THE AI LEADERSHIP PROGRAMME
AI is not another IT project. It is your operating model.
20-Session Programme · 25 Minutes Per Day · Free · For sitting executives
WEEK 1 OF 4 · THE INFLECTION · SESSIONS 1–5 · WORK AT YOUR OWN PACE
The Inflection · OrientationSESSION 1
Why AI Is Not Another IT Project
Why This Matters
I want to start with the mistake that costs the most. It is not choosing the wrong vendor. It is filing AI in the wrong drawer. An IT project has a scope, a go-live and a handover; an operating model has none of those, and AI is an operating model. Get the classification wrong and the budget, the owner, the metric and the accountability all go wrong with it. This is not a technology briefing. It is the first honest look at your own company, and it ends with you writing that look down.
25-Minute Module
0:00 – 5:00
The wrong question and the right one
- AI is not a product you buy; it is a property of how the company operates. A model is only as good as the data feeding it, the workflow acting on it and the governance containing it.
- So most boards ask the wrong question. Which AI tool should we buy? anchors you in vendor hype. The right one is yours: what must our company become so that AI can actually improve performance?
- THE DOCTRINE
Bolt the smartest tool onto a siloed, manual, low-trust operating model and you get faster confusion, not better performance. The tool inherits your operating model.
If an approval takes three weeks today, a recommendation produced in four seconds only shortens the wait before the same delay.
5:00 – 11:00
The most expensive misclassification available to a board
- An IT project has three properties: a defined scope, a go-live date, and a handover to whoever runs it afterwards. An operating model has none of them.
An IT project An operating model Scope Fixed and signed Expands as decisions are rebuilt Finish A go-live date No go-live — a new rhythm Owner Handed to IT Stays with the business Failure Late and over budget Delivered perfectly, nothing changes - THE DOCTRINE
Classify AI as an IT project and you will fund it, own it and measure it wrongly. Budget, owner, board metric and accountability all follow from the filing.
I have sat in that meeting: on time, signed off, handed over, everyone pleased — and nothing in the operation moved.
11:00 – 17:00
Six shifts, none of them technological
- Readiness is not a technology inventory. It reduces to six shifts, each a change in how the business operates rather than in what it owns.
From To Manual reporting Real-time visibility Reactive maintenance Predictive reliability Fragmented data Governed intelligence Isolated departments Integrated workflows Vendor dependence Owned capability Executive opinion Evidence-based decisions - ENERGY · THE WORKED CASE
- A compressor runs degraded for a month because nobody connected its vibration trend to the deferment ledger. The data existed the whole time; the definition, the owner and the rhythm did not.
- The barrel was not lost to geology. It was lost between two departments.
OUTSIDE ENERGY · THE PARALLEL- Banking. A lender knows its arrears position on the fifteenth working day after month end, from a report assembled by hand. The early signals sat in systems it already owned.
- By then the account is in collections rather than in a conversation.
None of the six needs exotic technology. All six need management courage, and each is an argument you must win internally.
17:00 – 21:00
What changes by 2027
- By 2027 the change is not that models get cleverer, although they will. It is that AI stops being a tool some people use and becomes the way decisions are made.
- The models are commoditised. The same frontier systems reach an operator in Port Harcourt and a major in Houston in the same week. If technology were the binding constraint, this would already be solved. It is not. The constraint is discipline: build the data foundation first, prove value in bounded pilots rather than slideware, keep human judgement in the loop where stakes are high, and scale only what demonstrably pays.
- THE DOCTRINE
AI will not save your industry. A leader who embraces it will. The models are commoditised; the discipline is not.
This programme is free because the constraint was never access to information. It is whether enough senior people will do the work.
21:00 – 25:00
Your four weeks, and Section 1 of your charter
- Week 1 the doctrine. Week 2 the operating model. Week 3 trust, control and the board. Week 4 the ninety days and your capstone. The finish line is a signed 90-Day AI Operating Model Charter for your own organisation: nineteen sections, one per session.
- Four house rules. Write about your own business, never a hypothetical one. No number without a baseline. No insight without a named owner. Report a blocker the same day. The pass mark is 70.
- Tonight you write Section 1: the drawer your organisation files AI in, and the one decision you would most want made better.
Most executives cannot answer the classification question about their own company without asking two colleagues first. That is the finding, not a weakness.
Key Terms
Operating model
How a company actually collects information, validates data, decides, governs risk and measures performance. Not the organisation chart — the machinery of deciding.
