The AI proof gap: could you explain that decision?

First, a question

Guess before you read on.

Grant Thornton asked nearly a thousand senior leaders in the United States one flat question this spring. Could your organisation pass an independent review of how it governs AI, within ninety days? Take a guess at the answer before you see it, because the number lands differently once you have committed to one.

Out of every hundred senior leaders, how many were fully confident they would pass?

50

Almost nobody guesses this one right, and the direction of the error is always the same. We guess the room is more in control than it is, because the rooms we sit in look calm.

Sound familiar

The same company, described twice.

Left is what gets said in the leadership meeting. Right is what somebody two floors down would tell you about the same thing. Both are honest.

In the leadership meeting

Two floors down

We sent round AI guidelines in the spring.

Nobody could name who enforces them, or what happens if you do not.

Our people are ready for this.

Depends who you ask. Your CIO and your COO give answers five times apart.

We approved the investment carefully.

The board signed the budget. It never wrote down what good governance would look like.

That decision was well founded.

Ask three months later how it was founded, and somebody needs a week to reconstruct it.

The uncomfortable part is that the left column is not spin. The people saying it believe it, because from where they sit that is what the company looks like. That is what makes this hard to catch from the top.

The AI leadership bottleneck is a leader approving work they can no longer judge. Seventy-eight in a hundred senior leaders are not fully confident their company could pass an independent review of how it governs AI within ninety days.1

Where this actually starts

Grant Thornton asked nearly a thousand senior leaders this spring. 78% lack full confidence they would pass an independent review of their AI governance within ninety days.1
Ask a CIO whether the workforce is ready and 39 in a hundred say yes. Ask a COO in the same kind of company and it is 7.1
Only 12% say their people are truly ready to use AI. Four in five call them “fairly” or “mostly” ready, which is a polite way of saying not yet.1
Leaders feel safer suggesting new ways of working than their own employees do: 81% against 67%, in the same survey of 20,000 people.2
The fix is four questions in one meeting you already have. They are further down and they take twenty minutes.

You most likely recognise one of these

You approved something last month and you could not now reconstruct why it was right
Your team is visibly faster and the queue in front of you has got longer
Somebody asked who is accountable for an AI-supported decision and the room went quiet
Your CIO and your COO gave you two different answers about the same readiness question

What you take away

Four questions for one leadership meeting, and the one that actually stings
The current research on what leaders can and cannot show, with the sample sizes
Why your CIO and your COO both believe their own answer
What to change first when the honest answer is that nobody could explain it

What it is

Approval used to mean you had checked it.

A leader signs things off. That has always meant a mix of two things: reading the work, and trusting the person who made it. The second part carries most of the weight and always has. What changed is that the first part quietly stopped being possible on a growing share of the work.

Grant Thornton calls the result the AI proof gap. Companies are spending heavily on AI and cannot show how the decisions get made or who answers for them. They asked nearly a thousand senior leaders across US industries this spring, and 78% were not fully confident they could pass an independent review of that within ninety days.1

Worth being precise about what that number is. These are self-reported feelings, and Grant Thornton says plainly that the word audit is used loosely here rather than meaning a real audit. So read it as a confidence reading, not a compliance failure. It is still the most direct answer anyone has published to the question of whether leaders can show their work.

AI deployment is simply outpacing the infrastructure that supports it. We see this pattern repeatedly with new technology: guardrails come after an incident occurs, not before. Tom Puthiyamadam, managing partner of Advisory Services at Grant Thornton Advisors1

I would add one thing to that. The guardrails arrive after the incident, and so does the honesty. Until something goes wrong, everybody in the room has a reason to assume somebody else understood the analysis.

Composite. Built from more than one conversation.

The numbers

Thirty-nine and seven, in the same company.

Grant Thornton asked whether the workforce is fully ready to adopt AI. The answer depends entirely on whose desk you ask at.1

CIOs and CTOs who say their workforce is fully ready

39%

COOs asked the same question

7%

Everyone in the survey, taken together

12%

Same survey, same question, different chair. The people closest to the technology are five times more optimistic than the people closest to the operation.1

Neither of them is lying. The CIO sees a company where the tools work and people are trying things. The COO sees the handovers, the exceptions and the Friday afternoons. Both readings are correct from where they sit, and a leadership team that has never put those two answers next to each other will keep making plans for the CIO's company.

46%

say AI underperforms because controls and compliance are not working. Only 11% think that is where the focus should be

1

3 in 4

boards approved a major AI investment. About half of those wrote down what they expect of AI governance

1

1 in 5

has tested a response plan for when AI gets something badly wrong, while nearly three in four are already running or scaling autonomous AI

1

And leaders feel safer than their own people

Microsoft asked 20,000 people who use AI at work, across ten countries including the Netherlands, and split the answers between leaders and everybody else.2 Three questions, three gaps, all in the same direction.

