A playbook series · Volume 2

Escape from theory land

Volume 2 · Working with AI

A practitioner’s playbook for leading AI adoption in complex, regulated organisations

Your team has the tools. The value comes from how it works with them.

Scroll to zoom in

01CREATE02CONSUME03REVIEW04TRANSFORM
Stop asking what AI can automate. Look at what you already do all day: you create, consume, review and transform knowledge. That is where AI fits.
You are hereYou are here: the map in one page

Your team has the tools. The value comes from how it works with them.

The problem

Most organisations have now given their people AI tools, and the usage numbers look good. Yet in most teams, little has changed in how the work gets done.

What is added

Horizon 1, where most people work today as doers and editors. Its chapters build two of the five enablers, skills (chapter 3) and knowledge (chapter 4), and look at a cost the map does not show: the attention and judgment of the people in the loop (chapter 5).

Volume 1 drew a map of AI adoption. This volume zooms in on one part of it. If you have not read volume 1, this page is all you need.

Three horizons. Where the change happens: in one person’s work (Horizon 1), in a process (Horizon 2), in the organisation (Horizon 3).

Five enablers. What must be in place to move safely: skills, knowledge, security, governance and operations.

In the bookYou are here: the map in one pagepages 3–5
011.2 Knowledge work is the whole job, not part of it

They are not a list of AI features. They are a description of your day.

The problem

Most people who struggle to find a use for AI are not short of ideas. They are looking in the wrong place. They look for an “AI project”, when the work AI is good at fills their own calendar from morning to evening.

What is added

This chapter gives you a simple way to see it. Four activities describe almost everything a knowledge worker does. Once you look at your day through them, the question changes from “where could we use AI?” to “where do I start?”

01CreateProduce something new: ideas, texts, analyses, proposals, code.Start by talking. Let the AI interview you.
02ConsumeTake information in: read, listen, extract what matters.Choose the format that fits the moment.
03ReviewJudge quality: check, proofread, spot gaps, give feedback.A coach before the committee.
04TransformTurn content into another form, for another audience.One insight, many audiences.

Figure 1 · The four activities of knowledge work

Take a one-hour presentation. Someone created it. The audience consumes it. A colleague reviewed it beforehand. And the recording can be transformed into speaker notes, a learning module or a written narrative. One hour, all four activities.

And nobody else will do the mapping for them. An IT department will never find your best use cases, because it does not live your day.

In the bookChapter 1 · Knowledge work, rethoughtpages 9–14
011.7 Knowledge is a cube

Keep the richest source. Every summary removes a face of the cube.

The problem

For most of our working lives, knowledge has been flat. It took the form of a finished document: a page, a report, a slide deck, written once for one audience and one purpose. If you needed something different, you started again.

What is added

AI changes the shape of knowledge. I describe it as a cube. You can zoom in for detail or zoom out for the big picture. You can turn it to see it from another side: the risk view, the customer view, the cost view. You can change the audience and change the medium. The same knowledge, many faces.

Zoom in or outFrom one detail to the big picture.
Turn itThe risk view, the customer view, the cost view.
Change the audienceExperts, executives, newcomers, customers.
Change the mediumMemo, slides, audio, FAQ, training.

Old shape New shape: one rich source, many faces

Figure 2 · Knowledge is a cube

Take the transcript of a one-hour working session. From that single source, you can produce minutes, a focused memo on one decision, a list of requirements, coaching feedback for the facilitator, a training asset for newcomers, or a narrative for people who missed it. None of these needs to be decided in advance.

Prompt · appendix B1

The week audit (chapter 1)
Interview me about my last working week. Go through four activities, one at a time: creating knowledge, consuming it, reviewing it and transforming it into other formats. For each one, ask me what I did, how long it took and what frustrated me. Ask one question at a time. At the end, list the five tasks where AI could help me most, and explain why for each.

Tip: run it on your own week first, then with your team, and look for the frustrations that come up more than once.

Three moves to make now

  1. Audit your own week first. Take thirty minutes, answer the four questions for yourself, and try one idea per activity before asking your team to do the same. People follow what their manager visibly does.
  2. Keep the richest source. Record working sessions with consent and within the rules, keep the transcripts, and stop deleting the full version once a summary exists.
  3. Collect your team’s answers in one place. Run the audit as a team, look for the frustrations that appear more than once, and give the best ideas to your champions to extract, package and spread.
In the bookChapter 1 · Knowledge work, rethoughtpages 18–23

Interlude

Twelve weeks in a day and a half

The write-up was long, and nobody was going to read it in the middle of a debate. So I asked for an infographic, put it on the big screen, asked for three minutes of silence and explained the model. Someone said: “It’s only the first morning, and the problem is solved.”

