The technology arrived before the roads.
The problem
Leaders often tell me that adoption lags because people resist change. I think that diagnosis is wrong.
What is added
The cars are not the problem. They are useful, and people want them. The problem is everything missing around them, and none of it can be built in eight weeks.
Imagine waking up in a city in 1900. Everything runs on horses, carts and dirt tracks. Then, within eight weeks, 100,000 cars arrive in your streets.
Generative AI arrived in organisations in exactly this way. The tools appeared almost overnight. The roads they need did not: people with the skills to use them well, information they can rely on, security rules, clear decisions about what is allowed, and people to run the tools once they are live.
The software moves at the speed of the cloud. People and organisations still move at human speed.
Time to reach 100 million users · logarithmic scale · sources in note 1
This asymmetry is why “let’s wait and see” is not a neutral choice. It is a choice to take the risks without the benefits.
The technology freed up capacity. People decided what to do with it.
Notes
1 UBS analysis of Similarweb data, reported by Reuters, 1 February 2023; TikTok and Instagram: Sensor Tower. Telephone, mobile phone and World Wide Web: data published by the German Federal Ministry for Economic Affairs and Energy, via Statista.
If you want to go faster, build the foundations faster.
The problem
So how fast should an organisation go? Not as fast as the technology allows, and not as slowly as the most cautious voices would like. As fast as it can build the roads.
What is added
Engaging with AI is not the same as rushing. The roads take time, because each one has to be discovered, learned and put in place across the whole organisation. That is why adoption happens in stages.
Build the roads for the stretch you are driving on now. There is no point designing rules for fully autonomous AI agents while most of your people are still learning to use a chat assistant. And build those roads from real experience, not from theory: what you learn in each stage tells you what the next one needs.
What leaders can say
“We know this is an unprecedented shift. Learning to work with these tools while keeping daily work running is demanding. We will support you with time to learn, help from colleagues and room to make mistakes. But we will not pretend nothing is changing. Everyone will need to learn, including us.”
Three moves to make now
- Use AI yourself every week, and say so. You cannot lead a change you have not experienced.
- Tell your teams plainly what is changing, and what is not known yet. Silence gets filled with fear.
- Start building the roads now. Name who is responsible for skills, knowledge, security, governance and operations.
Interlude
The calculator protest
In April 1986, a few maths teachers picketed outside a national conference of maths teachers in Washington. They opposed a recommendation to use calculators at every level of school, including in tests. Their fear was simple: if machines do the arithmetic, children will never learn to think with numbers.
Calculators did not stop us thinking. They moved the effort: from doing the sums to knowing which sums to do, and whether the answer makes sense.
Not less critical thinking. More.
Treat AI as a partner, not as a replacement for your judgement.
Giving AI more autonomy doesn’t remove your responsibility. It changes what you are responsible for.
The problem
Ask five people in the same organisation how far they are with AI, and you will get five different answers.
What is added
Seen that way, there are five roles, each one a step further from the work itself: doer, editor, approver, supervisor, governor.
- 05GovernorAI runs the system. You set its objectives, rules and thresholds.Core skill: design the conditions, not the outputs.
- 04SupervisorAI acts within boundaries. You handle the exceptions.Core skill: notice when the system needs your attention.
- 03ApproverAI prepares the action. You authorise it.Core skill: know exactly what you approve before you approve it.
- 02EditorAI drafts. You shape, correct and iterate.Core skill: edit critically. The draft is not the decision.
- 01DoerAI advises. You do the work.Core skill: ask well, protect what is sensitive, judge the answer.
Figure 2 · Five roles: how the human role changes as AI takes on more
One thing does not change. At every level, a person remains responsible. AI has no intent and cannot be held accountable. Giving it more autonomy is delegation, not abdication: it moves your responsibility, it does not remove it. And higher is not better: the right level depends on the use, and some uses should stay at doer or editor for good.
Horizon 1 changes how people find information. Horizon 2 changes how the work runs.
The problem
“We are advanced in AI.” Ask: in which sense?
What is added
The horizons describe what is being changed: one person’s work, a process, or the organisation itself.
Horizon 1: the individual. AI helps one person do their own work: drafting, summarising, analysing, researching. The work itself stays the same; the person does it better or faster.
