into this world. The emerging model looks more like a barbell: fewer operators, more product managers, data engineers, and workflow architects.

None of these decisions are incremental. All of them are institutional.

Two Strategic Paths

Across the market, two strategic paths are beginning to take shape.

The first is the incumbent path: services firms acquiring AI-native capabilities to modernize delivery.

This approach leverages existing client relationships and revenue bases. It allows for a staged transition. But it carries real risk. Integration is hard. Cultural alignment is harder. And in many cases, AI remains a layer on top of the existing model rather than a re-foundation of it.

The second is the inverse: AI-native firms acquiring services capabilities.

These firms start with a different assumption. They design around systems first and add human expertise where necessary. They are not constrained by legacy pricing models or revenue expectations.

Their limitation is different. They often lack deep domain knowledge, enterprise trust, and the ability to operate at scale within regulated environments—at least initially.

In the near term, both paths will coexist.

Over time, I believe the second model has structural advantages. It does not have to unlearn the past—and continued advances in AI will only accelerate that advantage, giving these firms a sustained tailwind against incumbents.

But in the current market, incumbents still have something that matters: distribution, trust, and cash flow. That buys them time—if they use it.

The F1 Analogy: Rebuilding the Car

This moment reminds me of Formula 1 more than traditional technology cycles.

The winning teams are not the ones that incrementally improve last year’s car. They are the ones willing to redesign the chassis when the regulations change—even if it means giving up short-term performance.

Most services firms are still tuning the engine.

The leaders will rebuild the car.

Second-Order Implications And Quest for Alpha

For investors, this moment requires a different underwriting lens.

Revenue durability can no longer be assumed from contract structure alone. Margins tied to labor efficiency are more exposed than they appear. Terminal value assumptions based on steady-state growth deserve scrutiny.

The gap between leaders and laggards will widen meaningfully.

Assets that successfully transition toward workflow ownership and system-led delivery will command a premium. Those that remain tied to labor scaling will face pressure—both on growth and on valuation multiples.

M&A activity will likely accelerate, not as consolidation, but as capability acquisition.

For services CEOs, this is not a technology transformation. It is a leadership one.

Your next COO will not look like your last. Increasingly, it will be someone who understands systems, data, and workflow orchestration as deeply as operations—because execution itself is becoming a technology problem.

It requires changing what you sell, how you price, how you deploy capital, and who you hire. It requires being willing to disrupt your own revenue model before the market does it for you.

And for enterprise buyers, the implication is equally clear.

Contracts will be repriced. The future model will blend base capacity with outcome-based pricing, with explicit mechanisms to share AI-driven productivity gains. Governance will matter more, not less.

The Demand Paradox

There is a temptation to see this as a contraction story for services.

History suggests otherwise.

When robotics entered manufacturing, it did not eliminate demand—it increased throughput, reduced cost, and expanded what the market could absorb—reshaping the nature of work.

The same pattern is likely here. In auditing, continuous monitoring and AI-assisted review will expand coverage. In healthcare, more claims will be worked, not fewer. In customer operations, more interactions will be handled, not less.

As the cost of executing complex workflows declines, enterprises will consume more of them. Work that was previously too expensive, too manual, or too slow will become viable.

The sequence is already familiar:

Human → Human with technology → Technology with human oversight → Technology-led execution

Closing

The services industry is not being displaced. It is being rebuilt.

From labor to systems. From scale to orchestration. From contracts priced on effort to outcomes priced on execution.

The moat is moving.

And the firms that endure will not be the ones that adopt AI fastest.

They will be the ones willing to redesign their business around it—even when that means letting go of the model that created their success.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

A recent commentary from Citadel Securities pushed back on the prevailing AI panic narrative. Adoption, they argue, is not vertical. Software job postings are not collapsing. New business formation remains resilient. Recursive technology does not automatically produce recursive economic deployment.¹

The models are advancing rapidly. But scale in the real economy is still constrained — by compute, energy, regulation, and capital.

