The return on silence

There is a scene everyone recognises and nobody tells the same way. On one side, a person doing things with AI that their organization has no idea about, waiting to be noticed. On the other, a leadership team that sees nothing, measures nothing, and wonders aloud why AI is producing no visible result in their house. Each blames the other's silence. Both have their reasons, and that is exactly the problem. What holds them there is not a misunderstanding. It is a mechanism, and that mechanism has a price you can calculate.

What nobody reports

Your people are not hiding anything from you. They have priced what speaking up costs.

Start with the measurement, because it is solid and it is global. In 2025, the University of Melbourne and KPMG surveyed 48,340 people across 47 countries about their relationship with AI. More than half of employees say they do not report their use, and have passed machine-generated content off as their own. Nearly one in two admits to using AI against company policy.^[1] In June 2026, a survey of 1,250 non-technical professionals at companies above five hundred million dollars in revenue found that two thirds had used unauthorised AI tools at work.^[2]

In France, the annual barometer run by Julhiet Sterwen with the polling institute Ifop, on 1,003 employees of large companies, measures the same thing from another angle. AI use went from 38% in 2024 to 62% in 2026. Over the same period, only a third of employees say they have any formal rules for using it.^[3] Use doubled. The frame did not follow.

I know this gap from somewhere other than the surveys. Among the three hundred and fifty or so professionals I train each year, the question that comes back most often is not technical. It fits in one line, and it is always asked quietly: can I tell my boss.^[21]

The problem is not ignorance. It is accounting.

Let us dispose of the lazy explanation, the one running through every article on the subject: management does not understand AI. It is false, and the French numbers say so bluntly. In the same barometer, 85% of managers use AI, against 44% of the people they manage.^[3] Managers are not behind on the tool. They are behind on something else entirely: the ability to see what their teams do with it, and to pay for it.

I have described elsewhere how this revolution happened from below, without permission and without a rollout plan, in The revolution that never happened, and how the worker now brings the machine, in BYOAI, when the worker brings the machine. Those two essays establish the fact. This one takes the question that comes next, which I had not treated: why does the fact stay invisible when everyone would gain from it becoming visible.

The three prices of silence

Staying quiet is not a character trait. It is a calculation, and it pays.

An employee who creates value with AI and says nothing is not sliding down a psychological slope. They are paying or avoiding costs, and picking the cheapest combination. Three of those costs are social, documented, and all three land on the person who speaks. None land on the person who listens.

The price of disclosure

Say you used AI. You will be read as lazy.

In May 2025, three researchers at Duke's Fuqua School of Business published four preregistered experiments on 4,439 participants in the Proceedings of the National Academy of Sciences. The finding is blunt: someone who uses AI is judged lazier, less competent, less diligent and more replaceable than someone using ordinary tools. Participants correctly anticipate that judgement, and report being less willing to tell colleagues and managers about their use. In the third experiment, 1,718 participants acting as hiring managers screened out candidates who used AI, with one exception I will come back to.^[4]

That same year, two researchers published thirteen preregistered experiments on more than three thousand people in Organizational Behavior and Human Decision Processes. Their conclusion has a name, the transparency dilemma: the person who discloses their AI use is trusted less than the person who keeps quiet.^[5] Disclosure costs. Not morally. Statistically.

What remained was the price in career currency. In spring 2026, a controlled experiment on 961 American knowledge workers had people evaluate an identical piece of work, presented sometimes with a mention of AI and sometimes without. The one who discloses is rated ten times lazier, and loses 24 percentage points of likelihood of being recommended for a high-visibility project.^[6]

Twenty-four points is the sticker price of honesty. Nobody pays it twice.

And yet the same study carries the way out, which is what makes it valuable. In companies that openly celebrate AI use, the penalty nearly disappears, to the point that people who disclose are rated more efficient than people who say nothing.^[6] The Duke experiment shows the identical flip on the evaluator's side: recruiters who use AI daily stop penalising candidates who use it, and start preferring them.^[4] The price is not a fact of nature. It is a fact of culture. Which means it can be changed.

The price of what nobody asked for

An organization does not recognise what it never ordered. This is not bad faith.

Suppose the employee speaks anyway. They then run into a second mechanism, older than AI and thoroughly documented. In 2012, three researchers published a set of experiments in Psychological Science on what they call the bias against creativity. The result is uncomfortable: people reject new ideas while sincerely claiming to want them, and the rejection switches on the moment they are placed under uncertainty. The last detail is the one that matters most. The bias degrades the very ability to recognise a creative idea when it shows up.^[7]

Translate that. The manager who does not hear their employee's proposal is not pretending. They do not see it. And they see it less the more uncertain the moment is. Generative AI is, in the life of organizations in 2026, the largest uncertainty machine available. The mechanism fires at exactly the moment it does the most damage.