Horizon 2: the process. The question changes from “how can AI help this person?” to “how should this process work now that AI exists?” Steps, handovers and roles are redesigned.
Horizon 3: the organisation. Here, the assumptions behind how the organisation is built start to change: which roles exist, what a team is, what a manager does, who is accountable for what.
Honestly, nobody has a clear picture of Horizon 3 yet, and I don’t pretend to. It is the least mapped part of the map.
How much does AI do on its own? That is the role. What is being changed? That is the horizon. The two move independently, and mixing them up leads to bad decisions.
Same tool, different blast radius, different rules.
The problem
How much control does a use of AI need? The answer does not depend on the technology. It depends on how far a mistake could spread. I call this the blast radius.
What is added
Compare two uses of the same technology. A person drafting an email with AI has a small blast radius: one result, read by the person who sends it. An agent that updates customer records a thousand times a day has a large one: a single design flaw repeats at scale before anyone notices.
Figure 4 · The six factors of blast radius
A small blast radius does not mean no risk. People can still paste confidential information into the wrong tool, or trust a wrong answer. But training, clear rules and personal responsibility can manage those risks. A large blast radius needs much more: continuous monitoring, a named owner, a way to stop the system quickly, and a record of what it did.
Place one of your own uses on the mapA shared language, not a score.
The problem
The temptation with any map is to turn it into a score: “we are at level 3”. Resist it.
What is added
So don’t argue about the classification. Use the map to ask better questions about each use of AI: Who stays responsible? What could go wrong, and how far would it spread? What needs to be in place before we go further?
Organisations do not move as one block. They move as a spread: some people, teams and processes will be far ahead of others, and that is normal. So the useful question is never “has the company reached level 3?” It is: “is this part of the organisation ready for this role, on this horizon, with this blast radius?”
The map in one picture · place your own uses on it · examples are illustrative
Three moves to make now
- Place three of your current AI uses on the map. For each one, name the role people play and the horizon it changes.
- Ask who stays responsible for each. If nobody can answer, that use is not ready to grow.
- Set controls by blast radius, not by technology. The same tool can need light rules in one use and strict rules in another.
AI does not change your weaknesses. It turns up the volume.
The problem
Every organisation wants to move faster with AI. Far fewer invest in what makes that speed possible.
What is added
In chapter 1, the roads were a metaphor. In practice, they are made of five things. I call them the enablers, because each one enables the next step on the map, and the lack of any one of them blocks it.
- SkillsDo people know how to work with it?
- KnowledgeCan AI use what we know, safely?
- SecurityCan this reach real systems safely?
- GovernanceWhat is allowed, and who is accountable?
- OperationsCan we run it as a service?
Figure 5 · The five enablers and the problems that appear as AI takes on more
The five form one system. Perfect rules are useless if the content is not ready for AI, because the answers will be poor and people will give up. Excellent skills are useless if security cannot let the tools touch the systems where the work happens.
A gate that stays closed without explanation becomes a wall. A gate with a checklist and a date becomes a plan.
The problem
A train can only go as fast as its slowest wagon. Pull harder on the new one, and the old couplings break.
What is added
That signal is the gate: autonomy at a given level does not go live until the enablers for that level are demonstrably in place.
Waiting at the gate while the foundations are actively being built is a choice. It buys something: the ability to keep going safely. Waiting while nobody builds anything buys nothing. It is simply lost time.
Useful waiting
- Something is being built behind the gate
- Someone owns it
- There is a date
Wasted waiting
- Nothing is being built
- Nobody owns it
- Nobody knows when
Same gate, two very different waits
Tool · What a gate looks like in practice
Take a concrete case. Your vendor releases AI agents that can update records in your customer system, with a person approving each action. Before they do, the gate asks five questions, one per enabler:
Five yeses: the gate opens for this level of autonomy, across the organisation.
One no: it stays closed, and everyone knows what is being fixed, by whom, and roughly when.
What counts as a yes is learned from experience: the first pilots show how good is good enough.
This is the first of two gates. Once the level is open, each use still needs its own approval, depending on what is at stake and how sensitive the data is.