Executive AI fluency
Judging where AI creates value and where it cannot, what data quality a use case needs, and how to measure its return.
Governed intelligence
Data with a named owner, a standard definition and a traceable path from source to reported number.
90-Day AI Operating Model Charter
The signed document this programme produces: diagnosis, ownership, governance standard, scoreboard and first pilot, in nineteen sections.
Self-Check — Before You Submit
1. Why is calling AI an IT project the most expensive misclassification a board can make?
- A. IT functions are under-resourced and deliver late
- B. IT projects are harder to fund than operating-model changes
- C. An IT project has a scope, a go-live and a handover; an operating model has none, so budget, owner and metric are all set wrongly
- D. AI software costs more than business cases assume
C. An IT project has a scope, a go-live and a handover; an operating model has none, so budget, owner and metric are all set wrongly — The error is one of category, not cost. Cost is the tempting answer, but such a programme can finish on budget, be handed over, and change nothing in the operation.
2. A proposal arrives: a twelve-month platform purchase, a data-science team, a proof of concept at month nine. As sponsor, what do you ask for first?
- A. The decision it must improve, what that costs you today, and who owns the resulting action
- B. A comparison of three platform vendors
- C. A larger budget, since under-funding kills these programmes
- D. An earlier board demonstration, so momentum shows sooner
A. The decision it must improve, what that costs you today, and who owns the resulting action — A technology question dressed as a value question. Comparing vendors feels rigorous and is the trap: it accepts the framing that choosing a platform is the decision.
3. Which statement is true of the six foundational shifts?
- A. They are sequenced, so vendor independence comes first
- B. None needs exotic technology; all need management courage
- C. They are upgrades a capable IT function can be given
- D. They apply to operators but not outside energy
B. None needs exotic technology; all need management courage — Each shift changes how the business operates, not what it owns. Delegation is the tempting answer and the failure mode: IT cannot overrule a senior colleague's silo.
Today’s Assignment
Write a short executive position on where your own organisation currently classifies AI, with evidence, and name the one decision you would most want made better.
- The classification, with evidence. Which drawer does your organisation file AI in — IT project, innovation experiment, or operating model? Prove it: where the budget sits, whose name is on it, which committee it reports to, what the last board paper asked for.
- The one decision. Name one recurring decision you would most want made better — a decision, not a system. Who makes it, on what information, how long from first signal to action, and what a late or wrong call costs. Label estimates as estimates.
- The honest part. One paragraph on what your own organisation gets wrong about AI. Not the industry. If nobody senior has taken a position, write that: it is a finding, not an embarrassment.
- The shift you are furthest from. Pick one of the six, say why in three sentences, and name who would have to lose an argument for it to change. Send it as Section 1, in four hundred to nine hundred words.
YOUR CHARTER
Section 1 — Where we are today: the honest classification and the one decision that matters most.
How this is marked — the rubric
| Names one classification and supports it with three pieces of organisational evidence, not intentions | 30 |
| Names one recurring decision, its decision-maker by role, and an elapsed time or cost figure, with estimates labelled | 30 |
| The honest paragraph names a failing in the student's own organisation, not a criticism of the industry | 20 |
| Names one of the six shifts and the role whose position would have to change | 20 |
| Total · pass mark 70 | 100 |
Submitting session 1
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 1, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
Classifying AI correctlyThe right executive questionThe six foundational shiftsBaseline disciplineCharter authorship
The Doctrine · Framework 1SESSION 2
The Operating-Intelligence Loop: Sense, Decide, Act, Learn
Why This Matters
Every executive wants to know where to start. You cannot know until you can see the loop. AI creates value through a loop — sense, decide, act, learn — and it pays only when the loop closes. That is why the world is full of dashboards that describe reality beautifully and change nothing about it. Today you find the exact stage where your own loop breaks, and put a number on the break. It is a diagnostic rather than a theory, and it is what will let you defend a funding decision.
25-Minute Module
0:00 – 5:00
Value arrives in a loop, not a moment
- Strip the vocabulary away. The organisation notices something. Someone decides what it means. Someone acts, with a deadline. And someone measures whether the action worked.
- Those four are sense, decide, act and learn — the Operating-Intelligence Loop. It is the unit of value in this doctrine. Not the model. Not the platform.
- THE DOCTRINE
AI creates value through a loop, not a moment. Value appears only when the loop closes.
A smoke alarm in an empty building has sensed perfectly and produced nothing. The value was never in the beep.
