AI changes the second half of that sentence. It changes what suppliers can produce, which gets most of the attention, and it changes what customers need to buy, which gets less.

Why the dependency was accepted

A solution contract bundles three things that are usually discussed as one. There is the physical asset and the network that services it. There is the risk the supplier takes on, availability guaranteed, downtime paid for by the supplier. And there is the information layer, the monitoring, the optimisation, the recommendations and the reporting that tell the customer what to do and when.

Customers did not accept the bundle because they loved dependency. They accepted it because the information layer was expensive to build and needed specialists they did not employ. A plant manager could in principle have run his own bearing analytics. In practice the model, the data pipeline and the people to interpret it cost more than the premium SKF charged for doing it. The convenience was priced in, and nobody had a cheaper option to compare it with.

What AI makes cheap

Capabilities that once needed a large budget and a specialist vendor are becoming cheap and adaptable on the customer's side. That does not mean every customer will build its own systems. Most will not want the burden or the risk, and many should not. But the option now exists where it did not, and an option changes a negotiation even when it is never exercised.

Three uses are already visible. A customer can optimise across several suppliers instead of inside one supplier's platform, which removes the quiet advantage of being the system the customer looks at every morning. A customer can test the supplier's recommendations, replacing the service interval and the spare parts forecast with its own model and seeing where the two disagree. And a customer can build partial substitutes, enough of the information layer to know what the rest should cost.

The customer may well keep buying from you after all three. It buys with better information than before, and the margin that sat in the convenience layer goes with it. The relationship the solution model promised turns out to have rested partly on an asymmetry, and the asymmetry is closing.

The three parts do not move together

This is where the two standard examples stop behaving alike. In the engine business the lock-in was never mainly informational. The airline cannot certify an engine, cannot stand up a global maintenance network and does not want the availability risk back. AI gives it a better view of what it is paying for and a stronger hand on scope and price at renewal. Leaving gets cheaper on the information part of the contract and stays as expensive as before on the rest, so the contract gets renegotiated rather than replaced.

Bearing inventory management carries more of its value in the information and logistics layer. The customer owns the machines, the sensors can be anyone's, and the logistics are close to a commodity. What the supplier keeps is the knowledge of how bearings fail in that installed base and the reliability of the service itself. The analytics premium is the part AI reprices first, because the problem it solved was a lack of tools rather than a lack of possibility.

The split runs inside each company's offer, not between companies. The question for a leadership team is which parts of its own offer solved problems the customer truly could not solve alone, and which parts were convenient mainly because the customer lacked the tools. The second category is repriced at the speed model costs fall, the first at the speed of certification, service networks and liability, which is far slower. Most solution portfolios contain both, priced as one.

Where the data sits

One asset is left out of the above and should not be. AI gives the customer the model. It does not give the customer the data if the data is generated by the supplier's sensors, held on the supplier's platform and licensed back under the service agreement. In many installed-base businesses that is the real moat, and it was built almost by accident.

Expect it to stop being accidental. Data access, portability and the right to run your own models on your own operating data will move from the appendix of the service agreement into the first negotiation round. A supplier that has quietly relied on holding the data should assume that position is now visible to the customer, because the customer has the tools to notice what it is missing.

The same shift works in the supplier's favour on one axis. AI lowers the cost of the model for everyone. It raises the value of the data the model is run on, and a supplier with ten thousand installations has operating data no single customer can assemble from its own plant. Failure patterns, the effect of a maintenance regime across hundreds of sites, what a reading means for this machine given what it meant for the other nine thousand. A customer can build the model. It cannot build that. So AI does not simply remove margin from the solution business. It moves margin out of the convenience layer, which the customer can now reproduce, and into the cross-installation layer, which it cannot. Suppliers that treat their data as something to hold will defend the first and lose it. Suppliers that offer the customer its own data freely and compete on what the whole installed base teaches will give up a small position to take a larger one.

What the argument misses

The argument treats the information layer as something the customer can now reproduce. That is true of the model and not yet of the judgment. Knowing what a reading means for one plant, with its history and its maintenance habits, is still mostly in the heads of the supplier's engineers, and a customer running its own model without that context will be wrong in ways it cannot see. How long that lasts differs by industry, and in the industries where the supplier's engineers are also the ones who turn up when something breaks it may last a long time.

It also assumes customers want the option. Many procurement organisations are built to buy outcomes, and the people in them have no incentive to take back a risk they successfully pushed out. The repricing described here will come from customers with operations people in the room, and it will come unevenly, so a supplier will see it first in a few accounts and can mistake that for a few difficult customers.