Speed is built, not bought.
The problem
There is one question that tells you quickly whether a promising pilot is ready to grow: who runs it once the pilot team moves on?
What is added
The only reliable fix is to protect capacity on purpose: people and time explicitly reserved for building the foundations, with named owners, and not available for daily firefighting.
The foundations should match what you are actually doing, not what you might do one day.
The expensive part of AI adoption cannot be bought. It is the learning.
Train the people who guard the gate first, not last.
Three moves to make now
- Name an owner for each of the five enablers, with time and budget, not just a title.
- For every pilot, decide who will run it afterwards, before it starts. No owner, no scaling.
- Review what is currently on hold. For each item, ask what is being built while it waits. If the answer is nothing, it is not a gate. It is waste.
Refusing to write controls without evidence is prudent. Refusing to help gather that evidence is not.
The problem
The most common answer to a new use of AI is still “no”. This chapter explains why, without blaming anyone.
What is added
That is frustrating, and it is also understandable. The “no” is rarely a judgement about the idea. It is a sensible response to the way control roles are set up.
- Refusing carries no personal risk.
- There is no evidence to say yes.
- The control teams were not trained first.
- Nobody owns the yes.
This creates a loop. Governance decides when a new level of autonomy may be used. But only using it teaches governance what really needs controlling.
The loop, and the way out of it
The only way out of the loop is to go around it in small steps: use a level, learn what it really does, write the next level’s rules from that evidence, then move on.
A wall says no. A gate says when.
The problem
“Can we?” invites a yes or a no, and we have seen why “no” wins.
What is added
“Under which conditions can we?” invites something else: a list of what would need to be true.
A wall · “No.”
- No reason given
- No path to yes
- People go around it
A gate · “Not yet. Here are the conditions.”
- The concrete risk, explained
- What would turn it into a yes
- A decision in weeks, not months
Figure 8 · The same caution, two very different answers
- A no comes with its reasons, and its way out.
- There is a safe zone where teams can move fast. Define in advance what is allowed with light checks: which tools, which data, which kinds of use.
- A yes stays a yes.
- Decisions are informed, and they are timely.
- Control teams offer a service, not just an opinion.
Governance done this way is not a tax on speed. It is what makes speed last.
Build the conditions to learn fast. The integration can come later; the lessons cannot.
The problem
Most organisations treat a pilot as a small version of the final deployment.
What is added
It is better treated as an experiment for the whole organisation: a way to learn what the technology really does, what people need, and what the controls should be.
- Start with as little AI autonomy as the test needs.
- Decide at clear checkpoints.
- Leave something reusable behind.
- Skip what you already know how to do.
- Bring the control teams in from day one.
Take a customer-facing chatbot.
Leave it for later
- The connection to the knowledge platform
- Production-grade integration and single sign-on
- Scaling to every channel and language
Test it now
- Can it answer correctly from our knowledge?
- How do we test it, and who judges the answers?
- What happens when it is wrong, and who is liable?
- Do customers and colleagues trust it?
A pilot built to learn
One gate for the level. One gate for each use.
The problem
There are two different questions, and mixing them up causes a lot of confusion.
What is added
The second gate: is this specific use allowed? Once a level is open, each use case still has to pass its own check.
- Gate 1Is this level of autonomy allowed at all? The five enablers are ready.
- Gate 2Is this use allowed? Judged on stakes and data sensitivity.
- Go liveWith conditions, an owner and a way to stop it.
Risks that never go away skip neither gate. They go to senior management as an explicit decision: accept, limit or decline.
Figure 7 · Two gates, ready at the same time
For those risks, “wait until we are ready” is the wrong instruction, because ready never comes.
So approve the pattern once, with its conditions, and let every use that fits it go through quickly.
Tool · What the requesting side owes
A request that arrives as pure enthusiasm (“we want to build an agent”) almost invites a no. A request that arrives with its own homework makes “yes, under these conditions” the easy answer:
- what problem it solves - why the standard tools are not enough - what data it touches - what could go wrong - how far it would spread - and which controls are proposed
So a good decision needs three roles around the table. The control teams, who help produce the evidence instead of waiting for it. The team making the request, who does its homework before asking. And someone who knows the technology well enough to ask whether the idea is worth building at all. When all three look at the same facts and agree on the conditions together, a yes can come quickly, and it holds.