It wasn’t. I said so straight away: the model was a baseline, and now it needed their expertise. The experts challenged every dimension and definition, giving their feedback to the AI to rework the model rather than editing it by hand. That was deliberate: feedback tells the AI why a change is needed, and that context may matter later. The AI had accelerated the work. The experts decided what was valid.

We did not win on paper. By our estimate, we had done in a day and a half what would normally take twelve weeks of full-time work. What we built was a prototype, not a product: turning it into production software would still need proper engineering.

AI got us to the starting line in a morning. The experts still had the last word.

In the bookTwelve weeks in a day and a halfpages 24–26
022.1 Three rules before any tool

The more convincing it looks, the more carefully you need to read it.

The problem

Most people use AI the way they would use a clever intern. They paste in a document, type “summarise this” or “make it more professional”, and take what comes back. They delegate when they should collaborate. It saves time. It also caps what AI can do for them at the level of the request.

What is added

Three short rules come before any tool. They are easy to remember and hard to practise.

  • Use your brain first. Before you ask the AI anything that matters, take a few minutes to form your own view. What do I think the answer is? What would a good result look like? If you ask first, the AI’s answer becomes your starting point, and it is very hard to think past a well-written answer. Your own view is also what lets you judge the output: without it, you have nothing to compare against.
  • Keep a human in the loop, always. AI has no intent. People define the objective, choose the information it works on, approve what it produces and remain accountable for the consequences. That does not change because the output looks convincing. The more convincing it looks, the more carefully you need to read it.
  • Keep learning. The tools change every few months. Features appear that make last quarter’s workaround unnecessary. Explore what your tools already do before asking for something new to be built. A good test before any development request: have we exhausted the value of what we already have?

One boundary sits under all three rules: use public tools for public information, and your organisation’s approved tools for anything internal.

In the bookChapter 2 · The craft of working with AIpages 27–28
022.2 A thinking partner, not an intern

AI takes part in the whole journey. I decide at each turn where it goes next.

The problem

“Summarise this” is a single request at the end of a piece of work.

What is added

The real change comes when AI is present at every step of the work, with your judgment between the steps.

  1. 1
    The sparkNo AI yet: a question worth exploring.
    You
  2. 2
    Talking it outSpeak freely; the AI captures and structures.
    You + AI
  3. 3
    Reviewing the recordEvery detail kept, nothing deleted.
    You check
  4. 4
    Reviewing the research planAdjust the plan before it runs.
    You check
  5. 5
    Discussing with the documentSelect passages, push back, ask for examples.
    You + AI
  6. 6
    UnderstandingYour own position emerges.
    You
  7. 7
    The final outputWritten with your refined instructions.
    You check

AI does the heavy lifting. You decide where it goes next.

Figure 3 · The flow: AI at every step, your judgment between the steps · checkpoint · you decide what comes next

This is not a recipe to copy. It is a way of seeing: AI takes part in the whole journey, and I decide at each turn where it goes next. Look at your own work, preparing a client meeting or a budget proposal, and ask where AI could help at more than one step.

You get your own assumption back, nicely written.

The method I use has three steps, with a human decision in the middle:

  • Options. Describe the goal and the constraints, and ask for several different approaches. Then stop: “Give me three distinct ways to approach this, with the main risk of each. Don’t write anything yet.”
  • Your choice. You read the options. One is too risky, one is solid, one rests on something that doesn’t exist. You choose, and you say why.
  • Execution. “We go with option two. Write it, and make sure you address the concern about timing.”

The human checks between every step. This is the part people are tempted to skip. When AI steps pass work to each other with no one in between, the result drifts towards the average. The unusual idea gets smoothed out. The bold choice becomes a safe one. Your check between steps is where your judgment and your own ideas stay in the work.

Prompt · appendix B3

The Architect: options before instructions (chapter 2)
Step 1 · Options
My goal is [goal]. My constraints are [constraints]. Give me three distinct ways to approach this, with the main risk of each. Stop there. Don’t produce anything yet.
Step 2 · Your choice. Read the three options. Discard what is too risky or rests on something that doesn’t exist. Choose, and say why.
Step 3 · Execution
We go with option [X]. Produce it, and make sure you address [the concern you have].

Standing instruction (in your AI tool’s settings)
Before doing what I ask, briefly list the trade-offs of my approach. If a better approach exists, say so before you start. Disagree with me when it is justified.

Tip: use options when you know the goal but want the path tested. Use clear instructions when you know exactly how the work should be done, or when you can’t yet judge the options.

In the bookChapter 2 · The craft of working with AIpages 29–36
022.6 Check the output: whoever signs, owns it

A reference you have not opened is not a reference.

The problem

A better request gives a better answer. It does not give a correct one. That is still your job.

What is added

The same asymmetry applies to every knowledge worker. When a document leaves your hands with your name on it, it is yours, whatever share of it the AI wrote. “The AI said so” is not a defence, and nobody will accept it as one.