The operating model has to follow

The sorting above is a pricing exercise before it is anything else. A solution priced as one bundle hides which part the customer is paying for, and when the customer can price the information layer itself the bundle unravels from the inside. Pricing the asset, the risk and the information separately, at least internally, shows where the margin actually sits and how much of it is exposed.

The service organisation follows the pricing. A service unit sized for a world where the supplier interprets everything is too large for a world where the customer interprets half of it. The people who remain need to be the ones who handle what the customer still cannot, the exceptions, the physical intervention, the liability.

Sales changes last and hardest. Part of solution selling was the creation of dependency, and sales organisations were paid to deepen it, even where the contract also transferred real risk and improved the asset. Selling to a customer that can check your numbers is a different job. The claim has to be verifiable, the outcome has to be one the customer can measure with its own tools, and the premium has to be explainable in terms of the first category rather than the second.

The economics above describe the direction, not the timing. Which year a given industry's convenience layer is repriced depends on model costs, on who owns the sensors, and on how many customers have the people to use the option. A supplier that reprices itself before the customers can is giving away margin. One that waits for the customers to do it is giving away the terms.

I agree with the diagnosis. I doubt the remedy. It puts the board where the decision shows up, which is rarely where it was made.

Where the decision was made

The week a tariff lands, management chooses between the options the sourcing footprint left open, and the footprint was decided years before. A supplier failure plays out inside the room set by how many sources were qualified and where stock was held. A price shock on energy lands on contracts and hedges signed long before it.

A board that joins the weekly loop adds a step to a decision that needs fewer steps. It arrives with less information than the people already in the room, and with the authority to override them.

Decisions taken in calm

What a board can do is decide in calm what has to hold during a disruption. These decisions concern operations, but they are meant to change slowly.

Exposure in operating terms. The largest share of volume the company accepts on one supplier, one market or one customer. How much margin it will give up to keep a customer through a disruption, and for how long.

Decision rights and thresholds. What management can do without asking, what comes back to the board, and how fast. Switching supplier at a cost of two points of margin for a quarter is management's call. Changing a commitment to a customer worth a fifth of revenue is not. The threshold has to exist before the disruption, because during it every decision feels like the second kind.

The assumptions under the strategy. The few assumptions that would break the plan if they proved wrong, each with a trigger that brings it back to the board when it moves. Energy prices, a tariff regime and the price point of AI-enabled competitors are typical.

None of this is new. What has changed is the cost of skipping it. In a stable period nothing tests these decisions. In a volatile one they are tested every quarter, and a board that agreed nothing in advance has to be consulted on everything.

Operating data as evidence

The day-to-day matters to the board as evidence about those decisions. The question about a weekly number is whether it pushes an assumption past its trigger or shows that a threshold was set in the wrong place. If it does, the board has work. If not, the board has information.

That distinction gets harder as data gets cheaper. Lawyers at Sullivan & Cromwell noted last year that, absent a conflict or improper purpose, a director's right to company information is essentially unfettered. The line is crossed by purpose, asking in order to oversee or asking in order to run. AI removes the cost of asking. A director can now hold the same dashboard as the COO, without the context the COO has.

Accountability

A board that shares the operating decisions also loses the position from which it judges them. When the CEO executes a decision the board pushed for, the board cannot hold the CEO to account for the result, and the CEO knows it.

BDO's 2026 survey of more than 200 US public-company directors suggests boards are finding this difficult. The share rating their collaboration with the CEO and management as highly effective fell from 69% to 58% in a year.

Ram Charan, Dennis Carey and Michael Useem described three modes for a board, taking charge, partnering and staying out of the way. Each is right at some point in a disruption. The failure is moving between them without saying so, which leaves management guessing which board it has this week.

What this does to the market for directors

Setting a threshold in operating terms, and reading an operating number as evidence about an assumption, requires directors who know the operations. That knowledge decays, and keeping it current takes time.

Ronald Gilson and Jeffrey Gordon described today's directors as thinly informed, under-resourced and boundedly motivated, and wrote of the monitoring board that it simply does not scale. Their Board 3.0 borrows from private equity, with a group of empowered directors who have more time, real equity and analytical support of their own. The portfolio of five or more non-executive seats was built for a monitoring board. If boards keep moving in this direction, I would expect fewer seats per director, higher fees, more working chairs, and more boards anchored by an owner close to the business.

Danish boards start with an advantage. The Companies Act already gives the board responsibility for the overall and strategic management of the company, and Nordic boards with an anchor owner have long worked with managers below the CEO. BCG found 77% of anchor-owned Nordic boards interacting with management below the CEO, against 53% where ownership is dispersed. A Danish board needs no new mandate. It needs different people and a different cadence for the decisions it takes in calm.