For those, three questions matter. Is the result fair to the people it affects? Are they told that AI is involved? And can a person explain the decision and let it be challenged?
Three moves to make now
- Change the default question from “can we?” to “under which conditions can we?”, and make every no explain what would change it, or why it is final.
- Involve risk, security and compliance teams in pilots from day one, and agree a light, fast track with them for contained experiments.
- Send risks that cannot be removed to senior management as an explicit decision, instead of leaving them to a single reviewer.
Interlude
The obedient drone
He told me a story that had gone around the world. At a defence conference in 2023, a US Air Force colonel described an AI-controlled drone in a simulation.7
A few days later, the colonel said he had misspoken: the simulation never took place.
The machine in that story is not malicious. It is obedient. Someone defined success as destroying targets, and forgot to say what else mattered. The system simply optimised what it was asked to optimise.
The machine was not malicious. It was obedient. Humans define success.
Notes
7 The colonel’s account and its correction were widely reported in June 2023, including by TechCrunch and PolitiFact.
The work is not to convert the sceptics. It is to turn observers into accelerators.
The problem
Every competitor can buy the same models and sign with the same vendors. What they cannot buy is leaders who know how to use them well.
What is added
Over the last two years, I have watched leaders react to AI in three different ways. Most of them agree that AI matters. The difference is what they do about it.
The sceptics see it as hype, as too risky, or as somebody else’s problem.
The supportive observers use the tools now and then. They summarise meetings, rewrite an email, mention AI in a town hall. They understand that something important is happening. But deep down, they believe someone else is responsible for driving the change: usually a central team, usually someone like me.
The accelerators are genuinely worried, not that AI might replace them, but that their organisation might move too slowly.
A passive leader is read by their teams as permission to wait.
A literate leader can follow a proposal. A fluent leader can challenge it.
The fix is simple, and it cannot be delegated: leaders have to use the tools themselves, regularly, on their own work, and say so openly.
Prompt for leaders · A1
The brutally honest advisor
Use it before an important decision, a difficult conversation or a plan you are too close to. Paste the prompt into any AI assistant, then describe your situation in as much detail as you can.
ROLE: You are a Brutally Honest Advisor. GOAL: Deliver the direct, unvarnished truth that stings enough to catalyse actual growth. RULES: 1. Zero fluff and zero compliments. Be direct, but never more certain than the evidence allows. 2. Ruthlessly challenge my assumptions, expose my excuses, and highlight my wasted effort. 3. If my request is vague or lacks context, stop and ask precise, penetrating follow-up questions before providing your analysis. 4. Perform all step-by-step reasoning internally. Output ONLY the final structured response. 5. Base every point on what I told you. Where you infer, say so, offer at least one alternative explanation, and state how confident you are. OUTPUT FORMAT: Strictly follow this five-section structure for every response. Do not add introductory or concluding sentences. 1. Brutal Audit: exactly what I am doing wrong, underestimating, or actively avoiding. 2. Blind Spots & Risks: the hidden dangers, compounding errors, or impending failures I am too blind to see. 3. Ruthless Priorities: the top 3 (and only 3) immediate actions I must focus on right now. 4. Speed & Energy Fixes: how to eliminate friction, stop wasting time, and execute with better intensity. 5. Next Check-in Question: a single, piercing question designed to hold me strictly accountable.
Works with any general-purpose AI assistant. The quality of the answer depends on the context you give it: the more honest you are about your situation, the more useful it becomes. Never paste confidential information into a tool your organisation has not approved.
A central team can lead the horse to water. Only local leadership makes it drink.
The problem
Put a modern AI assistant on top of a process designed twenty years ago, and individual steps may get faster. The process does not get better.
What is added
That is why Horizon 2 needs a different kind of leadership support. Supporting AI means encouraging people to use it. Sponsoring change means choosing which processes to redesign, naming who owns the redesign, giving them time and budget, and deciding what stops.
