Sixty in a hundred executives already use AI to help them decide.
Which decisions are still yours?
Judgment moving to AI is the slow, unannounced transfer of who actually decides. Nobody signed anything. A recommendation was accepted, then the next one, and by now the model proposes and the meeting confirms.
Deloitte asked more than nine thousand leaders about this. Sixty in a hundred executives use AI to support decisions, sixty-four in a hundred call the question very important to their success, and five in a hundred consider themselves to be leading on it.
Where it startsYou ask, it suggests, you decideWhere it ends upIt decides, and you review the ones it flags
Try this first
There are three settings, and most companies are on one by accident.
Every decision a machine touches sits at one of these three. The problem is almost never that a company picked the wrong one. It is that nobody picked, and the setting drifted upward one accepted recommendation at a time. Choose the one that matches how a real decision went last month.
What it buys you
Speed on the gathering, with the judgement still where it was. This is the setting almost everybody believes they are on, and it is the one that produces the least change in how a company actually decides.
What it costs
Very little, and that includes very little benefit. You are paying for a faster research assistant, which is worth having and is not what anybody bought the technology for.
On a Tuesday
You ask for the arguments on both sides, read them, and reach your own view. Twenty minutes saved, and the decision is unmistakably yours.
Right when
The decision is irreversible, contested, or one you will have to defend in a year. Amazon's language for these is one-way doors, and they deserve the slow setting.
Almost every company I ask says they are on the first one. Almost every company, when we look at a real decision from last month, was on the second, and had been for a while.
You have seen this
Which of these has already happened here?
Tick honestly. Nobody sees this but you, and every one of these is a normal thing that happened for a good reason.
Two or more is normal, and none of them is a failure. What they have in common is that the setting moved and nobody noticed it moving, which is the only part of this that is actually a problem.
Judgment moving to AI is the setting drifting from advising to deciding without anybody choosing.
Sixty in a hundred executives now use AI to support decisions. Five in a hundred consider themselves to be leading on how to do it well.1
If you read one part
Deloitte surveyed more than 9,000 business and HR leaders in 89 countries. 60% of executives now regularly use AI to support their decisions.1
64% call this very important to their current success and a similar number are taking steps. 5% consider themselves to be leading the way.1
Deloitte lists two findings that should worry a leadership team: people feel less ownership over decisions made with AI, and become more likely to be dishonest when delegating a decision to it.1
More than half of organisations, 57%, operate at low decision-making maturity, with few teaching decision skills at all.1
Outside work the habit is already forming: 16% of people across 23 markets have used AI that acts without human intervention, in the last six months alone.2
You know the feeling if
A meeting confirmed something rather than deciding it
Somebody justified an outcome with an output
You cannot name who answers for a decision made last month
You want to use more of this and would rather choose the boundary than discover it
What you take away
Three settings, with the honest price of each
A one-hour exercise that sorts your decisions into two piles
The research, with the samples and what it does not prove
The one document worth writing before you need it
What it is
The setting moves upward on its own, because moving it is never a decision.
Here is the mechanism, and it is almost polite. You ask for material and get a recommendation, because that is more helpful. The recommendation is good, so you take it. Next time you skip the material and read the recommendation. Six months later a meeting exists to confirm recommendations, and nobody in it can point to the week that changed.
What makes this different from ordinary delegation is that delegation has a moment. You tell somebody they own this now, they know it, and both of you can say when it happened. Here there is no moment, which means there is no point at which anybody asked whether this particular decision was one to hand over.
Nobody hands over a decision. They accept a recommendation, twelve times, and the twelfth one was the handover.
Paul Musters
Deloitte's research names two consequences that are worth reading twice. People feel less ownership over decisions made with AI, and they become more likely to be dishonest when they delegate a decision to it.1 Put those together and you have a decision nobody quite made, that nobody quite owns, arrived at by a process people are slightly less straight about than they would be on their own.
And where a company sits on this is a level question rather than a technology question, which is why two companies running the same tools end up in very different places.
01Campfire60%
02Wild West25%
03Blueprint10%
04Engine4%Decision rights written
05Ecosystem1%
Share of companies per level. Writing down which decisions a machine may make on its own, and who overrides it, is Level 4, Engine, where about 4% of organisations sit. Below that the boundary is wherever the last person happened to draw it. The shares per level come from emaho's own work with clients.
Below Level 4 the boundary is wherever the last person happened to draw it, which in practice means wherever the last helpful recommendation landed.
What the studies found
Widely used, widely important, almost nowhere managed.