  • Ask for the sources, and open them. A reference you have not opened is not a reference. AI can produce very convincing citations to things that don’t say what it claims, or don’t exist.
  • Make it argue against itself. “What is the strongest argument against this conclusion?” or “What would a sceptical expert say?” This counters the AI’s tendency to agree with you. Assembling a panel of experts with opposing views goes further in the same direction.
  • Do a blind read. Give the finished text to a fresh conversation without your brief, and ask what it thinks the main message is. If the answer is not what you intended, the text doesn’t say it yet.
  • Cross-check what matters. For facts and figures that carry a decision, verify with a second source, or a second model, before relying on them.

Match the depth of checking to what is at stake. A draft for yourself needs a quick read. A figure in a board paper needs its source. A message to thousands of customers needs more than one pair of eyes. Checking also has a human cost, which chapter 5 looks at.

In the bookChapter 2 · The craft of working with AIpages 37–40
022.7 Don’t default to agents

Complexity has to be earned.

The problem

Agents are the word of the moment. Many people’s first instinct, once they find a good use for AI, is to build one. Often, that is the wrong instinct.

What is added

An agent is like a horse wearing blinkers. It is deliberately narrowed: given fixed instructions, a fixed body of knowledge and a fixed purpose. That narrowness is exactly what makes it useful for a task that repeats in the same way. It is also what makes it the wrong tool for everything else.

Tool · Figure 4 · Before you build an agent

Before building an agent, ask three questions:

  1. Repeats often? Does the task really repeat, in the same form, often enough to be worth it?

  2. Works by hand? Does the method already work by hand?

  3. Has an owner? Who will look after it?

THE DEFAULT Conversation, research, notebook

ANY “NO”: STAY IN CONVERSATION

THEN BUILD An agent: a product to own

A broader question sits behind all three, and it belongs on every team’s wall: have we used everything our current tools can already do, before we build something new that we will have to maintain?

I have seen the opposite happen. A team spent months building a custom assistant to answer employees’ questions about an internal policy. It worked. But the standard AI assistant everyone already had could answer the same questions from the policy document itself, with no build at all. The months were gone, and the team now owned something that needed testing, maintenance, support and, one day, retirement. Every custom build creates those obligations. Complexity has to be earned.

Three moves to make now

  1. Write your view before you ask. For your next important request, write three lines on what you think the answer is. Compare them with the AI’s answer afterwards.
  2. Turn one habit into a standing instruction. Add a line to your AI tool’s settings asking it to list trade-offs and to disagree when it is justified.
  3. Map one piece of your own work as a flow. Pick a recurring task, list its steps, and mark where AI helps and where you check. Share it with your team: it is the start of a method others can reuse.
In the bookChapter 2 · The craft of working with AIpages 40–45
033.1 Purpose plus learning

It is a ramp, not a course.

The problem

Ask a busy professional to block twenty hours for a course on prompting and you will get polite agreement and an empty seat. Their calendar is already full of meetings and urgent problems. And classic training, theory first, exercises second, application someday, works badly for AI. The tools change faster than any course can be written. Theory without a real problem is also boring.

What is added

People do make time for one thing: a problem they care about. The formula is simple: purpose plus learning. Don’t learn AI for its own sake. Pick a real problem, personal or professional, and use AI to solve it. Then ask the AI to teach you along the way.

Prompt · appendix B5

The learning contract (chapter 3)
I have a real problem to solve: [describe it, and why it matters to me]. You are an expert in [the AI tool I use] and know all its capabilities. Build me a step-by-step plan that solves this problem and, at the same time, teaches me the capabilities of [the tool] I need, from brainstorming to delivery. For each step, tell me which capability we will use and why it fits, show me how to use it on my problem, then let me do it myself and give me feedback. Don’t hand me a ready-made solution. Start by asking me three questions about the problem and my current level, then propose the plan and wait for my approval.

Tip: choose a problem you care about, personal or professional. The motivation carries the learning.

With that framing, the problem becomes your excuse to learn. You ask the AI to solve two things at once: the problem in front of you, and the gap in your own skills. You kill two birds with one stone. You get the solution, and you build both literacy (knowing what AI can do) and fluency (knowing how to make it do it). The project pulls the skills in as they become necessary: you learn to organise information when it gets messy, to generate a working tool when you need one, and to give precise feedback when the result is not quite right. Each problem solved this way leaves you equipped to take on the next, harder one. It is a ramp, not a course.