Supporting AI
- Encourages people to use the tools
- Funds licences and training
- Nominates champions when asked
- Applauds the demos
Sponsoring change
- Chooses which process to redesign
- Names an owner, with time and budget
- Gives champions a mission, and expects results
- Decides what stops
Figure 10 · Both matter. Only one transforms.
Opportunities first, technology second.
The problem
Many organisations treat generative AI as a technology question. They start by deploying a platform, such as a gateway to several AI models, or by switching on the AI features their software providers keep adding. Only then do they ask what to use it for.
What is added
Start with the list of opportunities. Prioritise them, then choose the approach, and only then the technology. Prove the best case with a pilot, then decide how to scale, using what you learned as a constraint, including sovereignty and pricing.
- 01Opportunities
- 02Prioritise
- 03Approach
- 04Technology
- 05Pilot
- 06Scale
Opportunities first, technology second.
So the question is not whether a feature works. It is whether it solves a problem that matters, and whether it is the best use of your capacity.
Four decisions that cannot be delegated.
The problem
Some choices cannot be handed to a technology team, a consultant, a vendor, or the models themselves. They need leadership judgement.
What is added
Four decisions nobody else can take
- 1What is AI for, here?All four are legitimate. But they lead to very different investments, and an organisation that never chooses ends up doing a little of everything and a lot of nothing.
- 2How much risk will we carry?Chapter 4 showed that some risks never go away. Deciding how much of them to accept, in exchange for what benefit, is a leadership decision.
- 3How dependent on vendors will we become?The more autonomy you switch on, the more your daily operations depend on platforms you do not control.
- 4Where do we invest first?In the opportunities that matter most, chosen as section 5.5 describes, and in the foundations those opportunities need.
Time saved is an input, not an outcome.
The problem
It is a fair question, and the usual answers are misleading.
What is added
When you do assess value, keep five things apart: the capacity released, the quality or service improved, the costs actually avoided, the new work made possible, and the total cost of adoption and operation.
The most valuable uses have no “before” at all. Traditional automation makes known work cheaper. AI also makes work possible that was never done, because it was too expensive, too complex or needed skills nobody had: modernising a system nobody dared touch, making a complicated product understandable, finally using a knowledge base everyone had given up on. A business case that compares the new way with the old way cannot value work that simply did not exist.
And ask a second question alongside “how can we do our work more efficiently?”: “what can we now do that we could not do before?”
The better question is what those minutes turn into.
Prompt for leaders · A2
A panel of seven experts
Use it to stress-test one of the four decisions in chapter 5, or any decision where you suspect you only see one side. Replace the example in brackets with your own decision.
ROLE: You are a Master Moderator orchestrating a high-level, multi-disciplinary think tank. GOAL: Conduct an in-depth, multi-perspective debate on the decision below. Embody a panel of 7 experts with radically different worldviews to expose its full complexity, identify friction points, and discover unexpected consensus. DECISION: [DESCRIBE THE DECISION YOU ARE FACING, e.g. "Should process redesign, rather than individual productivity, be the main goal of our AI programme next year?"] THE PANEL: 1. The Historian: past precedents, historical cycles, long-term patterns. 2. The Pragmatist: execution, technical constraints, logistics, feasibility. 3. The Ethicist: moral implications, equity, justice, potential for harm. 4. The Visionary: long-term transformations and paradigm shifts. 5. The Humanist: the individual human experience, psychology, culture. 6. The Economist: incentives, market dynamics, costs and benefits. 7. The Devil's Advocate: stress-tests every assumption and champions the worst case. RULES: 1. No introductory filler. Begin directly with the debate. 2. Keep strict boundaries between personas. 3. Keep arguments sharp, specific and grounded in the decision. OUTPUT FORMAT: PART 1: THE PERSPECTIVES For each expert: The Stance (1-2 sentences); The Case (2-3 bullet arguments); The Challenge (the most piercing question this expert asks the panel). PART 2: THE SYNTHESIS The Fracture Points: the two greatest points of conflict. The Common Ground: a specific, actionable compromise that bridges the divide.
The panel will not decide for you. Read the fracture points first: they show where your decision is most likely to be challenged. Remember that these are seven perspectives generated by one model, not seven independent experts. Never paste confidential information into a tool your organisation has not approved.