Where it does not hold

In a company of thirty people the chair is often the most experienced operator available, and the board is part of the management bench whether anyone calls it that or not.

When survival is at stake, the board takes charge, and it should.

Events nobody wrote a trigger for are not covered either. The first weeks of the pandemic had none. What helps then is a board that can convene quickly and decide on incomplete information, which is a question of who sits on it.

Sources

• Kathryn Landis and Julie Kratz (2026), What to Do When Your Board Is Meddling in Operational Work, Harvard Business Review, 5 March 2026.
• Ronald J. Gilson and Jeffrey N. Gordon (2019), "Board 3.0: An Introduction", The Business Lawyer 74. The characterisation of directors and the private-equity model are theirs.
• BDO (2026), 2026 BDO Board Survey.
• Melissa Sawyer, Lawrence S. Elbaum and Jacob M. Croke (2025), Director Data Requests: The Line Between Oversight and Management, Harvard Law School Forum on Corporate Governance.
• Ram Charan, Dennis Carey and Michael Useem (2014), Boards That Lead: When to Take Charge, When to Partner, and When to Stay Out of the Way, Harvard Business Review Press.
• BCG (2016), How Nordic Boards Create Exceptional Value.
• Danish Companies Act (selskabsloven), § 115.

I started with principles, board notes, meeting recaps and decision records, with the aim of a system that knows how I think and tells me when I am not acting on it. For the first weeks it was mostly noise. It raised questions that did not matter and found patterns that were not there. Then it told me that I was again spending too little time with the stakeholders who mattered most for what I was trying to do. It was right, and I changed my calendar. That is one correction, on one person, and it is the kind of thing a good chief of staff would have said. What made me take it seriously is that the chief of staff would have needed to be in every meeting for a year to say it, and the system had been.

The organisation remembers the decision and loses the reasoning

A transformation programme is approved because the rationale is compelling, the benefits are attractive and management believes it can deliver. The people put in charge are capable; that is why they were chosen. Three years later it is struggling, and the account changes to poor management, unrealistic targets and the wrong people. Some of that may be true. The questions that are rarely asked are who made the original assumptions, what was known at the time, what changed, and whether the organisation supplied the conditions it had promised.

There is a structural reason. Major initiatives are approved on optimistic assumptions, because the people proposing them believe in them and because an ambitious case is approved more often than a cautious one. When the assumptions fail, accountability travels to the people executing. The business case usually still exists; what is missing is what was known at the time and who assumed what, and nobody has an incentive to reconstruct it, since the people who approved are the ones who would be examined and the people executing are the ones still in the building. The organisation moves on, and the next initiative is approved on remarkably similar assumptions. Outcomes are measured; the reasoning that produced them is not kept.

The drift nobody decided

The larger version of the same loss is drift. A company states an ambition to become customer-led. Over five years its decisions on hiring, incentives, technology, structure and cost each make sense on their own and collectively move it toward internal efficiency. Nobody intended it and nobody noticed, because nobody was present at all of the decisions. A CFO sees the investments, a CHRO the appointments, a COO the restructurings. The pattern lives in the sum, and the sum has no owner. The CEO owns the whole in theory; in practice nobody has the memory or the attention to hold every significant decision of five years in view at once.

This is where a system with the record becomes interesting. It cannot decide better than the executive. It can be the one party with the whole record in front of it, every decision, the assumptions behind it and the corrections since, and the one that can be asked whether the sum has started to contradict the stated direction. The same holds for the long bets in the previous piece. A board asked to review whether a thesis still holds can only do so if the thesis was written down when the money was committed.

Record first, challenge second

I would separate two functions that are usually sold together.

The record is the more dependable part. After an important meeting I write a short recap, or hand over the notes, covering what happened, what was decided, why, which assumptions and what remains open. It takes ten minutes. For a CEO the place to start is the executive committee, with its strategy papers, investment proposals, minutes, and the recaps of the conversations where the decision was actually made, which are usually before the meeting. Mail and calendar can be added later, but only with explicit rules for access, privilege, retention and who may see what. The system gets things wrong at first; the corrections improve the record and show it which things matter.

The challenge is the less dependable part, and I say that from my own corrections. In a live session the systems I run generate forward and drop their own premises on the way, so an assumption already in the conversation goes unused. Asking for a challenge while the system is working produces little. Given a closed record afterwards and one question, it finds the inconsistencies well. So the challenge works as a separate pass over the record, and it works best where the record is thickest. The judgment on what it finds stays with the person who owns the decision.