On top of the bias sits something more mundane: the instrument does not exist. The consultancy WTW went through the 2026 pay filings of the S&P 500. Eight per cent of those companies include any AI-related metric in their executive incentive plans.^[8]

If the top of the pyramid is not paid on it, nothing is going to come back down. We do not measure what we do not pay for, and we do not pay for what we do not measure.

None of this is new, and organizational research named it long before AI. In 1988, Dennis Organ defined organizational citizenship behaviour in terms that now read as a prophecy.^[20]

Individual behaviour that is discretionary, not directly or explicitly recognized by the formal reward system, and that in the aggregate promotes the effective functioning of the organization.

The employee automating in silence is waiting for recognition from a system whose own definition rules out what they are doing. I argued in Your organization captures nothing from AI that this incapacity is architectural rather than a matter of goodwill. We are in the same building here, but in a single room: the one where two people talk past each other.

The price of handing back time

Hand back the time you save. You will be handed more work to fill it.

Third cost, and the most concrete of the three. A 2026 American survey of 1,003 full-time employees found that two thirds stay logged in or look busy once their work is finished, spending close to five hours a week on the performance of being productive. More to the point, 64% say they have deliberately slowed down to avoid finishing early, because finishing fast creates further expectations.^[9]

The mechanism is old. AI makes it explosive. When the same task drops from two hours to ten minutes, the freed time has two possible destinations: it goes up, or it stays. The survey Anthropic ran on 81,000 of its users, published in April 2026, says where it goes. Among those who name a beneficiary of their gain, most name themselves. One in ten reports an employer using it to ask for more.^[10]

I have written elsewhere about what that time becomes once it stays with the person who produced it, and why productivity measures are structurally unable to see it, in AI's real payoff is not productivity. Keep only what matters here: handing back time earns you more work. That is a price, and it is paid by the person handing it back.

The fourth price, the one nobody counts

Doing takes ten minutes. Approving takes three weeks. Do the arithmetic they already did.

The first three prices are social. They explain why people stay quiet, but they do not explain why the best users, the ones with nothing to fear, the established and the well regarded, ask for nothing either. There is a fourth price, and it is not social. It is arithmetic.

For two centuries, making was expensive and slow, deciding was cheap and quick. Designing a tool, setting up a line, building software: months, a budget, a team. In that world, putting an approval step in front of every act was entirely rational. The procedure protected a heavy investment, and its cost, a few meetings, was negligible against what it protected. Every corporate scheme we know was calibrated on that ratio.

Generative AI crushes the first term and leaves the second untouched. Building a useful work tool no longer takes months. It takes an evening. Getting clearance to use it still takes three weeks, sometimes a quarter, because a decision goes through people, agendas and trade-offs that have not sped up by an inch. The ratio inverted. The procedures did not move.

Permission costs more than the act.

That is the whole thing. And when that happens, nobody asks for permission any more. This is not insubordination, and it is barely even a choice. It is the outcome of a calculation anyone runs in three seconds. Someone who has worked out that they can solve the problem before the end of the meeting where they were supposed to raise it is not going to raise it. They are going to solve it. I described that shift on the desire side in Where there's a way, there's a will: exposure to what is possible triggers wanting before motivation. What I would add here is that exposure to what is possible triggers bypassing too, and for the same reason.

A procedure that costs more than what it protects is no longer protection. It is a toll.

This fourth price is the only one specific to generative AI. No earlier technology collapsed build time this far while leaving decision time intact. The spreadsheet, email, the smartphone at work all demanded months of learning and rollout, which gave committees time to keep up. Here there is no delay left to occupy. And an organization whose decision time stays longer than its members' execution time stops, mechanically, being consulted.

Why none of the existing schemes work

Your schemes were designed for a world where making things took months.

There is a serious literature and there are serious programmes for surfacing employee initiative. Some are old, they work, they have proved themselves. Run them through the fourth price and see what is left.

Yesterday's schemes, run through the clock

A suggestion box takes three weeks to process what took ten minutes to build.

The most elegant is called Kickbox. Adobe launched it in 2013 and opened the method under a free licence two years later: a red box, a prepaid card loaded with a thousand dollars, no prior sign-off. The employee tests their idea however they like, and arbitration only arrives at the next stage, when they ask for larger funding.^[11] The principle is excellent. The mechanics are wrong for our subject: it opens with a two-day workshop, then a funding request. Two days to authorise what takes an evening.