Most commentary focuses on the first three.

Inside enterprises, the fourth may matter more.

AI is not failing in the lab. It stalls in the capital committee.

The 10–15% Question

Several months ago, at a CEO/CFO forum, I was asked a simple question:

How much capital should we allocate to AI?

I answered instinctively: 10–15%.

I meant transformative capital — not total IT spend. In most established industries, companies distinguish between maintenance capital required to sustain operations and growth capital intended to expand or reposition the enterprise, with maintenance often absorbing the majority of capital budgets and leaving a smaller share for true transformation.

But even as I said it, I knew the answer was incomplete.

A percentage assumes AI is a category. A budget bucket.

It isn’t.

AI is a capital thesis.

How Capital Governance Shapes Ambition

Across capital-intensive service industries — healthcare systems, airlines, energy, telecommunications, large financial institutions — capital governance evolved to protect stability.

Committees prioritize solvency, rating strength, regulatory posture, and downside risk.

Capital is allocated to physical expansion, equipment renewal, and infrastructure modernization. These assets depreciate predictably. Their replacement cycles are understood.

Technology investments, historically, have often been framed as necessary — compliance, modernization, operational continuity. Rarely as investments expected to generate measurable return.

That framing persists.

But AI does not behave like a traditional system upgrade.

The Opex Trap

Most large enterprises do not build foundational AI models. They buy capabilities — through SaaS platforms, usage-based APIs, and embedded copilots.

These purchases show up as operating expense.

Operating expense increases run-rate cost. It pressures margins immediately. It is scrutinized quarterly.

Capital investment, by contrast, is evaluated over a multi-year horizon. Returns are expected over time. Governance tolerates delayed yield.

When AI is funded primarily through operating budgets, it must demonstrate visible return quickly — or risk being cut.

When AI is funded through capital allocation, leaders can justify a longer horizon — building capability first and expecting returns over time.

The funding structure shapes the expectation.

Vendors Understand the ROI Problem

AI platform companies understand this tension clearly.

OpenAI, Microsoft, Google, Anthropic and others are not only advancing models — they are building enterprise partnerships, integration layers, and advisory ecosystems designed to drive measurable ROI. Major consulting firms have scaled AI implementation and value-realization practices.

The market understands that model capability is insufficient without adoption, workflow redesign, and economic capture.

The constraint is not awareness of potential.

It is alignment inside the enterprise — between incentives, governance, and financial expectations.

ROI Depends on the Digital Foundation

In many organizations, AI ROI disappoints not because the technology underperforms, but because expectations are misaligned with reality.

Productivity gains diffuse across departments. Throughput improvements are absorbed rather than monetized. Labor savings do not automatically become structural cost reduction.

Return depends on the condition of the digital foundation — the data, systems, integration maturity, and operating discipline that AI relies on.

Traditional capital-intensive industries track the “age of plant” as a proxy for asset productivity. Only a few organizations apply similar rigor to the age of their digital asset base.

Enterprise systems layered over a decade ago. Fragmented integration architectures. Data platforms built incrementally rather than intentionally.

AI deployed atop an aging digital foundation will underperform, regardless of spend.

Capital allocation sets priority. Priority shapes the digital foundation. The digital foundation determines yield.

Tech modernization becomes a prerequisite to AI.

The CIO–CFO Inflection Point

Too often, technology conversations with finance remain transactional — requests for tools, comparisons to peers, appeals to industry momentum.

The more important conversation is economic.

What lever are we moving?

Margin expansion?
Labor productivity?
Revenue capture?
Risk compression?
Asset utilization?

Is this maintenance, renewal, yield enhancement, or strategic repositioning?

Operating expense and capital allocation are not merely accounting categories. They represent different time horizons, risk tolerances, and definitions of success.

If AI is treated as incremental spend, it will be evaluated like incremental spend.

If AI is treated as capital deployment, it can be governed like capital deployment.

The Institutional Test

The real constraint on AI adoption is not model intelligence.