A useful side effect is that the system can tell which kind of decision is in front of you. Most decisions deserve less analysis and faster correction. A few deserve far more than they get; a senior appointment shapes capability and culture for a decade and often receives less scrutiny than a mid-sized capital request. The weight of a decision is rarely reflected in the process used to make it, and a record of past decisions and their consequences is the evidence for changing that.

Why this sits with the CEO

Strategic decisions play out over five to fifteen years; the people who made them rarely stay that long. The record is the one thing that can outlast them, and it outlasts them only if someone owns it. The CEO is the person responsible for connecting strategy, allocation, execution and the development of the organisation, and for whether the executive team learns at all. That responsibility is the reason to start there, and the reason it does not start as a company-wide programme. The CEO owns the discipline; the organisation has to end up owning the memory, or the record leaves with the person who kept it.

Where this does not hold

The record is only as good as what gets captured, and the decisions that matter most are often made in conversations nobody writes down; a thin record produces confident drift reports about nothing. A good decision can produce a bad outcome and a bad one a good outcome, and the system can separate reasoning from luck only if the reasoning was recorded in advance; recorded afterwards it is a rationalisation. A complete record is also a concentrated map of strategy, disagreement and personal judgment. Weak access control makes it a liability, and the knowledge that every remark will be retained may make executives less candid. The challenge function is, on my own evidence, unproven beyond one user who corrects it daily. An executive team that does not correct it will get a sophisticated confirmation machine.

Sources

• Asutosh Padhi, Brian Elliott, Daniel Swan, Kim Baroudy and Vik Sohoni (2026), The AI operating system: a new way for CEOs to run the enterprise, McKinsey, 13 October 2026. The source of "second brain" as a name for an enterprise AI operating system.
• Jesper E. Thomsen (2026), AI tilts capital toward what can be measured. The long bet loses on the business case.. The thesis review this record exists to serve.
• Jesper E. Thomsen (2026), An AI that reports nothing found cannot tell you whether it looked. The evidence that the live challenge fails and the retrospective pass works.

Two clocks

An organisation makes two kinds of decision. One kind can be checked against information that exists, and checked soon, such as margin, utilisation, conversion, working capital and cost to serve. The other kind rests on assumptions that take years to validate, such as which capabilities to build now, which markets will matter, which technology to commit to before its value is visible. The second kind can only be judged against information that does not yet exist, and no model changes that.

AI can move the first kind onto a continuous clock. Opportunities are found, modelled and executed as they appear, with a case attached. The second kind needs a different clock, one on which its assumptions can be examined without reopening the allocation every time a short-term alternative appears. In most companies that is the strategy round, because it is the only place such decisions can be argued at all. The company now runs two cadences for capital. The faster one produces a stream of documented proposals; the slower one produces a few that rest on judgment.

There is a further twist in the measurement itself. Much of what this quarter shows is the consequence of decisions taken years ago, the capability that was or was not built, the market that was or was not entered. We still use those results to judge today's management and to justify tomorrow's allocation. AI makes the numbers arrive faster and with more precision, and precision says nothing about which decision produced them.

The burden of proof moves

Put the two kinds of proposal in front of the same board. One arrives with a model, a payback and a comparison with last quarter, and the case can be refreshed every week. The other arrives with a thesis, a horizon and a person willing to stand behind it, and it looks much as it did last year. Boards do fund the second kind; they always have. What changes is that it now needs stronger sponsorship to get through, although nobody decided to shorten the horizon. The burden of proof moves on its own.

Two things push it further.

The fast clock does not reliably use its own history. A finance partner who has cut the same cost three times says so the fourth time. An agent optimising against a target may find the fourth cut and present it as cleanly as the first, even with the previous three in its context. Having the information is not the same as using it, and I have written elsewhere about how reliably that gap appears. The remedy is a constraint on what a proposal must contain. It has to show the cumulative cut to date and the capability it draws on, so the fourth cannot be written without showing the first three. It is a narrow design question.

The early AI cases I see also show value on cost before growth. An analyst producing ten analyses instead of two is not creating five times the value, and released capacity is easily absorbed into more activity. The visible gains come from stopping work and removing cost, which is the measurable side of the ledger. The organisation learns first that AI pays on that side, and the lesson becomes the bias.

The slow-clock review

The responsibilities do not change. The CEO still answers for this year's performance and for the business that has to compete in ten. What changes is that one side of that pair now arrives with better arguments every week, and the other side needs a scrutiny of its own kind to stay in the contest.