The best documented is German. The formal employee suggestion system there is institutional, and its numbers are published every year. In the 2012 panel: one hundred and sixty-four proposals per hundred employees, an average award of about six hundred and forty euros to the person who proposed, and roughly seven hundred million euros in declared savings.^[12] Proof enough that an award system can run for decades. But its cycle is submission, review, decision, payment.

The most generous is American and very recent. In May 2026, a large audit firm launched cash awards of up to five thousand dollars for AI use with measurable results, across a population of eighty thousand people.^[13] The intent is right and the money is serious. The design flaw sits elsewhere: nominations come from managers. So the scheme rewards, by construction, the people management could already see. It misses precisely the population it claims to reach.

The most solid, finally, is the oldest. Gain sharing, formalised in 1935 as the Scanlon plan, organises employee improvement proposals, their review by joint committees, and the sharing of the resulting savings on a formula known in advance, with the employee share running around three quarters. Review literature and longitudinal studies document a durable drop in grievances and absenteeism.^[14] French law already carries the same principle for inventions: article L611-7 of the intellectual property code requires an employer to pay additional compensation to an employee who invents in the course of their job.^[15] The idea that value created beyond the wage gets paid for is neither exotic nor militant. It is in the statute book.

But a sharing agreement gets negotiated, a committee convenes, a financial year closes. All of these mechanics are good, and all of them add latency to the one activity whose entire advantage is the absence of latency. They are not insufficient. They are out of time.

The same remark applies to the answer that comes to mind first, the amnesty survey. Surfacing usage under a written no-penalty guarantee does work, and security specialists now recommend it over outright bans.^[16] But let us be honest about what it produces. A survey buys visibility. It buys no gain. It is an instrument of measurement, not an instrument of exchange. The organization finally sees what is happening. The employee has still received nothing.

What survives, and why

Two mechanisms hold. What they share is that they only arrive afterwards.

Run the filter the other way and ask what resists. Two things only, for a single reason: they are downstream. They require no prior permission, so they slow nothing down.

The first is two-tier compensation. The fast tier is named recognition: the practice carries the name of whoever found it, and whoever spreads it gets a role, protected time and training. The slow tier is gain sharing on a collective formula, written before it is owed. The second is sharing the time back: a rule stating in advance what share of saved time returns to the organization and what share stays with the team. Neither asks anyone to wait for clearance before acting.

Out of that comes the rule that governs everything else, and the one line to keep if you keep only one from this essay. We are not moving from an uncontrolled system to a controlled one. We are moving from authorisation to recognition. The organization stops approving acts one by one, which it no longer has the capacity to do at the rate they arrive. It sets a perimeter once, and it pays for results afterwards.

Then comes the reflex objection, the one that lands the moment anyone says the word bonus: paying would kill the motivation. It is a respectable fear and it is wrong. In 2017, three economists published a field experiment in the Review of Economics and Statistics, run inside a large technology company, with nineteen teams randomly assigned to treatment and control. Rewards substantially raised the quality of ideas. They broadened participation, and the broadening persisted after the reward stopped. The authors found no crowding out of intrinsic motivation. One caveat, and it is a design caveat: rewards did not raise total volume, because more people proposed and each proposed less.^[17] Which is to say, you do not steer a scheme like this on contribution counts.

As for the time, giving a share of it back now has a name and an unlikely champion. In April 2026, OpenAI published a policy paper urging companies to pilot a thirty-two hour week at full pay, under the term efficiency dividend: what the machine frees should come back in part to the person doing the work, and not only to the operating margin.^[18] You may find the messenger interested. You will struggle to find the argument wrong.

The trap of the individual gain

Keep it all. You keep the ceiling too.

Now for the uncomfortable part, and I am addressing it to the reader who recognises themselves in the quiet employee, because nobody else is going to say it to them.

Someone who produces a great deal with AI and says nothing has three exits. The first is going independent. It is real, it works for a few, and it means carrying a risk most people neither want nor can carry, often for excellent reasons that have nothing to do with courage. The second is selling the skill on the labour market. It pays: an analysis of more than a billion job postings across twenty-seven countries puts the wage premium on AI skills at 62% in 2026, up from 57% the year before.^[19] But that premium is collected by leaving, not by staying. The third is keeping the time and saying nothing.

Almost everyone picks the third, because it demands nothing. And it is the least profitable of the three. Time saved in silence does not accumulate, cannot be attested, cannot be transferred and cannot be argued. It buys comfort, which is not nothing, and I have argued elsewhere that comfort is in fact the main gain of this technology. But comfort is not capital.