Leaders

It feels safe to suggest a new way of working with AIAsked of leaders, then of employees at the same companies.

81%

Employees

Same question, further downFourteen points lower, and this is the friendliest of the three.

67%

Leaders

Reinventing how the work is done gets rewarded, even without resultsThe question that says what a company actually pays for.

21%

Employees

Same question, further downHalf. And 21 was not a comfortable number to begin with.

10%

Microsoft's 2026 survey of 20,000 AI users in ten markets, fielded between February and April. Bars are the share answering yes.2

That last pair is the one I would take into a meeting. Ten people in a hundred believe this company rewards them for changing how the work gets done. Their leaders think it is twice that. Every plan built on the leader's number is a plan for a company that behaves better than yours does.

How it starts

It is not a bad leader. It is a queue nobody redesigned.

The work got harder to read and the calendar did not change

A proposal that used to take a person two days now arrives in an afternoon, with more in it. The slot in which you look at it is the same forty minutes it was in 2023. You are reading twice the material in the same time, so you read for shape instead of substance, and shape is exactly what these tools are good at producing.

Nobody hands their boss something they cannot defend. They hand you something that looks like they could. Paul Musters

Saying you do not follow it costs you something

This is the one nobody writes in a survey. Admitting in a leadership meeting that you cannot judge the analysis in front of you feels like admitting you are behind. So you ask a question about timing instead, everybody nods, and the thing gets approved. I have watched that exact move in rooms where three people were doing it at once.

Governance is somebody else's job until it is nobody's

Three in four boards approved a major AI investment. Around half of them wrote down what they expect of the governance around it.1 The money has an owner and the accountability does not, which works fine right up to the first difficult phone call from a customer.

Settle the company-level question too

Anonymous, and it takes three minutes. Your level, what it is costing you, and what changes one step up.

Do the Culture Level scan

One thing to try

Four questions. One meeting you already have.

Take one decision the leadership team made in the last month with heavy AI input. A pricing call, a market read, a hiring plan. Ask these four out loud, with no notes on the table.

1

Who made this, and with what? Not which team. Which person, and which tool. If nobody in the room knows precisely, that answers the next three as well.

2

Which assumption in here did not come from our own numbers? There is always one. Somebody should be able to name it in a sentence. If the room goes quiet, you have found where the risk sits.

3

Who read the reasoning, and what were they looking at? Ask that person directly rather than asking the room. The answer is usually narrower than everybody assumed, and usually about the numbers rather than the logic.

4

If a customer asks about this in three months, who answers? One name. Not a team, not a function. This is the question that turns the other three into something you can act on.

Twenty minutes, and you will learn more about your own governance than a policy document will tell you in a year. Do it on a decision that went well, not one that went wrong, or the meeting becomes about blame and everybody stops answering honestly.

One warning from running this with clients. The first time, somebody will get defensive, and it will usually be the person who made the analysis. Say up front that you are testing the process rather than the person, and mean it, because you will need that person to answer honestly the second time as well.

Level by level

Your CIO and your COO are both describing themselves.

The reason those two answers were five times apart is that neither of them was measuring. They were each reporting what their own part of the company feels like, which is honest and useless for planning.

An Operating Profile in use. Personality type and AI level in one profile, with the agents that fit it.

Two things per person, filled in by the person themselves: how they think and work, and how far along they are with AI. Fifteen minutes each, filled in by the person themselves. What comes out is a picture per team rather than an opinion per executive, and the first thing most leadership teams learn from it is which of them was closer to right.

It also tells you something about yourself, which is the part leaders skip. If you cannot judge the output, that is a thing you can change in a quarter, and knowing where you stand is the start of it.

About emaho

emaho measures one Operating Profile per person: personality type and AI level in a single profile. On that we build a personal set of AI agents that fit how that person works, inside the tools they already use. Fifteen minutes to complete, first profile free, built for companies between 20 and 500 people.

See your own profile, free Find your AI Culture Level

Fifteen minutes per person. No credit card, no strings.

The answer

Two things this week. Two this quarter.

Split it that way deliberately. The first two cost you almost nothing and tell you where you stand. The other two change something, and they need the first two to have happened.

This week

Run the four questions on one recent decision, in a meeting that already exists. Twenty minutes. You are not hunting for a problem, you are finding out how far the honest answers get before they run out.

