AI adoption challenges

What goes wrong when companies start working with AI.

Twenty-five things that break inside a company once people start using AI, each one sitting at the level where it starts to bite. Pick your level and see which are yours.

25challenges · one map

Reading as

Which of these sounds like your company?

One choice, no questions. The challenges that start biting there will light up.

Swipe the map sideways

Ask yourself

How to read the map. The line is the hype cycle and the bands underneath are the five AI culture levels, so each challenge sits where it starts to bite.

The full set

All twenty-five, at the level where they start

Read the level you are on and skip the rest. The ones above you are not your problem yet, and knowing they are coming is most of what you need from them today.

Level 1

Campfire

AI is around and the work has not changed. The problems here come from outside: what the market does while you wait.

1

The winner-take-more effect

Everyone in your market bought the same tools. Six companies in a hundred turned them into a lead that widens every quarter, and it is not the tools doing it.

The companies pulling ahead did not hire better people. They changed how the work is organised around them, which makes this a people decision before it is a technology one.

Ask yourselfWhich competitor had a better year than we did, and do we know what they changed?

Ask yourselfIf a competitor were pulling ahead on this, would we notice it from the inside?

See what the leaders did differently

2

The entry-level squeeze

Nobody gets fired. The junior role simply stops being refilled, and employment for 22 to 25 year olds in AI-exposed work already runs 19% behind.

The junior you are not hiring this year is the senior you cannot hire in 2030. Every bench you have ever built started with somebody being allowed to be slow.

Ask yourselfWho do we promote in 2030, and are they in the building yet?

Ask yourselfWho here is genuinely learning the job right now, and who is only producing output?

See what happens to your bench

3

The AI leadership bottleneck

Work now arrives faster than the people approving it can judge it. The queue forms at the top, and the decisions coming out of it get thinner.

Your best people are waiting on a manager who can no longer assess what they made. From where they sit, that looks like distrust rather than overload.

Ask yourselfWhen did I last approve something I could not have made myself?

Ask yourselfWhich of our managers can still judge the work their team now produces with AI?

See how to clear the approval queue

Level 2

Wild West

Everyone does their own thing and nobody can see it. This is the crowded level, and most of what goes wrong here is survivable.

4

AI culture drift

The unwritten rules about AI changed without anyone deciding. Eight in ten people now suspect a colleague of using it to look busier than they are.

Your team already knows what is normal here. Nobody ever wrote it down, so everyone guessed, and the guesses do not match each other.

Ask yourselfWhat is normal here now that would have raised eyebrows a year ago?

Ask yourselfIf I asked three people what is allowed with AI, would I get one answer?

See the four questions to answer in public

5

Agents without rules

Something in your company is making decisions under nobody's authority, because nobody ever wrote down what it may settle on its own.

Every person here has a job description that says what they may decide. The agent doing comparable work has none, and nobody notices until it decides something large.

Ask yourselfWhat can our agents decide right now without asking anyone?

Ask yourselfWho answers for a call an agent made that a person would have escalated?

See what to write down first

6

AI sycophancy

The model agrees with whoever is asking. Eleven leading models backed the person 49% more often than humans did, including plans that were clearly harmful.

The tool your people trust most is the one least likely to tell them they are wrong, which quietly removes the friction that used to make the work good.

Ask yourselfWhen did a model last tell me my plan was wrong?

Ask yourselfWhere did a second opinion get replaced by something that always agrees?

See how to make it disagree with you

7

The middle manager squeeze

Every AI change lands on the same layer, and that layer is already thinner than it was. Manager engagement fell nine points in three years.

Wider spans, fewer layers, and now the job of making AI work inside the team. All of it on the same shoulders, with nothing taken off in return.

Ask yourselfWhat came off my managers’ plate when AI arrived?

Ask yourselfWhich manager here is holding our AI adoption together on their own time?

See what to take off their plate

8

The usage gap

You are planning for a company that stopped existing. Around three quarters of frontline employees now use AI regularly, and most leaders guess a fraction of that.

The adoption number in your report and the number of people using AI every day are not the same number, and every policy you write sits on the wrong one.