WhoWhat they needHorizon
Every knowledge workerKnow what the AI tools at their desk can do and bring them into their daily work; frame a request, give context, judge the answer; know the risks and what is not allowed; stay accountable1
Managers and team leadsThe above, plus spotting use cases in their team’s work and protecting time to learn1
ChampionsFluency, plus the ability to extract, package and teach what works1 and 2
Process owners and product leadersEnough practice to redesign a process and judge what AI can reliably do in it2
Risk, security, legal and complianceEnough practice to assess real risks and set proportionate requirements1 and 2
Engineers and architectsDeep technical skills to build, integrate and operate AI in production2

The risk and control row is the one organisations most often forget.

Figure 5 · Literacy by responsibility

In the bookChapter 3 · Learning and adoptionpages 46–52
033.4 Stop selling wow, start selling phew

Sell the relief, not the wow.

The problem

Licences are paid for, the tools work, and usage stays flat. Ask why, and the most common answer is: “We just don’t have the time.” It is tempting to treat that as an excuse. Sometimes it is simply true.

What is added

The answer is to stop selling the wow, what the AI can do, and start selling the phew, the relief it brings:

  • Bring AI into the tools people already use, so they feel the benefit before they measure the effort.
  • Show relief, not features. Instead of a spectacular demo, show how AI removes the task people hate most at the end of their week.
  • Protect the time saved, explicitly. If a team saves time, give part of it back for deep work or learning. Chapter 5 returns to this.
  • Give people a safe place to be beginners, in small groups or one-to-one, not in front of their whole department.

The answer is to move from push to pull. Instead of pushing a catalogue at people and hoping they find what they need, let the training be pulled by the problems they actually have:

Adoption succeeds when people stop seeing AI as one more thing to learn and start seeing what it gives them. Going home on time is part of it, but only part. The bigger gain is in the work itself: more room for creativity, better quality, more resilience when things go wrong, and skills that keep people in demand.

In the bookChapter 3 · Learning and adoptionpages 52–56
033.5 Finding use cases as a team

The people doing the work already know where it hurts.

The problem

Many managers are told to “implement AI” and find themselves in front of an empty page, wondering which of their team’s hundreds of tasks to start with. The mistake is thinking the manager has to find the answer alone. The people doing the work already know where it hurts.

What is added

Chapter 1 gave individuals four questions about their own week. The method below, which I call the Tomorrow Protocol, does the same for a whole team, and uses AI to make listening possible at scale.

  1. 1
    Know what is possibleMap the tools you have
  2. 2
    Open with trustFind the rocks in their shoes
  3. 3
    Prepare the questionsAI drafts a five-whys guide
    AI does the heavy lifting
  4. 4
    Let people speakAlone or in small groups
  5. 5
    Voices into notesAI structures each recording
    AI does the heavy lifting
  6. 6
    Find the patternsAI clusters: a heat map
    AI does the heavy lifting
  7. 7
    Decide togetherThe team chooses where to start

AI prepares, formats and clusters. The decisions, at the start and at the end, stay human.

Figure 6 · The Tomorrow Protocol: from empty page to shared priorities

The protocol has one limit worth knowing. It tells you where it hurts. It does not tell you the best way to fix it.

So treat the protocol as the first half. Use it to choose the problems. Then put the two halves of the answer in the same room: the people who know the work and the people who know the tools, on the same real problem, in a build day or a hackathon (see 3.3). That is where the best solutions come up, and often the problems nobody had listed.

In the bookChapter 3 · Learning and adoptionpages 57–60
033.6 Champions with a mission

What counts is whether the work changes where champions are.

The problem

Every organisation already has AI enthusiasts. On their own, they help the colleagues sitting next to them, and their discoveries stay local.

What is added

A champion programme turns that energy into a structure that spreads what works. It is a permanent capability, not a launch campaign.

  • A centre and a network. A small central team sets the direction, builds shared material and keeps the link with leadership. The champions, spread across teams, adapt that material to their colleagues’ real work, support them day to day, and bring back feedback and success stories. The centre gives consistency; the network gives relevance.
  • Coaches, not IT support. Champions show colleagues how to work differently. They are not a help desk for access problems or broken tools. If that line isn’t drawn clearly from the start, they will spend their time on tickets.
  • Recruit the sceptics too. Choose people who communicate well and are respected by their peers, and include thoughtful sceptics: they understand the limits and risks of AI, make the programme more credible and help write guidelines people trust.
  • Plan for people leaving. Champions change roles or leave. Give newcomers an experienced buddy, and make leavers hand over what they built and knew.
  1. 01
    Community healthAre champions active and engaged?
  2. 02
    Champion effectivenessAre they doing the right things, according to their colleagues?
  3. 03
    AdoptionDo teams with an active champion use AI more, and better?
  4. 04
    Business impactNot only time freed: work never done before, better quality, fewer errors and exceptions, new ideas.

Each layer only counts if the next one moves

Figure 7 · Measure what matters, in four layers

The number of champions is the least interesting of these measures. What counts is whether the work changes where they are.

To judge what an AI produces, you need to know the subject without it.