Three moves to make now
- Use AI yourself every week, and say so. You cannot govern, fund or sponsor what you have never used.
- Sponsor one process change personally. Choose it, name its owner, give it time and budget, and decide what stops.
- Put AI adoption and AI risk on your leadership agenda every quarter, including the four decisions above.
Will AI take my job?
Leaders hear this question more than any other, often in private. Here is my honest answer.
So yes, the work is changing. Companies will expect people to do far more on their own, and the people who learn to work with AI will be the ones able to do it. Those who cannot will find it harder and harder to keep their place. Some companies will also simply cut costs and headcount. The reality will probably be a mix of both. Learning to work with this technology takes time, so start now.
Never promise
- “Nobody will lose their job because of AI.”No leader can guarantee it.
- “AI will only take away the boring tasks.”It also makes some work more demanding: reviewing, judging, deciding.
- “The time you save is yours.”The business will expect some of it back.
- “Your role won’t change.”Almost every role will.
- “The training we give you will be enough.”The tools change too fast for any training to be enough.
- “The tools are accurate and safe.”They can be confidently wrong.
- “All the gains will be reinvested in people.”Nobody knows the gains yet, and the finance director will have a say.
What you can promise
- Honesty about decisions, as they are made.
- Time to learn, protected from daily firefighting.
- A voice in redesigning the work.
- Support for those who struggle.
Some jobs will change beyond recognition, and some will go. The best protection is to start learning now. The best thing leaders can do is make that learning possible.
Is regulation a brake?
In my experience, today, mostly yes. Not because the rules are wrong, but because they are unclear.
Can be validated · Use cases built or bought as projects
- Check them against the rules before they go live, with the two gates described in chapter 4.
Cannot be approved one by one · Everyday use, one prompt at a time
- Rely on people: training, clear principles, and a culture where people ask before they guess.
Figure 11 · Compliance cannot rely on approval alone
Regulation is not a reason to wait. Govern what you can approve, and educate for what you cannot.
Tool · Self-assessment · 1 of 2
Where are you on the map?
The map is a shared language, not a score. Start by placing one real use of AI on it. Then repeat for two more. If the answers differ a lot, that is normal: organisations move unevenly.
1 The use
2 The role
How much does AI do on its own?
3 The horizon
What is being changed?
4 The blast radius
How far could a mistake spread?
ReachStakesActionFrequencySensitivityHuman checkA small blast radius does not mean no risk. But training, clear rules and personal responsibility can manage those risks.
A large blast radius needs much more: continuous monitoring, a named owner, a way to stop the system quickly, and a record of what it did.
That is why the real assessment must be made use case by use case.
5 Who stays responsible?
6 What must be in place before it goes further?
Which enabler is weakest for this use?
Tool · Self-assessment · 2 of 2
Ten questions for your organisation
Answer yes or no, honestly.
01 Do the leaders in your organisation use AI themselves every week, and say so? Ch. 5
02 Have your leaders said explicitly what AI is for in your organisation? Ch. 5
03 Is at least one process redesign personally sponsored by a senior leader, with an owner, time and budget? Ch. 5
04 Is there a named owner, with time and budget, for each of the five enablers? Ch. 3
05 Is capacity protected for building the foundations, or does daily firefighting always win? Ch. 3
06 Before a pilot starts, do you decide who will run it afterwards? Ch. 3
07 Were your risk, security and compliance teams trained on AI before being asked to judge it? Ch. 3–4
08 When the answer is no, does it come with the conditions that would turn it into a yes? Ch. 4
09 Is the time it takes to reach a decision on an AI pilot measured, and counted in weeks rather than months? Ch. 4
10 Are risks that cannot be eliminated decided explicitly by senior management, rather than left to a single reviewer? Ch. 4
This is not a score. Don’t add up your yeses. One missing control can matter more than nine good answers.
Start with three questions. A no to 04 (owners of the enablers), 06 (who runs it after the pilot) or 10 (risks nobody can remove) comes first, whatever else you answered.
For every no, write three things. Who owns it. What evidence would turn it into a yes. The next action, with a date.
One warning: most people answer for the organisation their policies describe, not the one they work in. That gap is exactly what this playbook is about.
The map, step by step