It is alignment — between funding, expectations, and the condition of the digital foundation.

If AI is funded as incremental operating expense, it will be judged on incremental timelines. Leaders will expect visible impact within quarters. That may be appropriate for narrow automation. It is unrealistic for structural transformation.

If AI is funded through capital allocation, the organization signals something different. It signals that capability will be built before yield is harvested. It accepts staged return and multi-year accountability.

Neither approach is inherently right or wrong.

What matters is whether the funding structure matches the ambition.

How we fund AI defines what we expect from it — and when.

But there is another constraint beneath this one.

AI depends on the quality of the digital foundation.

In Formula 1, a championship-caliber power unit cannot compensate for an unstable chassis. The engine may be world-class. The lap time will not be.

AI behaves the same way.

If expectations are immediate but the digital foundation is fragile, disappointment is predictable.

This is why some organizations conclude that AI “doesn’t deliver ROI.”

It often does.

But return is downstream of capital discipline and digital maturity.

The question is not simply how much to spend.

It is whether funding structure, time horizon, and digital readiness are aligned.

AI is not a budget line.

It is a capital decision — and capital decisions define ambition, accountability, and the timeline of return.

¹ Citadel Securities, 2026 Global Intelligence Crisis, 2026.
https://www.citadelsecurities.com/news-and-insights/2026-global-intelligence-crisis/

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

If you are not slightly overwhelmed by AI right now, you are not paying attention.

Boards are asking about it. CEOs are referencing it in earnings calls. Vendors are embedding it into every roadmap. Internal teams are experimenting faster than governance structures can absorb.

And yet inside most enterprises, AI is still being treated as a deployment question.

It isn’t.

It is an institutional readiness test.

The real question is not, “What is our AI strategy?”

It is: Is our institution structurally prepared to absorb intelligence at scale?

The Structural Shift

AI collapses the distance between idea and execution. It lowers the cost of experimentation. It increases the speed of software delivery by an order of magnitude.

That acceleration is destabilizing (like a F1 race car).

Most enterprises were designed for slower cycles — annual budgets, multi-year roadmaps, tightly gated releases.

AI breaks that cadence.

The constraint is no longer technical capability.

It is governance capacity.

Legal teams must evaluate contracts that include evolving models, training data exposure, and ambiguous liability boundaries. Enterprise risk must assess model behavior, not just cyber posture. Finance must determine how to measure productivity gains that don’t fit traditional ROI templates.

Meanwhile, functional leaders are energized — but unevenly prepared for change. Some are ready to redesign work. Others want to “install AI” without altering incentives or accountability.

From the CIO seat, the question is not how fast we can deploy.

It is whether we can govern what we deploy.

A Three-Layer View of Enterprise AI

In practice, AI readiness rests on three institutional layers.

1. The System-of-Record Layer

Your core platforms — ERP, EMR, financial systems, CRM — will increasingly embed AI natively. Summaries, copilots, anomaly detection, workflow automation.

This is the lowest-friction starting point.

The guidance here is simple: engage your major software vendors now. Understand what capabilities are already available within your licensed platforms. Turn features on deliberately. Pilot inside controlled domains. Strengthen data governance and access controls before scaling.

This layer is about compounding value from assets you already own.

It is also about discipline. Avoid bolting on external solutions to solve problems your core systems will address within 12–24 months.

Stability here determines leverage everywhere else.

2. The Targeted Value Layer

This is where disproportionate ROI lives.

Point solutions that remove a measurable constraint: automated prior authorization workflows, AI-assisted coding, revenue leakage detection, intelligent scheduling optimization, contract analysis acceleration.

Narrow scope. Clear metrics. Explicit accountability.

The guidance: fund high-impact use cases that produce visible results within months — but avoid long-term architectural lock-in. Many point vendors will be leapfrogged as core platforms catch up. Structure contracts accordingly. Preserve exit flexibility.

Treat these initiatives as capital allocation decisions, not experiments.