Silence protects. It accumulates nothing.

The market has already started deciding for the undecided. In the survey on unauthorised use quoted above, three professionals in four say they would move to an employer that developed their AI skills better, rising to four in five at the largest companies.^[2] That number is not a warning to employees. It is a warning to leadership: the value they cannot see, somebody else is currently buying.

And to be fair, the clarity has to be turned back on the person waiting. Expecting recognition for a contribution you never declared is not discretion. It is a contract signed by one party. The organization never agreed to it, for the good reason that it was never shown to them. Silence is rational, and it remains a losing bet over time. Saying so is not submission.

Ten moves that make it pay

Nobody needs more goodwill. Everybody needs an honest price.

What follows is not a charter, and it is certainly not an invitation to be transparent. You do not repair an asymmetry with morality. These ten moves ask nobody to be a better person. They change prices: the price of disclosure, the price of novelty, the price of time, the price of permission. They hold together, and the order matters, because each one makes the next possible.

One warning, and it is a serious one. These ten moves are a grid, not a protocol. Their dosage depends on what you are looking at: size, legal status, the share of seasonal contracts, the state of the management line, what existing agreements already allow. The move that unlocks a twelve-person organization is not the one that unlocks a three-hundred-person company. And nobody picks that dosage without first measuring what circulates and what goes unsaid, organization by organization. That is the work I do with my firm. Applying the ten moves in order without that measurement would repeat the very mistake this essay describes: deciding in the place of the people who know.

What the organization decides once

Set a perimeter. You will not have to decide a thousand times.

First move, decide a perimeter rather than acts. Publish what is forbidden, what needs an opinion, and what is free, and state that anything unnamed is free. A perimeter is decided once. An authorisation is requested every time. This is the only known answer to the latency asymmetry, and it conditions everything after it: as long as the rule is to ask, the rule will be bypassed.

Second move, sign the guarantee before asking the question. The no-penalty commitment is written, signed and circulated before the first survey goes out. A verbal guarantee weighs nothing against a penalty measured at twenty-four points, and employees know that better than their leadership does, since they are the ones paying it.

Third move, pay for the tool without demanding the result. Reimburse the personal subscription, capped, with no usage audit, and no consideration other than logging the practice in the shared register. It is the only move that settles an entire asymmetry for the price of a pizza a month, and it removes in passing the most legitimate grievance a team can hold: today they are personally financing their employer's means of production.

Fourth move, write the sharing rule before knowing the gain. The profit-sharing formula and the time-sharing rule are published in advance. A rule decided after the gain is no longer a contract, it is a favour, and everyone works that out within a second. This is what sharing plans have held on to for ninety years, and it is why they hold.

What managers do every week

A manager who does not use the tool penalises the person who does. Without meaning to, without knowing it.

Fifth move, practise before evaluating. This is not an exhortation to be modern, it is the direct consequence of the Duke experiment: the same evaluator, depending on whether they practise, either screens out or prefers the same candidate. An organization that opens a reporting scheme before putting its management line into the practice has installed a bias and called it judgement.

Sixth move, ask about the method and not only the result. As long as how the work got done is not an ordinary item in a team meeting, alongside the numbers, it stays invisible. And what stays invisible does not get paid. The question runs to five words, it costs nothing, and it can be asked every week: "how did you get that result?"

Seventh move, answer within the week, including to say no. Silence after a disclosure kills reporting schemes far more reliably than refusal does, because it reads as disavowal and it teaches people not to try again. A published deadline that is honoured is worth more than a promised bonus.

What the employee has an interest in doing

Stop waiting. Recognition is not earned. It is prepared.

Eighth move, report a repeatable practice rather than a feat. The framing decides the reception. A feat invites judgement, with every chance of it going badly. A method invites reuse, and reuse is what protects its author. What is told as an achievement gets judged. What is shown as a procedure gets copied.

Ninth move, aim for reuse rather than recognition. A practice picked up by five colleagues is a fact, verifiable and dated. Expected recognition is nothing, and it cannot be argued anywhere. It is also the only indicator that routes around the bias against novelty, since it asks nobody to judge an idea, only to observe that it was copied.

Tenth move, negotiate at the moment of the gain, not at the annual review. Leverage exists while the practice is visible, reused, and while its withdrawal would be noticed. Twelve months later it has faded into the scenery and the conversation turns into a request. Of every move on this list, shifting that calendar is the one that changes the most for the person who applies it.

We do not pay for what we do not look at. We do not say what it costs to say. Between the two there is no misunderstanding. There is a price.