  • Without AI, test the foundations: the knowledge someone needs to spot a wrong answer.
  • With AI, test how they work: how they frame the request, what context they give, what they check and what they reject.

Three moves to make now

  1. Pick your own project and sign the contract. Choose one real problem you care about, personal or professional, and use the learning-contract prompt to solve it. Tell your team what you learned.
  2. Run the Tomorrow Protocol with your team. Start by listening for the rocks in their shoes, and make the decision on priorities together.
  3. Give your champions a mandate. Name them, protect part of their time, include at least one thoughtful sceptic, and give them a direct line to someone who can act.
In the bookChapter 3 · Learning and adoptionpages 61–67
044.1 Where do I put this?

If your organisation cannot say which source is the reference, AI will guess.

The problem

AI has no such head. When it searches your content, it finds all five copies and all three versions, and has no way to know which one the team actually trusts. It will pick one and answer with complete confidence. If it picks the old one, the answer is wrong, and nothing in its tone will tell you.

What is added

This is why one question matters more than any other when you connect AI to your knowledge: for each type of information, which source is the reference?

A simple convention helps, and many organisations use a version of it: me, we, many. Personal drafts live in your own space. Active collaboration happens in the team’s shared space. The final reference is published where everyone who needs it can find it, moved there, not copied.

MeYour own spaceDrafts
MOVE
WeThe team's spaceActive collaboration
MOVE
ManyPublished referenceFor everyone who needs it

MOVE IT, DON’T COPY IT: EVERY COPY IS ONE MORE VERSION TO CONFUSE THE AI

Figure 8 · Me, we, many

In the bookChapter 4 · Knowledge foundationspages 68–70
044.2 Ready for people is not ready for AI

Intelligence is not the same as context.

The problem

A common belief in leadership meetings goes like this: “Let’s connect all our documents to the AI. It is smart enough to figure it out.”

What is added

People read documents with a lifetime of context. They read around gaps, recognise an outdated version, and understand what a picture is meant to show. AI reads what is actually there.

A person readsAn AI reads
A pictureUnderstands its intent from the page around it.Has to guess what it shows.
A large tableLinks each number to its heading at a glance.Can lose the link, especially across pages.
A bold line used as a titleSees a heading.Sees ordinary text.
“Contact the IT lead”Knows who that is.Has no idea.
The 2019 and 2024 policiesKnows which one is current.May pick the old one.
Three pasted copiesTrusts the one the team uses.Finds all three, picks one.
“The new policy”Knows which policy is meant.Cannot tell which, or since when.

Figure 9 · The same document, read by a person and read by an AI

The good news is that most fixes are about format, not content. You rarely need to rewrite a document. You need to make what is implicit explicit.

In the bookChapter 4 · Knowledge foundationspages 70–73
044.3 Don’t police, curate

A document is only as ready as its weakest point.

The problem

Faced with the mess, the classic answer is to enforce strict rules on every document before AI touches anything. In a large organisation, that rarely works.

What is added

The alternative is to curate: apply the rules only where AI will actually read, and raise the bar step by step.

LevelWhat it looks like
1 Not yet for AINo clear owner, no review, sensitivity unknown, duplicates, content that is unstructured or out of date
2 Ready for pilotsOwner named, key information recorded, reviews happen, but nothing is formalised; some formatting gaps
3 Ready for productionOwner changes and reviews follow a written procedure; accurate, clearly scoped, well structured, every image described
4 Self-maintaining (an aspiration for now)Ready for production, plus automation: information about the document, reviews and broken links maintained by the system

Four levels of readiness.

A document is only as ready as its weakest point. One missing element, an owner who left, an undescribed diagram, caps the whole document at the lower level.

Tool · appendix B6 · Is this document ready for AI? (chapter 4)

Answer each question for one document. A document is only as ready as its weakest answer.

Governance

  1. Is one person or team named as owner, visibly?

  2. Is the date of the last review recorded, and within the rhythm (yearly for important content, every two years for the rest)?

  3. Is its readiness level recorded?

  4. Is its sensitivity recorded, and is that level approved for the AI tool that will read it?

  5. For a policy, process or report: is the date it takes effect recorded?

Content

  1. Is it accurate, checked by someone who knows the subject, and are all links working?

  2. Does it say, at the top, what it covers, for whom, and what it does not cover?

  3. Is every implicit reference made explicit (“the new policy” becomes the policy’s full name and date)?

  4. Are internal terms and acronyms defined, and internal prerequisites linked?

  5. Is it the only copy of this knowledge, with other documents linking to it?

Format

  1. Is it in an editable, structured format, not a presentation-only file?

  2. Are headings real headings, used in order?

  3. Do tables have headers, repeated on long tables?

  4. Does every image or diagram have a description that makes sense without seeing it?

  5. Are view and edit rights defined separately?

any “no” in governance, or content that is inaccurate, duplicated or locked in a presentation-only file, means not yet for AI: keep it away from AI.