That scrutiny has two properties. It runs on the slow clock. The allocation is made where the long decisions are made, and the question of whether to continue is asked there, not every time the fast clock produces a better case. And the question it asks is whether the thesis still holds, rather than whether the bet has paid back, which by construction it cannot have. A thesis review has three outcomes, and they call for different responses. Either the assumptions hold and execution is slower or dearer than planned, or the environment has moved and the thesis needs rewriting, or the original reasoning was wrong and continuing compounds it. The second and third are reasons to stop or change. The first is a reason to examine whether the route to the objective is still the right one, since a capability can stay strategically necessary while building it in-house stops making economic sense and buying or partnering takes over. It is not a reason to abandon the objective, and it is the outcome the fast clock would use to do exactly that.

A board reading an AI-enabled plan should ask where in it the decision that cannot be measured sits, who signed for it, and when its thesis is next reviewed. That review needs the original reasoning to exist somewhere retrievable, which is the subject of the next piece.

When the tilt is useful

Where the business is in distress, the measurable is survival, and the tilt is correct. Where the industry's own clock is short, the ten-year bet may not exist in a form worth protecting. And the argument assumes the long bet stays unmeasurable. If AI makes scenario work and capability valuation good enough that the thesis arrives with something resembling a model, the contest gets fairer on its own. I do not expect that soon, because the missing input is the future, not the analysis.

Sources

• Asutosh Padhi, Brian Elliott, Daniel Swan, Kim Baroudy and Vik Sohoni (2026), The AI operating system: a new way for CEOs to run the enterprise, McKinsey, 13 October 2026. The description of continuous management, fewer layers and executives spending more time on judgment is theirs; the two-clock argument is this piece's.
• Jesper E. Thomsen (2026), An AI that reports nothing found cannot tell you whether it looked. The finding that a model does not reliably use what sits in its context.

That is different from removing blockers. Blocker language assumes there is a faster organisation underneath, held back by things in the way. If that were true, removing them would work permanently. Often it does not. The obstruction returns under another name because the conditions that produced it are still there. The speed you see is the organisation working as it is built. There is no faster version waiting to be released. There is a different organisation that would have to be built. Blockers stand in the way of that work. Cadence is the ground it stands on.

Two definitions

Cadence is the time from something that calls for a decision to a resource being committed, measured on the decisions that determine whether the organisation adapts. Recognising a problem has its own clock, and so does executing once the decision is taken. The clock that matters is commitment.

Resting cadence is the cadence the organisation goes back to once the extra pressure is off. Every organisation has one. A crisis, a new CEO, a transformation office or a consulting engagement can move the organisation away from it for a while. How far it moves is displacement. How long it takes to come back is the reversion half-life. Neither of those is a change in resting cadence. Only a change in what produces it is.

Efficiency is something else again, output per unit of input inside one process. It gets mistaken for cadence all the time.

What holds the resting cadence in place

Resting cadence is not something an organisation inherits and then carries. It is produced continuously by a set of generators that reinforce each other. They are who has the right to decide and how many layers approve, the meeting and reporting calendar, the budget cycle and how resources get allocated, incentives, who gets promoted and who succeeds whom, what information is available to whom, how much risk and failure is tolerated, who owns the company, and who actually holds power.

Four things follow from that.

Speeds inside the organisation do not add up. Cadence is set by the slowest element on the path every real decision has to pass through, and that path is normally governance. Cutting a development cycle to one week changes nothing if the board approves every six weeks. That is a gain in efficiency, not in speed. It is also why transformation programmes can report real improvements and still change nothing.

What builds up over time is apparatus. Adding a control costs one incident and one decision. Removing it needs somebody to put their name on the risk that remains, and nobody volunteers. In the meantime, behaviour routes around the controls. The organisation gets formally slower and behaves the same, and the gap widens the older it gets and the more incidents it has been through. The end state is easy to recognise. Governance is heavy and eats time without governing anything, while the real decisions travel through channels that are on nobody's org chart.

So cadence does not live only in the formal apparatus. That is the part you can redesign, but the informal system often restores the old pace. A redesign alone rarely reaches who actually holds power.

And the whole set-up reproduces itself through people. Promotion and succession favour those who are comfortable at the current pace, and who read faster colleagues as reckless. Leaders get pulled into the organisation's rhythm at least as much as the other way round.

Why a faster unit inside the company does not spread

A unit running at a different cadence inside the organisation is often rejected rather than absorbed, and strong performance does not protect it. Putting it in a separate structure may only make the rejection cleaner.

I have run both versions. At Maersk, X-Leap in 2010 and the later Simplification programme were both placed outside the line, both delivered, and both were wound down once they came to be seen as running on the side. In both cases the break point was a fast IT organisation up against legacy waterfall IT. The faster body did not lose on performance. It lost on cadence.