5:00 – 11:00
The four stages, and what each looks like when it is real
Stage What it is What it looks like when it is real Sense Signals from sensors, reports, records The signal arrives while the window to act is open Decide Insight ranked into an owned decision One ranked recommendation, not a list of findings Act A named owner, a deadline, a workflow The action lands in someone's day, with a date Learn The outcome measured and fed back Next quarter's decision is measurably better - The decide row is the one most programmes fudge. An insight is not a decision. A ranked list of nine anomalies is not a decision. A decision has a chosen option and an owner.
- THE DOCTRINE
An insight with no owner is not intelligence. It is an interesting finding, and interesting findings do not move a number.
Sensing is the easiest stage to buy and to demonstrate, and the only one that produces nothing on its own.
11:00 – 17:00
Four ways a loop breaks
Break pattern Where it breaks What you see Blind Sense No signal, or one that arrives too late to act on Analytical Decide Excellent analysis, many dashboards, no chosen option Orphaned Act A decision is taken and nobody owns the action Amnesiac Learn Actions happen, nothing is measured, nothing improves - ENERGY · THE WORKED CASE
- A vibration trend on a critical compressor is captured hourly and reviewed weekly. The work order it raises is not ranked against production impact, so it sits behind three routine jobs. The unit trips, and a modest repair becomes a deferment.
- Sensing worked. Deciding and acting broke.
OUTSIDE ENERGY · THE PARALLEL- Logistics and shipping. A carrier's system flags a vessel that will miss its berthing window. Operations see the exception, but the customer-commitment decision sits with commercial, who hear it at the morning call.
- Same break, different building. The signal was early and the owner was absent.
- In both cases the fix is not a better model. It is a named owner and a ranking rule.
A blind loop and an orphaned loop need completely different money. Spending one on the other is how budgets die quietly.
17:00 – 21:00
Decision latency: the number everything rides on
- Decision latency is the elapsed time from a signal existing to a validated action being taken on it. Not the time to produce a report. The time to act.
- It is the one number that captures all four stages. A faster model that does not reduce it has improved nothing you can bank.
- THE DOCTRINE
Measure decision latency: the elapsed time from signal to validated action. Everything else in an AI programme is a means to that end.
Take the last five occurrences and report the range, never the average. Averages hide the events that cost you money.
21:00 – 25:00
Before you close the laptop
- Measuring latency takes an afternoon: the timestamp when the signal first existed in any system — not when it was reported — and the timestamp when the action was authorised.
- Choose the loop you will map tonight. One that already costs money and that you personally hear complaints about. Not something merely interesting.
- Section 2 of your charter is Our broken loop — the stage that fails, and what it costs. Most leadership teams believe they have a sensing problem and in fact have an acting problem.
You are not looking for the loop with the biggest number, but the one you can close first and prove.
Key Terms
Operating-Intelligence Loop
The four-stage cycle by which an organisation turns signals into value: sense, decide, act, learn. Value appears only when the loop closes.
Decision latency
The elapsed time from a signal existing to a validated action being taken on it. The single variable an AI programme should be judged on.
Loop closure
The point at which an action's outcome is measured against a baseline and fed back, so the next decision is better.
Orphaned insight
A finding, alert or recommendation no named person is accountable for acting on. Operational noise, however accurate it is.
Self-Check — Before You Submit
1. An operator has excellent condition data, a well-attended weekly analytics review, and interventions that genuinely happen — but never measures whether they reduce downtime. Which break pattern is this?
- A. Blind
- B. Amnesiac
- C. Analytical
- D. Orphaned
B. Amnesiac — Sensing works, decisions are taken and actions happen; the learn stage is missing, so nothing compounds. Analytical is tempting given all the analysis, but analytical means no option is ever chosen.
2. Loop A: the signal exists within an hour and the decision is clear, but the action is unowned and slips for weeks. Loop B: nobody has a signal until a monthly report. Both cost the same to fix. Which do you fund first?
- A. Loop B, because sensing is foundational and must come first
- B. Both at once, so the programme shows breadth
- C. Neither, until an enterprise data platform exists
- D. Loop A, because an owner and a deadline are cheap, fast and provable, and they close a loop you already have
D. Loop A, because an owner and a deadline are cheap, fast and provable, and they close a loop you already have — Loop A is orphaned: the expensive stages already work and the missing piece costs a memo, not a platform. Loop B is tempting, but a new signal in an organisation that ignores its existing ones just produces a second unowned alert.