Governance complete but gaps in content or format means ready for pilots.

All “yes”, with owner changes and reviews following a written procedure, means ready for production.

Reading the result: any “no” in governance, or content that is inaccurate, duplicated or locked in a presentation-only file, means not yet for AI: keep it away from AI. Governance complete but gaps in content or format means ready for pilots. All “yes”, with owner changes and reviews following a written procedure, means ready for production.

In the bookChapter 4 · Knowledge foundationspages 72–75
044.5 Treat wrong answers as bugs in the knowledge

Bad knowledge does not stay still: it multiplies.

The problem

When an AI gives a wrong answer, the reflex is to blame the AI, or to rewrite the question until the answer looks right. Both miss the point.

What is added

Very often, the AI answered faithfully from a document that was wrong, outdated or unclear. The bug is in the knowledge.

At team level, the answer is simple: every reference document has a named owner, a person or a team accountable for keeping it right.

Flawed sourceAI answerSaved as a new documentMore flawsLEFT ALONE: IT MULTIPLIESConnectDetectFix at the sourceTest againFIND AND FIX

Figure 10 · Two loops: left alone, or found and fixed

Two habits keep this going after launch. First, make every “thumbs down” from a user lead to someone looking at the source, not just at the answer. Otherwise people simply close the window, and you learn nothing. Second, build knowledge into how work ends: a project is not closed until its key findings are written up in a form the AI can use. That prevents new debt from piling up behind the old.

Prompt · appendix B10

The knowledge health check (chapter 4)
You are a knowledge quality reviewer. Review the documents in [folder or space]. For each document, check: is an owner named? Is the last review date recorded and within [12 or 24] months? For policies and processes, is the date it takes effect recorded? Is a sensitivity level recorded? Does it contain vague references (“the new policy”, “the IT lead”) or acronyms that are never defined? Are there images or diagrams with no description? Does another document say the same thing, or the opposite? Give me a table: document, problems found, suggested readiness level (not yet for AI, ready for pilots, ready for production). Then list the five fixes that would help most. Flag problems; don’t change anything.

Tip: a starting point to adapt. The AI flags; people decide.

Tool · Try this · The ten-question test

Before connecting AI to a team’s knowledge, or when an existing assistant disappoints, spend an hour on this.

  1. Collect ten real questions the team is asked often. Take them from emails, tickets or the newest team member.
  2. Include some hard ones: a question that needs a table, one that needs two documents together, one where an older version exists.
  3. Ask the AI all ten, using the sources it will really use.
  4. Score each answer: right, partly right, wrong.
  5. For every answer that isn’t right, find the source document and note the reason: wrong version, unreadable table, image without text, missing context, unclear owner.
  6. Fix the three most common reasons first, then run the ten questions again.

Fix the three most common reasons first, then run the ten questions again.

    The result is a short, concrete list of what to fix, based on evidence rather than opinion.

    Your answers, to copy
    In the bookChapter 4 · Knowledge foundationspages 75–78
    044.7 Access is a knowledge question

    Sensitivity beats readiness.

    The problem

    The most valuable knowledge in an organisation is often the least documented. Why a platform was built this way. Which workaround matters and which is obsolete. What the expert would try next. It lives in people’s heads, and it leaves when they do.

    What is added

    One rule sits above all the others: sensitivity beats readiness. A perfectly structured, fully reviewed document is still off limits to an AI tool if its sensitivity level is not approved for that tool.

    AI offers a simpler route. Don’t start with a blank template. Let the expert talk. Record a series of conversations, with their consent and within your organisation’s rules, and keep the full transcripts. Then use AI to derive what the team needs from that rich source: a playbook, a testing approach, a set of frequently asked questions, onboarding material for newcomers. This is chapter 1’s cube in action: one rich source, many outputs, decided afterwards.

    In the hackathon story, I asked the AI to reason across the thousands of documents I was allowed to access. That phrase matters. An AI assistant working on your behalf sees what you can technically reach. Not what you should see: what you can reach.

    • Check permissions before you connect a source, not after the first incident.
    • Keep test and technical content out of reach. Logs and test files can contain instructions that confuse an AI, or data that was never meant to be shared.
    • Don’t build one agent with access to everything. It is tempting: one assistant that knows the whole organisation. In practice, it combines every permission problem, every outdated version and every unclear source into a single point of failure. Narrow agents with clearly defined sources are easier to trust, to test and to fix.