Deloitte's 2026 Global Human Capital Trends research ran with Oxford Economics across more than 9,000 business and human resources leaders in 89 countries.1 Three of its numbers describe the state of this cleanly.
60%
of executives now regularly use AI to support their decisions
That third number is the shape of every new discipline before anybody builds it. What makes this one urgent is the first number, because the use is already at sixty per cent and the practice is at five.
Deloitte adds a finding from its own decision intelligence research that explains a lot of it: more than half of organisations, fifty-seven in a hundred, operate at low decision-making maturity, with few teaching decision skills or providing tools to support decisions.1 There is an irony in there worth sitting with. Many organisations are busy teaching AI how to decide while assuming the humans already know how.
The habit is already forming outside work
EY's 2026 AI Sentiment Report surveyed 18,152 people across 23 markets about the previous six months. It is consumer research rather than workplace research, which is exactly why it is useful here: it shows what people are getting used to before they bring it to the office.2
Have used AI that acts without human intervention
16%
Let AI refill shopping carts or handle banking tasks automatically
11%
Have used AI agents to buy products on their behalf
10%
Have used a self-driving vehicle or an autonomous taxi
9%
Share of 18,152 people across 23 markets reporting each, in the six months to March 2026. Bars are shares of all respondents.2
Small numbers, and that is the point. A year ago they were rounding errors. The people arriving in your Monday meeting are getting comfortable with machines acting on their behalf in their own lives, which is how a professional norm forms without anybody proposing one.
The same survey finds two thirds saying human oversight remains essential, and the same two thirds worried about systems being hacked.2 EY's own summary of it is blunt and correct: adoption is moving faster than confidence.2
Two limits worth naming. The EY figures are about consumers rather than employees, so read them as the weather rather than as your forecast. And the Deloitte research is a self-reported survey of leaders, run by a firm that sells advice on decision governance.
Why it happens
Every handover was somebody being helpful.
A recommendation is more useful than material
Every tool in this space is built to be more helpful over time, and more helpful means closer to an answer. So the default drifts from here are the considerations towards here is what you should do, and it drifts because users prefer it. Nobody in that loop is doing anything wrong, and the loop moves in one direction only.
Accepting a good recommendation is not a decision anybody records
If you had proposed handing this class of decision to a system, there would have been a meeting, a risk conversation and probably a document. Accepting a recommendation that happens to be right takes four seconds and leaves no trace. Twelve of those add up to the same transfer with none of the scrutiny.
The version with a meeting gets governed. The version that happens twelve times on a Tuesday does not.
Paul Musters
Neither side of the handover was prepared for it
Deloitte lists managers not being prepared to act as supervisors of AI, and many executives lacking the AI literacy to contribute to oversight.1 So the people who would notice the setting moving are the people least equipped to describe what they are looking at, and the ones who could describe it are two levels away from the Tuesday where it happens.
And at company level?
Anonymous, three minutes, six questions. The level, what that is costing you, and what moves one step up.
This takes an hour with your leadership team and it uses a distinction Amazon has been using for years, which Deloitte cites as a working example: one-way doors and two-way doors.1 Write down the twelve or fifteen decisions your company actually made last month, then sort them.
1
Two-way doors
Reversible, cheap to undo, made often. These are the ones to hand over deliberately and quickly, and most companies are far too slow with them. Speed here costs almost nothing and buys a lot.
2
One-way doors
Hard or impossible to reverse. Pricing, a market you enter, a person you let go, a customer you walk away from. These stay on the first setting, with a named person, and the slowness is the feature.
3
The ones you argue about
Pay attention to these rather than to the sorting. A decision two people place in different piles is a decision nobody has agreed the stakes of, and that disagreement is worth more than the exercise.
Then write, next to each one, at which of the three settings it was actually made last month. Most teams find at least two one-way doors sitting at the second setting, and finding that in a meeting is much cheaper than finding it in an outcome.
One thing that makes this go badly. Do not start from what you want the policy to be. Start from what actually happened in the last four weeks, because the gap between those two is the entire finding.
In the five levels
The right setting depends on the decision and on the person making it.
A policy that sets one boundary for the whole company will be wrong in both directions at once. It will hold back the people who could safely run further, and it will give the same room to somebody who reads a recommendation as a conclusion.
An Operating Profile in use. Personality type and AI level in one profile, with the agents that fit it.
Everything gets measured per person: the way somebody thinks and works, and how far along they are with AI. Those two together tell you who can be trusted with the third setting on which class of decision, which is a far more useful answer than one company-wide rule and a hope.