3. What exactly does decision latency measure?
- A. The elapsed time from a signal existing in any system to a validated action being taken on it
- B. The time a model takes to return a prediction
- C. The gap between the monthly report and the meeting on it
- D. The average age of the data in the dashboard
A. The elapsed time from a signal existing in any system to a validated action being taken on it — Latency runs end to end, from when the signal existed rather than when it was reported. Model response time is the tempting answer and is almost never the binding constraint: the delay sits in review and approval.
Today’s Assignment
Map one real loop in your own organisation end to end, name the stage where it breaks, and cost the break in your own units.
- Describe the loop in four short paragraphs, one per stage, naming the system or person that performs each. Where a stage does not happen at all, say so plainly rather than describing what ought to happen.
- Name the break. Which of the four patterns does your loop show, and what single piece of evidence proves it? One pattern, not three.
- Measure the latency. For the last five occurrences: the date the signal first existed in any system, and the date the action was authorised. Report the range and the worst case. If the timestamps do not exist, name the system that should hold them.
- Cost the break in your own units — barrels, cases, claims, shipments, customers, downtime hours, naira. Show the arithmetic in one line and label every assumption.
- The honest part. One paragraph: which stage your leadership team believes is broken, and which stage actually is. Send it as Section 2, in four hundred to seven hundred words.
YOUR CHARTER
Section 2 — Our broken loop: the stage that fails, and what it costs.
How this is marked — the rubric
| All four stages described for one real loop in the student's own organisation, with each performer named and missing stages stated as missing | 30 |
| A single break pattern named and supported by one specific piece of evidence | 20 |
| Decision latency reported across at least five occurrences with a worst case, or absent timestamps identified by system | 25 |
| The cost of the break given in the organisation's own units, with arithmetic shown and assumptions labelled | 25 |
| Total · pass mark 70 | 100 |
Submitting session 2
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 2, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
Loop diagnosisDecision-latency measurementOwnership of actionCosting a breakEvidence over opinion
The Doctrine · Framework 2SESSION 3
Garbage In, Confident Garbage Out: The Data Foundation
Why This Matters
Every executive has heard garbage in, garbage out. The version costing money today is worse, and most boards have not noticed the upgrade. A current model fed poor data does not fail loudly. It returns a clean, fluent, well-formatted answer that is wrong, and sounds certain doing so. Where decisions carry consequences, that is worse than a blank screen. This session gives you the instrument that fixes it — the data contract — and asks you to sign exactly one. It is the least glamorous session of the week and the one with the longest shadow.
25-Minute Module
0:00 – 5:00
Confident garbage
- The old failure mode was bad data in, obvious nonsense out, and somebody catches it. The new one is bad data in and a fluent, well-formatted answer out — wrong, and certain.
- Modern systems hide the fault. There is no ragged edge to warn a reader — no missing field, no silly figure. The wrong output looks exactly like the output that is right.
- THE DOCTRINE
A model fed untrustworthy data does not fail loudly. It becomes confidently wrong — which, where decisions have consequences, is more dangerous than a blank screen.
If a recommendation reached your inbox tomorrow, what would stop you acting on a wrong figure in it? Usually only the seniority of whoever sent it.
5:00 – 11:00
The precondition, and the trap inside it
- Two mistakes are available here and they are opposites. The first is to build models on data nobody owns. The second is to announce that nothing can begin until the data is perfect.
- Both are fatal, and the second is more respectable, which is why it kills more programmes. Waiting for perfect data is a decision to do nothing, taken in a tone of prudence.
- THE DOCTRINE
Perfection is not required to begin. Build a trustworthy enough foundation around the decisions that matter most, and expand from there.
Never build the enterprise data lake first. It is a two-year programme with no operating result, and a finance director will be right to defund it.
11:00 – 17:00
The data contract: the unit of executive work
- For each critical data domain you sign a data contract with five parts. It fits on one page and it is a management document, not a technical one.
Part What it answers A weak answer Named owner Who is accountable? 'Operations' — a function, not a person Standard definition What does this mean, everywhere? Two teams, different meanings Validation rule What is checked, and where? Checked later, by hand, in a spreadsheet Quality threshold What level is acceptable, and who is told? A target with no alert Reporting standard What form, how often, to whom? On request - Contract the domains your consequential decisions run on — production, maintenance, equipment, procurement, contract, compliance, cost, workforce, client or tender. You will contract one, properly, not twelve.
The word contract is deliberate. A contract has a counterparty who can be held to it, which is exactly what a data policy lacks.