    Three moves to make now

    1. Name the reference sources. For your team’s five most important types of information, write down which source is the reference and who owns it. Where you can’t answer either, you have found your first problem.
    2. Run the ten-question test on one real use case, and fix the three most common causes of wrong answers at the source.
    3. Replace the next handover template with a conversation. When an expert moves on, record them talking (with consent), keep the transcript, and build the handover from it.
    In the bookChapter 4 · Knowledge foundationspages 79–85

    Interlude

    I was wrong about the dog

    Then I felt the pain myself. When I ask the people I work with for something, they now deliver so fast that I have entered a race of reviews. AI helps with the review, but I remain accountable for the result, so I still have to think, steer and give feedback.

    My days used to mix tasks that needed different levels of concentration. Now my brain is in review mode all the time, because everyone around me works at that speed. I had become the bottleneck. Not because I was slow, but because judgment does not speed up the way production does.

    I can feel my brain overheating, and I need to cool it down. By walking the dog. Part of the time saved has to go into letting the brain cool down and the mind wander. That is when I am most creative, and when I come back to AI with better ideas.

    Attention is the new bottleneck.

    In the bookI was wrong about the dogpages 86–87
    055.2 Checking is harder than creating

    The work has not disappeared. It has moved.

    The problem

    When someone uses AI to produce work, their role changes. They stop being the creator and become the reviewer of what the machine produced.

    What is added

    There is a paradox at the heart of AI productivity. Producing a draft now takes seconds. Checking it still takes a person, with full attention. If a team produces several times more output, the checking load grows in proportion, while the number of people and hours available to check stays the same. The work has not disappeared. It has moved.

    The checking itself is tiring in a particular way. I call it the toggle tax. All day long, the reviewer switches between two opposite states: trusting the AI’s fluent, confident answer, and doubting every word of it. Each switch costs a little. You don’t feel it in the moment. You feel it at the end of the day, or the end of the month, as a loss of sharpness you can’t quite explain.

    Not every output deserves the same scrutiny, and pretending otherwise burns people out. Match the depth of review to what is at stake:

    Low stakes

    An internal summary for your own useThe author's own check

    Medium stakes

    A regular team reportOne review by the team lead, and it stops there

    High stakes

    Anything for clients, regulators or the boardA full review chain, with named people at each step

    Figure 12 · Match the depth of review to the stakes

    Producing a draft now takes seconds. Checking it still takes a person, with full attention. As output grows, the effort shifts from producing to reviewing.

    Before AI

    ProducingReviewing

    With AI, same team, more output

    ProducingReviewing

    Illustration of the shift, not measured data

    Figure 11 · Where the effort moves

    We audit the data. Who audits the auditor?

    In the bookChapter 5 · The human costpages 88–92
    055.3 Cognitive debt: brain first, again

    Thinking first, then using AI, kept the brain working.

    The problem

    We have always handed mental tasks to tools: notes, calculators, maps. AI is the most powerful and effortless of these tools ever built. The question is what happens to the skills we stop using.

    What is added

    The order mattered. Thinking first, then using AI to improve the result, kept the brain working.

    In practice, that means changing how you ask. Instead of “give me the answer”, ask the AI to guide you: “Ask me three questions that will help me sharpen my own argument” or “Find the three weakest points in my reasoning.” You stay the pilot, not the passenger.

    Prompt · appendix B7

    The brain-first routine (chapter 5)
    Before you open the AI, write three lines: what do I think? What would a good result look like? What am I unsure about? Then ask the AI to challenge, not replace:
    
    Here is my plan. Act as a sceptical competitor: find its three biggest flaws and ask me three questions I haven’t considered.
    Propose three different structures for this document. For each one, ask me a question about my own experience that would make it more authentic.
    My definition of success here feels vague. Ask me five questions, one at a time, to help me define it.
    Explain how this process works, then tell me which two people I should talk to about how it really works in practice.
    
    Afterwards, compare: what did the AI add? What did it get wrong? What did you know that it didn’t?
    In the bookChapter 5 · The human costpages 92–93
    055.4 Protect attention and the time saved

    Value the quality of a judgment over the speed of a reply.

    The problem

    AI removes waiting. The file that took thirty seconds to load, the walk to the printer, the minute before a meeting starts: those small gaps have been shrinking for years, and instant AI answers remove more of them. It is often in those gaps that the mind digests what it just did, and where ideas come. Without them, work becomes one continuous stream of input.

    What is added

    None of this requires new technology to fix. It requires deliberate habits, and leaders who model them:

    • Keep real breaks. After a long stretch of focused work, take a break with no screen and no input: a walk, a coffee, a window. Which rhythm works best is still an open question: ninety minutes of focus, then twenty minutes with no input? Each team has to find its own.
    • Protect your best hours. Keep the first hours of the day for creating or for the reviews that matter most. Group routine AI reviews into one block later in the day, rather than switching all day long.
    • Alternate modes. Balance periods of reviewing AI output with periods of creating without it. The skills of original thinking stay sharp only if they are used.
    • Give saved time back to thinking. Don’t automatically refill the time AI saves with more tasks. How much of it should go back to deep work, learning or simply thinking? A quarter? Half? It is a question worth asking your team explicitly.
    • Slow down on purpose. If you lead a team, stop answering every message instantly. Show that you value the quality of a judgment over the speed of a reply.
    In the bookChapter 5 · The human costpages 94–95
    055.5 Don’t lose each other

    The AI should not be the destination. It should be the route to the right conversation.