It also gives you a way to move somebody deliberately. Handing a person more room becomes a thing you do on purpose, with a level attached and a date, rather than something that happens to them over eleven quiet Tuesdays.
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.
Fifteen minutes per person. No credit card, no strings.
What works
Write the list before you need it.
One page, and it is the only artefact on this page worth making. It lists your recurring decisions, and next to each one it says which setting it runs at, who overrides it, and what has to happen for it to move up a setting.
Three columns and about twenty rows. Deloitte's framing of the same idea is to classify choices and pre-assign owners, data, guardrails and speed for each category, which is the enterprise version of exactly this.1 A company of a hundred and fifty people needs the page rather than the framework, and the page takes an afternoon.
What the page changes immediately
The setting stops drifting, because moving a decision up a level now requires editing a line that somebody will see. That single piece of friction is the whole mechanism. It converts a drift that happens by accident back into a choice that somebody makes on purpose.
It also gives your people permission to go faster on the two-way doors, which is where most of the value has been sitting untouched while everybody worried about the other pile. Written boundaries usually speed a company up rather than slowing it down, because the caution people apply in the absence of a boundary is applied to everything equally.
And one habit to attach to it
Revisit the page every quarter and ask which lines moved without being edited. That question surfaces drift in about ten minutes, and it is the only reliable way I know to find it, because drift by definition leaves no record of itself anywhere else.
One thing to stop while you are at it. Stop accepting an output as a reason. When somebody says the analysis said so, the useful next sentence is asking what they would have decided without it, and whether they can say why the analysis was right rather than that it was confident.
Which one surprised you?
If you sort last month's decisions, send me the one that turned out to be running at a higher setting than you thought. Those are the most interesting sentences I get, and I will tell you what I usually see behind them. No pitch. One honest answer.
The cost that arrives first is a one-way door that got opened at the wrong setting. A price change, a candidate, a customer, decided at the speed of a reversible call because nobody had written down that this one was different. That is survivable once, and it is the kind of thing an insurer asks about afterwards.
Underneath sits ownership. Deloitte's research finds people feeling less ownership over decisions made with AI.1 A decision nobody quite owns is a decision nobody defends in month four, when it becomes inconvenient and somebody asks why you chose it. You will hear a version of we ran the numbers, and that sentence has no author.
And the slowest one is capability. If more than half of organisations already teach almost no decision skills,1 then every year the setting drifts upward is a year in which fewer people practise deciding. The judgement you will need for the one-way doors is a muscle, and it is currently being trained by nobody.
Asked and answered
Questions about who decides now
What does it mean when judgment moves to AI?
It is the unannounced transfer of who actually decides. Nobody hands a decision over. A recommendation is accepted, then the next one, and after a few months a meeting exists to confirm recommendations rather than to reach conclusions. Unlike delegation to a person, this has no moment, so nobody ever asked whether this particular decision was one to hand over.
How many executives use AI to make decisions?
In Deloitte's 2026 Global Human Capital Trends survey of more than 9,000 business and HR leaders across 89 countries, 60% of executives said they now regularly use AI to support their decisions. In the same survey 64% called AI and decision-making very important to their current success, and 5% considered themselves to be leading the way on it.
Do people take less responsibility for decisions made with AI?
Deloitte's research lists exactly that among the distinct challenges AI brings to decision-making: people feeling less ownership over AI-made decisions, and becoming more likely to be dishonest when delegating decisions to AI. In practice that produces a decision nobody quite made and nobody quite owns, which is the one that gets abandoned in month four.
What does emaho do about judgment quietly moving to AI?
emaho makes the setting explicit at the level of the person. Everyone gets an Operating Profile, personality type and AI level in one profile, and on that we build a personal set of AI agents that fit how they work, with the limits written into the agent rather than into a policy nobody opens. The first profile is free, fifteen minutes per person, built for companies between 20 and 500 people.
What are the levels of AI involvement in a decision?
Three settings are enough for most companies. It advises, meaning it gathers material and you decide. It recommends, meaning it proposes an answer and you approve. It acts, meaning it decides within limits and you see the exceptions. Most companies believe they are on the first and, looking at a real decision from last month, turn out to have been on the second for a while.
What are one-way and two-way doors?
It is Amazon's distinction between decisions that are hard or impossible to reverse and decisions that are easily undone, cited by Deloitte as a working example of a decision framework. Two-way doors should be handed over deliberately and fast. One-way doors, such as pricing, entering a market or letting someone go, keep a named human and the slowness is the feature.