Then ask your CIO and your COO the same readiness question separately, and put the two answers on one slide. That slide is worth more than the next three strategy sessions, because it shows the leadership team that they have been planning for two different companies.

This quarter

Give one decision type an owner. Pricing, hiring, or whatever your riskiest AI-supported call is. One person who can be asked afterwards how it was made and who has the standing to send something back. Start with one type rather than all of them, because a rule that covers everything gets applied to nothing.

And close your own gap, out loud. Pick the kind of analysis you approve most often and spend two hours learning where these tools tend to be confidently wrong on it. Say in the meeting that this is what you are doing. The interesting thing is what happens next, which is that two or three other people admit they were planning the same.

What did question four turn up?

Send me the answer you got when you asked who would take the customer call. I will tell you what I usually see at that point and what I would change first. I answer these myself, and usually within a day.

Message me on WhatsApp

The price of leaving it

The queue gets longer and your best people stop joining it.

The first cost is speed, which is the opposite of what everyone expects. A leader who cannot judge the work slows down in front of it. The team ships faster into a queue that moves at the same pace it always did, and the frustration lands on both sides of the desk.

Then there is the decision you cannot defend. Nearly three in four companies are already running or scaling AI that acts on its own, and one in five has tested what they would do when it gets something badly wrong.1 The gap between those two numbers is where the expensive Tuesday lives.

And the one that compounds quietly. People notice when approval has become a formality, and they adjust: they stop bringing you the hard calls and start bringing you finished things. You will feel that as your team getting more autonomous, which is a pleasant way to lose sight of the work.

Questions people ask

Questions about the approval queue

What is the AI proof gap?

The AI proof gap is the distance between what a company spends on AI and what it can show about how the decisions get made. Grant Thornton named it in its 2026 AI Impact Survey of nearly a thousand senior leaders in the United States, where 78% lacked full confidence they could pass an independent review of their AI governance within ninety days. Grant Thornton notes these are perceptions rather than an actual audit result.

How do leaders stay in control of decisions made with AI?

By being able to reconstruct one. Take a recent decision and ask who made the analysis and with what, which assumption did not come from your own numbers, who read the reasoning and what they looked at, and who answers if a customer asks in three months. Four questions, twenty minutes, and the fourth one is where most leadership teams find the gap.

Why do leaders become the bottleneck when AI speeds up the team?

Because the work got harder to read while the calendar stayed the same. A proposal that took two days now arrives in an afternoon with more in it, and the slot for reviewing it is still forty minutes. So the reading shifts from substance to shape, and shape is exactly what these tools produce well.

How many companies could pass an AI governance review?

Few are confident they could. In Grant Thornton's 2026 survey of nearly a thousand US senior leaders, 78% lacked full confidence their organisation would pass an independent review of its AI governance within ninety days. Only 12% said their workforce is truly ready to adopt AI, and 81% described it as only fairly or mostly ready.

Why do a CIO and a COO give different answers about AI readiness?

Because they are each describing their own part of the company honestly. In Grant Thornton's 2026 survey, 39% of CIOs and CTOs said their workforce is fully ready to adopt AI, against 7% of COOs. The technology leaders see tools that work and people experimenting. The operations leaders see the handovers, the exceptions and the deadlines.

Who is accountable for a decision made with AI input?

The person who signed it, which is why the question is worth settling before it is tested. In practice the workable rule is one named person per decision type rather than a policy covering everything: somebody who can be asked afterwards how the call was made, and who has the standing to send an analysis back.

How do you know whether you can still judge the work your team hands you?

By measuring where you sit, not where the company sits. emaho builds one Operating Profile per person, personality type and AI level in a single profile, and that includes you. It tells you which kinds of output you can still evaluate confidently and where you are approving on trust, which is a normal place to be and a fixable one.

What should a board expect from AI governance?

Something written down. Grant Thornton found that three in four boards approved a major AI investment while 52% had set clear governance expectations and 54% had built AI risk into ongoing oversight. The money has an owner well before the accountability does, and the gap between those two is where the difficult phone call lives.

Is it a problem if leaders cannot judge AI output themselves?

It is normal and it is fixable, and it stops being fine when it goes unnamed. A leader who cannot evaluate an analysis will approve it on trust, which works until it does not. The practical move is to pick the analysis type you approve most often and spend two hours learning where these tools are confidently wrong on it.

What is the difference between an AI policy and AI governance?

A policy says what people may do. Governance is who decides, who checks, who answers and what happens when something goes wrong. Grant Thornton found 46% of leaders saying AI underperforms because controls and compliance are not working, while only 11% think risk and compliance is where the focus should be. Most companies have the first and not the second.

How do you prepare for an AI incident before it happens?