When resting cadence does change

The element that binds the organisation's speed is often controlled by people whose authority comes from it. Changing it may mean giving up some of that authority, so the proposal rarely comes from the people usually asked to make the change.

Resting cadence changes when control of the binding element changes, when the organisation's link to its environment changes, or when slack is reduced until the environment's clock becomes binding. By ownership I mean control over the constraint, not only ownership of the company. Attention, pressure and programmes can change the pace for a time. They do not change the resting cadence unless they move that control.

That makes ownership events unusually effective. A buyout, an IPO, a delisting, a merger, an activist coming in, a founder leaving or a forced recapitalisation changes several generators at once. Near-death often has the same effect because it reallocates control and burns the slack that was covering the mismatch, not simply because it motivates people.

Private equity is the strongest confirming case, and a better one than the celebrated turnarounds because the intervention is repeated. PE can change the operating cadence of ordinary companies by changing the reporting calendar, capital allocation, the board, incentives and slack at the same time, from outside. It reaches the informal system because it changes who holds power and who is still in the building.

The same logic explains why a domain hire can fail on cadence. A Chief Digital Officer brought in for digital expertise may have no authority over anything that produces the organisation's speed. I made this argument when two such roles were being designed, and I still think those outside hires were the wrong answer. In those cases, understanding how the organisation worked was the binding qualification. Domain knowledge was not.

What this means for strategy

A strategy that needs a faster cadence than the current one is incomplete until it says how the generators will change. Cadence is a constraint you have to deal with before the strategy, not an execution problem you deal with after. A board reading a plan that assumes a speed the organisation has never sustained should ask which generator changes, who owns that change, and what risk moves to whom.

Fast is not better

Cadence is calibrated to the environment and the risk profile. Pharma and financial organisations are often slow for good reason, because the cost of a mistake shows up over long horizons. Other organisations carry too much risk appetite, and what they need is to slow down. The argument runs both ways and takes no position on which direction is right. It explains why speed is stable. The test of whether anything has changed is whether the organisation keeps operating differently after the extra pressure is gone.

The gap was visible with the PC, the internet and the smartphone. AI makes it harder to excuse because first-hand contact is now so cheap.

Why a briefing does not work

You cannot ask a qualified question about something you have only had summarised, any more than you could have understood the internet from a briefing. The summary carries the conclusions and not the feel for what the thing does easily and where it breaks, which is where the question comes from.

Twenty minutes with a terminal is not expertise. It is enough to remove the excuse for having no first-hand contact at all. A board that stays two years behind on the fastest-moving technology in the business is not governing it.

The no

Operating the technology is easy to mistake for a route to enthusiasm. The highest use of it is the no.

Everything is interesting at the start. The hard, valuable judgement is seeing which technologies will not scale, or do not fit the actual problem, before the organisation spends two years and its credibility finding out. Close up, you can see what the technology needs to deliver its value, and ask whether the business, or the industry, can provide it.

That is two pieces of understanding held at the same time. One is what the technology requires, the other what the industry's trust structure can supply. Hands-on use gives you the first only. Feel on its own produces the enthusiast, who has used the thing and holds half the judgement.

TradeLens

The nearest example is TradeLens, the blockchain shipping platform Maersk built with IBM, launched in 2018 and wound down by early 2023.

A blockchain network of this kind delivers its value through broad, neutral, shared adoption. A network co-owned by Maersk, then the largest container carrier, struggled to look neutral to the carriers it competed with. Several large carriers joined after a governance change in 2019. Maersk's stated reason for closing TradeLens in November 2022 was that full industry collaboration had not been achieved. The governance change did not close the gap between what the technology required and what the industry's trust structure could provide.

The people who built TradeLens had the technology side, from years of operating it on real work. Their yes is the evidence that hands-on use is half the judgement. The industry side was available at a distance to anyone in container shipping. The judgement that saves the most money is the two held together, before the spend.

The board standard

On any technology fast and material to the business, my view is that at least one person at the table should have operated it on real work rather than a demo, and know the industry well enough to hold the other half. Whether one is enough is the board's call. If the capability costs twenty minutes, the case for stopping at one is weak, and one director who has used it is still, for the rest of the table, a description. On AI-native boards I apply it to myself. If I do not understand this, I have no role there.

None of this makes the director the expert. You should rely on people more intelligent than you, and the technical questions are theirs. Whether the business or the industry can supply what the technology needs is the director's question, and someone proud to say "I am not an IT expert" cannot ask it.