How do you stop AI from quietly taking over decisions?
Write one page listing your recurring decisions, which of the three settings each runs at, who overrides it, and what has to happen for it to move up a setting. The page works because moving a decision up now means editing a line somebody will see. That single piece of friction turns a drift that happens by accident back into a choice somebody makes.
Where can I read about the other AI challenges inside companies?
How mature are organisations at decision-making generally?
Not very, which is the underlying problem. Deloitte's High Impact Decision Intelligence research found that 57% of organisations operate at low decision-making maturity, with few teaching decision skills or providing tools to support them. Deloitte notes the irony directly: many organisations teach AI how to decide while assuming humans already know how.
Are people letting AI act on their behalf outside work?
Increasingly. EY's 2026 AI Sentiment Report surveyed 18,152 people across 23 markets and found that 84% had used AI in the previous six months, with 16% reporting use of AI systems that act without human intervention. Ten percent had used AI agents to buy products for them, 11% let AI refill carts or handle banking, and 9% had used a self-driving vehicle or autonomous taxi.
Does trust in AI keep up with how much people use it?
No, and EY says so plainly: adoption is moving faster than confidence. In its 2026 survey two thirds of respondents worried about AI systems being hacked or breached, the same proportion said human oversight remains essential, and nearly three quarters feared no longer being able to tell what is real from what is AI-generated. Those concerns are not slowing adoption.
Where this goes next
Moving the same boundary
Each one is a different way authority shifts without anybody signing anything.
One boundary for the whole company is wrong in both directions at once. Fifteen minutes per person tells you who can safely run further and who needs the recommendation to arrive with its objection attached.
Put every decision in one of the three boxesMachine decides, machine proposes, person decides. An afternoon of work, and the arguments stop.
Set the boundary per person, not per companyOne setting for everybody is wrong in both directions at once. A level per person tells you who can safely run further.
Attach a date to every moveA boundary that moves without a date moves over eleven quiet Tuesdays, and nobody ever signed off on it.
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.
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.
Deloitte, 2026 Global Human Capital Trends, chapter 3, AI and the future of human decision-making, by David Mallon, Julie Duda, Stefano Besana and Maya Bodan, published 4 March 2026. Research conducted with Oxford Economics among more than 9,000 business and human resources leaders across many industries and sectors in 89 countries, supplemented with separate worker, manager and executive surveys and more than 50 interviews. Source of: 60% of executives now regularly using AI to support their decisions; 64% of respondents considering AI and decision-making very important to their current success with a similar number taking steps, while only 5% consider themselves to be leading the way; the listed challenges including people feeling less ownership over AI-made decisions and becoming more likely to be dishonest when delegating decisions to AI, managers not being prepared to act as supervisors of AI, executives lacking sufficient AI literacy to contribute to oversight, and insurers being reluctant to cover corporate AI use; the finding from Deloitte's High Impact Decision Intelligence research that 57% of organisations operate at low decision-making maturity with few teaching decision skills; the observation that many organisations teach AI how to decide while assuming humans already know how; the recommendation to classify choices and pre-assign owners, data, guardrails and speed per category; and the reference to Amazon's one-way and two-way door model as a working example of that. Deloitte sells consulting on decision governance and organisation design. The chapter also cites a Gartner projection that half of business decisions will be augmented or automated by AI agents by 2027, which is a forecast rather than a measurement and is not used on this page.
EY, 2026 AI Sentiment Report, announced 26 March 2026 in London. Survey conducted by EY Studio+ among 18,152 people aged 18 and over across 23 markets, using random stratified sampling with quotas on age, gender and location based on local census data, with results statistically weighted so each market carries equal weight. EY notes that in markets with lower internet penetration, including India, the Chinese mainland, Brazil and Japan, results represent internet-enabled people rather than the general population. Source of: 84% of respondents having used AI in the previous six months and 16% reporting use of AI systems that act without human intervention; 9% having used self-driving vehicles or autonomous taxis, 10% having used AI agents to purchase products on their behalf and 11% allowing AI to automatically refill shopping carts or manage banking tasks; 66% worrying about AI systems being hacked or breached and 66% saying human oversight remains essential; and the statement that adoption is moving faster than confidence. This is consumer research rather than workplace research, and it is used here as context for the habits people bring to work.
Numbers are quoted as published. The recognition list and the quotations attributed to Paul are drawn from client situations rather than transcripts. The share of companies per level comes from emaho's own work with clients.