Ask yourselfSay our adoption number out loud, then go and check it.

Ask yourselfDo we know who uses AI daily, or only who has a licence?

See how far off the real number is

9

The human cost of AI

One in a hundred laid-off workers is let go because of AI. Almost everyone still working has quietly worked out whether they are next.

Nobody told them what this means for their own job, so they filled it in themselves. What people invent in that silence is always worse than the truth.

Ask yourselfHas anyone here been told plainly what AI means for their job?

Ask yourselfWhat are people telling each other about this that they will never tell us?

See how to have the conversation

10

Shadow AI

Real work is running through personal accounts you cannot see. Unapproved use went from 15% to 45% of employees in a single year.

The people hiding it are usually your best ones. They get more done that way and they do not want to be told to stop, so they say nothing.

Ask yourselfWhere does the real work happen, and can we see any of it?

Ask yourselfWhat would somebody have to believe to tell us honestly which tools they use?

See how to get it back in the open

11

Workslop

Work arrives looking finished and holding nothing. Forty in a hundred desk workers were sent one last month, and each takes about two hours to repair.

The thinking one person skipped lands on whoever opens the file, and that person never gets credit for the rescue. They just get slower.

Ask yourselfTwo hours per rescue, times the people here. What is that a month?

Ask yourselfWho on this team keeps quietly fixing other people’s work?

See what it costs you per person

12

The context gap

The advice is sensible and it would fit any company in your sector, which means your competitor is getting the same answer from the same machine.

Good advice for a company that is not yours, applied to your team by somebody who assumed it fitted because it sounded right.

Ask yourselfWould this advice be any different for the company across the street?

Ask yourselfWhat does a model need to know about us before its advice is worth following?

See what makes advice fit here

Level 3

Blueprint

The rules are written down and the work still moves the way it always did. These are the problems that survive a policy.

13

The missing translation layer

You removed the layer that used to tell you the request itself was wrong. 44 in a hundred employees say management layers were cut, and that job went with them.

The people who pushed back on a bad brief are gone. Briefs now go straight to build and come back exactly as asked, which is the whole problem.

Ask yourselfWho used to tell us the request itself was wrong, and who does that now?

Ask yourselfWhere does a bad instruction get stopped in this company today?

See who has to ask the question now

14

Nobody owns the agent

Something is doing the work of a person with no name under it, so when it gets one wrong there is nobody to ask and nothing to correct.

You would never let a person start with no manager, no title and no review. An agent did exactly that, and it is already sending things to clients.

Ask yourselfIf the agent gets it wrong tomorrow, whose name is on it?

Ask yourselfWho does this agent report to, and who notices when it stops?

See how to put a name on it

15

Automation bias

Nobody decided to stop checking. The review is still in the calendar and it stopped catching anything some months ago.

The person doing the review is not lazy. Being right ninety-nine times is what teaches you to stop looking on the hundredth.

Ask yourselfWhen did we last catch something the machine got wrong?

Ask yourselfDoes anyone here still get the time to check, or only to sign off?

See how to make the review real again

16

AI answer convergence

Three companies in your market ask the same model the same question and get the same fashionable answer. Only one of you can win with it.

Everyone prepares the same way now, so every meeting starts in the same place and the disagreement that used to make it worth holding is missing.

Ask yourselfWhere does our answer differ from our competitor’s, and can we say why?

Ask yourselfWho here still turns up with something the model would never have said?

See where to keep your own answer

17

The AI productivity paradox

Three years in, nine in ten executives report no measurable productivity gain. Everyone is faster and nothing arrives sooner.

Your people are working harder and cannot name one thing that got easier. That is where belief in the whole programme quietly goes.

Ask yourselfEveryone is faster. What actually arrives sooner?

Ask yourselfWhat did AI take off someone’s plate here, honestly?

See where the gains disappear

18

Organisational AI context

It has read almost everything ever published and it has never met your company, so what you get back is the average answer for your industry.

Everything you know about how this place actually works lives in people’s heads, where no model and no new joiner can reach it.