    The problem

    It has become very easy to stay at your desk and work with AI instead of with colleagues. The AI is always available, never busy, never impatient.

    What is added

    The answer is not to use AI less. It is to use it as a map that points back to people.

    Using AI is not something to hide; you remain responsible for what you send.

    No one really knows. Anchor on value, not on technology.

    • Remember that no one really knows. Not the vendors, not the analysts, not the people speaking loudest about technology in your meetings. Everyone is experimenting.
    • Anchor on value, not on technology.
    • Keep experimenting, on purpose. Small, time-boxed experiments are the honest answer to uncertainty. They replace the fear of missing something with evidence of what works.
    • Choose a few trusted sources and let the rest go. Following everything is not possible, and pretending otherwise is exhausting.
    • Say it out loud. When a leader admits they can’t keep up either, it gives everyone else permission to learn instead of pretending.

    Three moves to make now

    1. Set review tiers for your team. Decide, together, which outputs need only the author’s check, which need one review, and which need a full chain. Write it down.
    2. Protect one block of thinking time. Put it in your calendar every week, and keep it free of AI reviews and messages. Tell your team why, and encourage them to do the same.
    3. Ask where the saved time went. At your next team meeting, ask what people did with the time AI saved them. If the answer is “more of the same”, agree what part of it goes back to deep work or learning.
    In the bookChapter 5 · The human costpages 95–102

    Tool · Self-assessment

    Two short checks: one for you, one for your team. Score each statement honestly: 0 = not yet, 1 = sometimes, 2 = consistently. The point is not the total. It is the lowest scores, and the chapter each one points to.

    How are you working with AI?

    You

    Statement · 0 · 1 · 2 · Chapter

    1. I use AI in at least three of the four activities (create, consume, review, transform) in a normal week.

      1
    2. I keep the richest source (the transcript, the full report), not only the summary.

      1
    3. I write my own view before I ask the AI anything important.

      2, 5
    4. For important work, I ask for options before I give instructions.

      2
    5. I open the sources of anything I send under my name.

      2
    6. This month, I learned something new with AI by solving a real problem.

      3
    7. I know which document is the reference for the information I use most.

      4
    8. I protect time for deep work, and I know where the time AI saves me goes.

      5

    Reading the score (out of 16)

    Total – (out of 16)

    0 to 5: start with chapter 1 and your own week.

    6 to 11: you have the habits; work on the lowest line.

    12 to 16: you are ready to help others; chapter 3 is for you.

    Your lowest lines

      And your team?

      Your team

      Statement · 0 · 1 · 2 · Chapter

      1. We have a list of use cases that comes from the team’s own work, not from a catalogue.

        1, 3
      2. We share the prompts and flows that work, in one place everyone knows.

        1, 2
      3. Our champions have time set aside, a mission, and a direct line to someone who can act.

        3
      4. Our risk, security and legal colleagues have been trained with the same hands-on approach as everyone else.

        3
      5. Every document our AI relies on has a named owner and a review date.

        4
      6. When the AI gives a wrong answer, we trace it to the source document and fix it there.

        4
      7. We match the depth of review to what is at stake.

        5
      8. Part of the time AI saves us is explicitly given back to thinking and learning.

        5

      Reading the score (out of 16)

      Total – (out of 16)

      0 to 5: start with the Tomorrow Protocol (chapter 3) to find where it hurts.

      6 to 11: individual gains are real; turn them into shared ways of working.

      12 to 16: you are ready for Horizon 2; volume 3 is for you.

      Your lowest lines

        Your answers, to copy
        In the bookHow are you working with AI? A self-assessmentpages 103–104

        The map, zoomed in

        You are here: the map in one pageDoerEditorApproverSupervisorGovernorHorizon 3Horizon 2Horizon 1Volume 1 drew a map of AI adoption.You are hereHorizon 1 · Doer · EditorHorizon 1Doer · Editor01Create02Consume03Review04TransformKeep the richest source.Your viewfirst.You checkcheckpoint · you decide what comes nextWhoever signs,owns itTHEN BUILDAn agent:a product to ownFinding use cases as a teamreferenceownerWhere the effort movesBefore AIWith AI,same team,more outputProducingReviewingIllustration of the shift,not measured dataLow stakesMedium stakesHigh stakesbrain first, againProtect attentionand the time savedSkillsKnowledgeSecurityGovernanceOperations1234
        New at this stepAdded earlierTo come