Write down what you would do, then test it once. Nearly three in four organisations in Grant Thornton's 2026 survey are piloting, scaling or running autonomous AI, while one in five has tested a response plan for AI failures. Most already have incident playbooks; they simply have not been adapted for work a machine produced.

Do leaders and employees experience AI differently?

Consistently, and always in the same direction. Microsoft's 2026 Work Trend Index surveyed 20,000 AI users across ten markets and found leaders more likely than employees to say it feels safe to suggest new ways of working with AI (81% against 67%) and that reinventing work is rewarded regardless of outcome (21% against 10%).

Does better AI governance actually improve results?

The two travel together. In Grant Thornton's 2026 survey, companies with fully integrated AI were nearly four times more likely to report AI-driven revenue growth than those still piloting, 58% against 15%. That is a correlation rather than proof of cause, and the direction is consistent with what shows up in other 2026 surveys.

Where this goes next

Where this leads next

Each one is a different part of the same question: who can still tell whether the work is right.

18

Oversight erodes

The same thing one level down. A firm whose whole product is checking shipped a report with invented sources, and every review layer worked exactly as designed.

Read this one 06

The usage gap

Why your CIO and your COO gave different answers. What leadership thinks daily AI use looks like, and what it actually is.

Read this one 01

Nobody owns the agent

Question four, applied to the things that act without a person. Which named person answers when an agent sends something out under your name.

Read this one

All 25 AI challenges, and what to do about each

Twenty-five things that break inside a company once people start using AI, each with the research behind it and the level where it starts to bite. This one starts to bite at Campfire, and 2 of the others start there too.

Open the full set

Your next step

Which of your two answers was right?

Your CIO said 39, your COO said 7. Neither of them was measuring. Fifteen minutes per person gives you the actual picture, per team, and it starts with your own.

Ask both of your answers to show their workingNeither 39 nor 7 came from a measurement. Better to find that out in the meeting than after it.
Fifteen minutes per person, then one pictureYou get the actual spread per team instead of two confident guesses that cancel each other out.
Start at the topIf your leadership team cannot judge AI output themselves, everything below them is being approved by people who cannot check it either.

Start with your own profile, free Ask me anything first Plan a 30-minute call

First profile free · no credit card · built for companies of 20 to 500 · you decide what your team gets to see

Not ready to put your team in anything yet? Start with the level of the company instead. The Culture Level scan is six questions, three minutes, and asks nothing of you.

Paul MustersFounder of emaho

Paul Musters

Fifteen years of leadership and team development in Dutch scale-ups. That practice now sits in software: one Operating Profile per person, with agents that actually fit. He writes these pages from what he runs into with clients, not from a research summary.

LinkedIn · paul@emaho.world · WhatsApp

Sources and numbers used on this page

Grant Thornton, 2026 AI Impact Survey, press release of 13 April 2026. Self-reported responses from nearly 1,000 senior business leaders across multiple industries in the United States, collected in early 2026. Source of: 78% lack full confidence they could pass an independent AI governance audit within 90 days; 12% say their workforce is truly ready to adopt AI and 81% describe it as only “fairly” or “mostly” ready; CIOs and CTOs are five times more likely than COOs to say the workforce is fully ready, 39% against 7%; 46% say AI underperforms because controls and compliance are not working while only 11% think risk and compliance should be the main focus; three in four boards approved a major AI investment while 52% set clear AI governance expectations; 51% say strategy is the biggest driver of AI return while 22% of operations leaders have a fully developed AI strategy; nearly three in four are piloting, scaling or running autonomous AI while one in five has tested a response plan for AI failures; companies with fully integrated AI report AI-driven revenue growth at 58% against 15% for those still piloting. Quotes from Tom Puthiyamadam and Sumeet Mahajan come from the same release. Grant Thornton states that these are respondents' perceptions and that references to an audit are conceptual rather than an actual audit or a determination of regulatory compliance.
Microsoft, 2026 Work Trend Index Annual Report, published 5 May 2026. Survey by Edelman Data x Intelligence among 20,000 knowledge workers who use AI at work, across ten markets including the Netherlands, 2,000 per market, fielded 18 February to 7 April 2026. Source of: leaders are more likely than employees to say it feels safe to suggest new ways of working with AI (81% against 67%), that their manager creates space for AI experimentation (78% against 59%), and that reinventing work with AI is rewarded regardless of outcome (21% against 10%). All survey items are self-reported, and Microsoft notes the relationships shown are statistical associations rather than causal effects.

Numbers are quoted as published. The four lines in the two-column recognition block are composites drawn from client situations rather than transcripts. Nothing on this page is legal advice.