Ask yourselfWhat do we know that a model could never look up?

Ask yourselfHow long before somebody new here knows the things nobody writes down?

See the ten things to write down

19

Coordination cost

Teams with heavy AI use finish 21% more work while the time it waits for review nearly doubles. The gain lands in a queue instead of in the market.

Your people are faster and the work still waits on somebody, which is more demoralising than being slow was.

Ask yourselfHow long does finished work sit before anyone touches it?

Ask yourselfWho became the bottleneck here, and do they know we made them one?

See where the waiting happens

20

The AI reality gap

59 in a hundred executives say they communicate a clear AI vision. Eight in a hundred employees agree that they heard one.

The plan you communicated and the plan they heard are two different plans, and they are acting on the second one every day.

Ask yourselfAsk three people to describe our AI plan, separately, and compare the answers.

Ask yourselfWhat do people here think our AI plan means for them personally?

See what your people actually heard

Level 4

Engine

Ownership and measurement are arranged. What is left is harder to see and more expensive to get wrong.

21

The measurement blindspot

Eight in ten people say AI made them faster and fewer than four in ten companies can show it in the numbers. You cannot keep funding what you cannot prove.

You can feel the change in the team and you cannot put it in a report, so the budget goes to the department that can.

Ask yourselfIf the board asked for proof tomorrow, what would we put on the slide?

Ask yourselfWhat are we counting that tells us nothing, and what should we count instead?

See what to measure instead of logins

22

The outdated org chart

Nobody left and the work moved anyway. The chart still describes a company that stopped existing about a year ago.

People are doing jobs that are not their job. It shows up as vague performance reviews now and as burnout later.

Ask yourselfPoint at a piece of work, ask who owns it, then check the chart.

Ask yourselfWhose real job no longer matches the title we review them against?

See how to map where the work went

23

Judgment moving to AI

60 in a hundred executives already use AI to help them decide, and 5 in a hundred believe they are leading on how. The rest are improvising in private.

Your managers supervise decisions nobody trained them to supervise, and they are not going to be the first to say so out loud.

Ask yourselfWhich decisions here are still ours to make, and where is that written down?

Ask yourselfWhat would a manager need to know to overrule a model with confidence?

See which decisions to write down

Level 5

Ecosystem

People and agents run as one system. The shortest list, and the one nobody has a playbook for yet.

24

AI skill atrophy

Sixteen pilots flew a simulator with the automation switched off. Their hands were fine and their picture of the situation was gone. Your reviewers are in that seat.

The people who could still do the work by hand are the ones who have not done it in a year, and they are exactly who you rely on to catch a mistake.

Ask yourselfWhich piece of our work could nobody here do by hand today?

Ask yourselfWho here still knows why we do it this way, and not only how?

See which skill you would miss first

25

The agent inventory

You can name every person who works here. Try naming the agents, then try naming who owns each one and what it is allowed to do.

Every person here has a manager and a review date. The agents doing real work have neither, and at least one of them belongs to somebody who left.

Ask yourselfWho could give me a complete list of our agents by Friday?

Ask yourselfWhat happens to the agents somebody built once that somebody leaves?

See where the ones you forgot are hiding

How this set was made

Where each of these came from

Published research first

Every figure on a challenge page names its source, its sample and its method. Where the only source for a number is a company selling the remedy, we say so and leave the number off the page.

Then what shows up in the room

A challenge earns a place when it turns up in more than one company and somebody recognises it before it is explained. That is a lower bar than research and a higher bar than an anecdote, and we mark which is which.

Placed at a level, not ranked

The level is the point at which a challenge starts to cost something. It is a judgement, based on where we have seen each one appear, and it is the part of this page most likely to move as the set grows.

Last checked against sources on 8 September 2026. Two challenges were added in September 2026: skill atrophy and the agent inventory, both at level five.

What to do with this

Three ways to use the set, in rising order of effort

1

Read your level

Pick the level above that makes you wince and read the three or four challenges that sit there. Fifteen minutes, no forms, nothing to install.