The limit

Not every technology is fast, material and yours to govern, and not every director must code. What the rule asks is that you operate the specific technologies reshaping your specific business well enough to ask a qualified question, see the link to the business, and tell the ones that will scale from the ones that will not.

It also does not hold where the industry is not yours. Contact with a technology in an industry whose trust structure you do not know buys a qualified question, and not the no.

A board is not democratic in the literal sense. Voting power follows shares, rights may differ between share classes, and the nomination process may not offer much of a choice. But it is built on some of the same principles as representative government. Authority is conferred through election or appointment. Members exercise judgment collectively and can be removed or not re-elected. Elective governance is a better description than democracy. If nomination and election are mostly ceremonial, the selection mechanism is weak. That is further evidence of the selection problem. It does not turn later silence into a psychological safety problem.

That matters because dissent is not a favour granted by the chair. It is part of the mandate. A board is not a group of advisers who offer opinions and leave management to decide. Its job is to govern. It appoints and oversees management, decides the matters reserved to it and carries responsibility for the decisions it makes.

The team work is useful where it belongs. An employee depends on the manager they may need to challenge. A non-executive director is chosen partly to stand outside that hierarchy. Independence, judgment and the willingness to become unpopular are qualifications for the role, not conditions the rest of the board must create afterwards.

Two failures

The first failure belongs to selection. A member may sense that something is wrong but be unable to turn the discomfort into an objection that will hold up in the room. That is a competence problem. Another member may see the problem clearly but decide that the next mandate or the relationship with the chair matters more than saying so. That is an independence problem. No amount of psychological safety produces an objection that does not exist, and it should not be needed to release one that does.

The second failure belongs to the chair and the board's process. The chair cannot supply a director's nerve. The chair can make sure a material objection, once raised, is heard, tested and resolved. That means allowing the time, requiring a position, preventing management or a dominant shareholder from closing the discussion, calling a vote when needed and recording dissent.

The comparison with parliament is useful here. A member is expected to speak and vote without first being made comfortable. The institution provides a route from dissent to decision through speaking rights, motions, votes and a record. Those are procedural protections, not psychological safety. A board needs its smaller equivalents.

This gives a clean division of responsibility. If a director sees the objection and stays quiet, that director should not have been selected. If the director raises it and the board never deals with it, the chair failed to run the decision.

The channel outside the format

The most expensive channel is not in the meeting format at all. A board decides on curated information. What reaches the agenda depends on what the executive dares to put on the table, which in turn depends on the CEO's reading of how the board will receive it. No decision format can deal with an item that was never tabled.

It is the chair's responsibility for a plain reason. The CEO sits in the same room as the members and reads the same chair. The price the chair sets on a member's objection is the price the CEO reads on tabling a difficult item, and that is the only route from the meeting to the channel outside it.

What to select for

The selection signal is the candidate who has said no while a pipeline was at stake. That candidate has delivered the proof references cannot. References on this point are skewed rather than absent, because the witnesses to a dissent are often the people it annoyed. So weight references from situations where the objection cost the person something.

Money is not the currency. Fee dependency disqualifies, and the governance codes already codify that, but it is a floor that removes candidates; it does not explain the variance among those who pass. My working hypothesis, and it is a hypothesis, is that the currency is pipeline and relevance, meaning the next mandate, the deal flow and the rotation. That channel is weakest for established seniors, where personality and competence carry the variance, and strongest for the member building a portfolio of mandates, whose next seat comes on a recommendation from the people in the room.

The chair is often the principal recommender for that next seat. That is how the chair's price reaches the members. The room has little repetition with which to change the price; the chair's word when another seat is filled carries it much further.

The chair's test

A chair's job is not to make directors feel safe. It is to ensure that a material objection reaches the decision and is dealt with. The price of dissent is still set by you, at the point of decision, and collected in what you say about a member when the next seat is being filled. Putting psychological safety on the agenda says nothing about that price.

For a board member, one question from a quarterly check I have set up for the boards I sit on bears on this. Would I hire the current CEO for this job today, knowing what I now know? If no, or a hesitation, what is the plan, because that answer does not improve by waiting a quarter. "Don't know" counts as red. A red answer that stays private is the silence described above, so the check carries a rule. Two red answers on the same company two quarters in a row become an agenda item at the next board meeting, not another observation.

Limits of the claim

The phenomenon is real. Groupthink and curated information happen on boards, and I am not arguing otherwise. The claim is that psychological safety confuses a failure of office with a feature of the environment. It gives selection failure a more flattering name and gives the chair the wrong job.