Every deployment should answer:
What constraint are we removing?
Who owns the outcome?
How will we measure it?

Without that clarity, AI becomes expense disguised as innovation.

3. The Agentic Reimagination Layer

This is the most misunderstood — and the most powerful.

Beyond augmentation lies redesign.

Entire sub-functions can be reimagined as AI-led and human-governed systems. Not incremental productivity gains, but structural workflow transformation. Think revenue cycle redesign. Supply chain planning. Clinical documentation ecosystems. Corporate FP&A.

Here, technology is not the bottleneck.

Change management is.

Decision rights shift. Escalation paths change. Supervision models evolve. Incentives must be rewritten.

The guidance: move deliberately. Engage business leaders early. Frame this as operating model transformation, not software deployment. Invest heavily in communication and retraining. Measure adoption as rigorously as technical performance.

Most organizations try to jump here first.

Few are ready.

The Executive Test

Boards are asking a simple question: “What are we doing about AI?”

That question requires three answers:

A 30-second answer that signals control.
A three-minute answer that demonstrates prioritization.
A thirty-minute answer that reveals governance, capital discipline, and operating model alignment.

If you cannot articulate all three, you are not ready — regardless of how many pilots are running.

AI is everyone’s agenda, not just the CIO’s.

Senior executives cannot outsource fluency. Use it yourself — not as a search engine, but as a collaborator. Draft with it. Analyze with it. Pressure-test decisions with it.

Personal fluency sharpens enterprise judgment.

What Happens Next

Over the next three to five years, we will see divergence.

Some organizations will accumulate fragmented tools, unclear accountability, and rising risk exposure. Costs will increase. Value will be difficult to prove.

Others will treat AI as governance and capital discipline. They will fund fewer initiatives, demand measurable outcomes, clarify decision rights, and integrate AI oversight into enterprise risk.

Those institutions will compound advantage.

For CIOs, this is a structural moment.

The enterprise needs someone who understands systems of record, governance mechanics, capital allocation, and operating model design — simultaneously.

That is no longer an operational role.

It is architectural.

AI will evolve quickly. This perspective, written in early 2026, will age.

What will not age is this:

Technology moves fast.
Institutions move slowly.

The leaders who win will not be those who chase every capability.

They will be those who redesign their institutions to absorb capability without losing control (like a good F1 driver).

AI is not testing your technology stack.

It is testing your institution.

I write here in a personal capacity. The perspectives shared are my own and do not represent my employer or affiliated organizations.

The CIO is no longer a back-office operator. In many organizations, the role is becoming a first-class executive seat — shaping growth strategy, influencing capital decisions, redesigning operating models, and increasingly, defining how AI will (or won’t) reshape the institution.

Technology is no longer support infrastructure. It is the lever.

Across industries, when boards and CEOs ask how to evolve, differentiate, or grow, the answer almost always runs through technology. Not as a project. Not as a system implementation. But as a structural advantage.

The previous generation of CIOs built the digital backbone. They digitized workflows, modernized infrastructure, professionalized IT, and earned the organization’s trust. They quite literally defined Information Technology inside the enterprise. That foundation is their legacy.

Our generation inherits something different.

We are not being asked to digitize. We are being asked to architect the enterprise itself — data models, governance structures, AI strategy, capital prioritization, platform choices, and the institutional courage to change.

That’s a very different mandate. And it raises a question I think about often:

What will be our legacy?

This publication is my attempt to think that through in public.

These are my working notes — lessons from mentors who shaped me, CIOs I’ve worked alongside, leaders I’ve admired, mistakes I’ve learned from, and perhaps most importantly, the expectations I hear from CEOs, boards, physicians, operators, and peer executives.

Some posts will be about strategy. Some about AI. Some about operating models and institutional design. Some about where CIOs get in their own way.

If the last generation built the digital foundation, this generation will define whether technology becomes institutional leverage.

This is my diary as I try to understand what that really means.