Go to the full set
2

Ask the questions out loud

Every challenge ends in a question. Take three of them into your next leadership meeting and see which one makes the room go quiet. That is your first one.

Find the questions
3

Talk it through

If one of these is the one keeping you awake, half an hour on it is worth more than a proposal. That conversation costs nothing.

Book a 30-minute call

Questions

What people ask about this set

What are the AI adoption challenges?

They are the twenty-five things that reliably break inside a company once people start working with AI. Not the technology failing, the company failing to absorb it. Each one is written up with what it looks like from the inside, what the research says, and what to do in the first week.

Why twenty-five and not more?

Because these are the ones we could source. Every challenge on the list has either published research behind it or a pattern we have seen in more than one company. Plenty of other things go wrong with AI. They are not here because we could not stand behind them yet.

What are the AI culture levels?

Five stages a company passes through as AI moves from personal habit to how the work is organised: Campfire, Wild West, Blueprint, Engine and Ecosystem. They describe how a company works with AI, not how much software it has bought.

Why is each challenge tied to a level?

Because the same problem is not equally urgent everywhere. Agents without rules cannot hurt a company that has no agents. Putting each challenge at the level where it starts to bite turns a list of twenty-five worries into the three or four that are actually yours this quarter.

Which level are we on?

Pick the description that makes you wince. The chooser on this page gives you five of them with three signs each, and the map then shows which challenges start biting at that level. It takes about a minute and nobody has to fill in a form.

Can we be on two levels at once?

Usually, yes. A technical team at Blueprint and a commercial team at Campfire is the most common split we see. Take the level of the part of the company where the work actually happens, because that is where the cost lands.

Is this list for a CEO or for HR?

Both, and they read it differently, which is why the map has a switch. A CEO tends to see the competitive and financial edge of a challenge. A Head of People tends to see who ends up carrying it. Same challenge, two entirely different sentences.

What does the shape of the map mean?

The line is the Gartner hype cycle, drawn from left to right: expectations climb, they fall, and they settle. The vertical bands are the AI culture levels. Each challenge sits on the line at the level where it starts to bite, so the crowded part of the curve is the crowded part of the journey.

Where do the numbers on this page come from?

Every figure on a challenge page names its source, its method and, where it matters, who paid for the research. Where a number only exists in vendor marketing, we say so and leave the number off. That happens more often than you would like.

How is this different from an AI readiness assessment?

A readiness assessment scores you against a model of what good looks like. This starts from what is already going wrong. You do not need a benchmark to recognise a manager who has stopped being able to judge the work, and the recognition is what makes anything happen.

What should we do first?

Read the three or four challenges at your level, and pick the one where the question at the bottom made somebody in the room uncomfortable. Discomfort is a better prioritisation signal than severity, because it means it is real and it means it is yours.

Do we need to fix all of these?

No, and trying is how the effort dies. Most companies have three that matter and twenty-two that do not, at least not yet. The value of the full set is knowing which ones are still ahead of you, so they are not a surprise when they arrive.

How often does the list change?

It grows when something new shows up in enough companies to be a pattern rather than an anecdote. Two of the twenty-five arrived in September 2026. Every page carries the date it was last checked against its sources.

Is any of this specific to the Netherlands?

The research is international and the patterns hold everywhere we have looked. What is local is the pace: Dutch scale-ups tend to reach the Wild West level fast and stay there longer than they intend to, because nobody wants to be the one who slows everybody down.

Who is behind this?

Paul Musters. He works with founders, CEOs and People leaders at scale-ups and innovative SMEs on how their companies work with AI. The set is written from that work and from published research, and every page says which is which.

What does it cost to talk to you about this?

Nothing for the first conversation. If a challenge on this list is the one keeping you awake, the useful next step is half an hour on it, not a proposal.

Who wrote this

Written from the work, not from a survey

Paul Musters

Paul Musters works with founders, CEOs and People leaders at scale-ups and innovative SMEs on how their companies work with AI. Every page in this set is written from that work and from published research, and says on the page which part is which.

This page is available in Dutch as well: de 25 AI-challenges.