The incentive channel is a hypothesis. If the price of dissent turns out to be set inside the room after all, the selection signal weakens and the chair's route to the members runs through the meeting rather than the seat market.

Everything above is argued for boards. Whether it carries to other episodic forums, such as investor circles or advisory boards, is untested. They share the missing substrate, but the incentive channel may run differently, and I have not checked.

I run a small fleet of AI agents over my own notes and correspondence, because that is the domain where I can judge the output. The two findings below come from a month of my own corrections. I think both apply to any AI reporting, monitoring or compliance deployment.

An AI "no" is easy to treat as a weaker version of an AI "yes". That gets the shape wrong. A positive claim usually arrives with the thing it found attached. A negative claim arrives with nothing, and the output often does not tell you what was actually searched.

Absence reported as fact

Retrieval returns a subset of the corpus. The model treats the subset as the population. It then generates a negative claim ("nothing outstanding", "no matches") with exactly the confidence it gives a positive one, and nothing in the output separates "not in what I retrieved" from "not true".

Five instances in one month, five different agents, each caught by a correction of mine. The agent that watches bills reported no new bills while a bill from the 14th sat unprocessed; my whole correction was "from the 14th?". The agent that tracks correspondence reported ten days of silence on a thread. It was three. A weekly review raised as still open a matter we had already cleared up. None of the five read differently from a run where the claim was true.

The one agent in the fleet I have never corrected, across three runs and roughly sixty findings, is the one whose output is a quotation, a line number, an issue and an action. That is a small sample, but it cannot say "nothing" without pointing at what it claims to have read.

The control case suggests a distinction. A rule that asks the model to judge itself ("check whether you have searched everything") did not help inside the pass. A constraint on the shape of the output ("every negative reads 'not found in [named source]'", "quote the line") makes an unsupported sentence harder to pass and gives a downstream check something concrete to verify. The quotation still has to be checked against the source; the model can invent that too.

The model does not join its own premises

Working through a board financing note for a company I chair, in a long session with many established facts, the assistant quoted the CEO's own travel gate, then flagged the absence of a pre-agreed exit criterion. Its own text, earlier in the session. There were several more in that session, and in each the missing input was already in its context, written by itself. (A further miss, an option set that did not map onto a structure set out days earlier, I leave out; the structure may not have been in view.)

This failure produces negatives too. "No exit criterion agreed" is an absence reported as fact with the record in front of the model. So the first fault has two causes, and the output-shape constraint reaches one of them. A quote requirement makes the model point at what it read, not read what it wrote.

The observable failure is that the model did not use a premise already in its context. I do not know whether the mechanism is attention dilution, retrieval, or something else. What matters operationally is that putting a fact in the context did not make the model use it.

Fourteen standing rules and nineteen operating principles were in context during that session, and none fired. An instruction of that kind asks the same pass to produce and critique at once, and in this case it did neither reliably. The same model, given the transcript afterwards, diagnosed the misses precisely. Same model, different mode. The candidate fix is architectural. It is a second pass, with the finished artefact as input and a clean context. A pass of that kind exists in my fleet, still unvalidated.

The second pass carries the first finding one level up. Its usual output on a clean artefact is "nothing found". Unless that sentence is under the same source constraint, naming the artefact and the lines checked, the reader holds a negative with nothing attached, from the pass meant to catch them.

What to ask a vendor

Two questions worth asking a vendor of AI reporting, monitoring or compliance before signing.

• Show me a negative result. Where is the quotation and location of what was read, and what system trace proves the source was retrieved? A generated list of sources does not count. The demand is for both an output that can be checked and a trace outside the model's own prose.
• What reads the output before you do, and is it a separate call? A second pass with the artefact as input and a fresh context is one answer. An instruction in the same call to review the work is not, however many rules it contains.

A separate argument holds that an agentic layer in front of the old stack lowers programme risk in a migration. This is the risk it puts back. The layer reports on the systems behind it, and its negatives are the ones nobody checks.

Limits of the finding

The first failure is proportional to how much of the corpus the agent can see in one pass. Where the whole record fits in front of the model, or the agent is constrained to quote, the retrieval cause disappears, though the join cause does not. The claim is about systems that summarise more than they can read in one pass.

The second failure is worse in long sessions under production pressure, and absent or much reduced in retrospective analysis of a closed artefact with one question in front of it. That is the reason to expect a second pass to work. The pass has found misses known to be there; untested is its false-negative rate where nobody knows whether a miss exists, and that is the test. Does it surface a join I would otherwise have had to supply, or does it return reformatting? If reformatting, the mechanism is misdiagnosed and the problem is retrieval depth, not reasoning.