17:00 – 21:00
Data archaeology, and the three questions
- On old assets and old businesses the documentation does not match what is there. So the first work is not modelling but data archaeology — reconstructing an accurate record of what exists before anything is trusted on it.
- ENERGY · THE WORKED CASE
- On a brownfield asset that has changed hands twice, two teams report downtime on different clocks and deferment against different reference rates. Both figures are defensible, and they cannot be added together.
- Until the definition is one definition with one owner, every portfolio number built on it is an opinion.
OUTSIDE ENERGY · THE PARALLEL- Healthcare delivery. A hospital group cannot agree what a readmission is: one site counts thirty days from discharge, another from closure of the episode, a third excludes planned returns. The board reviews the average of three definitions.
- Nobody lied. There was simply no contract.
- THE DOCTRINE
Three questions a fluent executive can ask and expect real answers to: show me our model inventory; who owns this data; where did this reported number come from.
Ask those three this week and time the answers. If the answer to 'who owns this data' is a department, you have your finding.
21:00 – 25:00
Before you close the laptop
- Write down the five parts of a data contract. You will be asked for them again — by me, and eventually by an auditor.
- Then pick the one data domain your Session 2 decision depends on. If you are choosing between three, the decision is not specific enough: narrow it. Section 3 of your charter is The one data contract we will sign first.
- Pop Quiz 1 — The Precision Paper opens after this session: five questions across the loop, the classification error and the data-foundation precondition.
Choose an owner who will be annoyed by the contract. An owner comfortable with it has not understood what they are signing.
Key Terms
Data contract
A one-page agreement covering a single data domain: a named owner, a standard definition, a validation rule, a quality threshold and a reporting standard.
Critical data domain
A category of data that consequential decisions depend on — production, maintenance, procurement, contract, compliance, cost. The unit you contract, one at a time.
Data archaeology
Reconstructing an accurate record of what actually exists on an old asset or inside an old business, because the documentation no longer matches reality.
Model inventory
A list of every model in use: what it does, what data it uses, who owns it. If you cannot be shown one, you are not governing models.
Self-Check — Before You Submit
1. Why is confident garbage out more dangerous than an obviously broken output?
- A. It costs more to compute for the same result
- B. It is harder for a vendor to fix under warranty
- C. The wrong answer is clean, fluent and well formatted, so nothing in its appearance prompts a challenge
- D. Modern models are wrong more often than older methods
C. The wrong answer is clean, fluent and well formatted, so nothing in its appearance prompts a challenge — The danger lies in presentation, not frequency. A ragged output gets challenged; a fluent one that looks exactly like a correct answer gets acted on. Lower accuracy is the tempting answer and misreads the point.
2. Your team proposes an eighteen-month data lake, arguing that partial data produces unreliable models. Your priority decision depends on six fields in two systems. What do you do?
- A. Fund a data contract for the domain those six fields sit in, with a named owner and a deadline, and hold the lake
- B. Approve the lake; the team is right about partial data
- C. Approve the lake but halve its budget and timeline
- D. Reject both and revisit AI when those systems are replaced
A. Fund a data contract for the domain those six fields sit in, with a named owner and a deadline, and hold the lake — Build a trustworthy enough foundation around the decision that matters most, then expand. The team's argument is the respectable trap: true in general, and it buys two years with no operating result.
3. Which of these is a complete data contract?
- A. A named owner and a standard definition
- B. A quality threshold, a dashboard and a monthly review
- C. A standard definition, a validation rule and a dictionary entry
- D. A named owner, a standard definition, a validation rule, a quality threshold and a reporting standard
D. A named owner, a standard definition, a validation rule, a quality threshold and a reporting standard — All five parts are load-bearing, and the two most often missing are the ones with teeth: the threshold that says who is told when it is breached, and the named individual accountable.
Today’s Assignment
Write one data contract for the single data domain your priority decision depends on, and name the person who will own it.
- Name the decision and the domain, one sentence each. If more than one domain is genuinely required, choose the one whose failure would hurt the decision most, and say why.
- Write the contract — all five parts, on one page. Owner: a person and a role. Definition: written out in full, in the words you would put in front of two teams that disagree. Validation rule: what is checked and where. Threshold: the level, who is alerted, what follows. Reporting standard: form, cadence, recipient.
- Run the three questions in your own organisation this week — model inventory, data ownership, the provenance of one reported number — and report what you were told, and how long each answer took.
- The honest part. One paragraph on the definition your organisation does not actually agree on. Name the word, and the two functions that use it differently. Send it as Section 3, in four hundred to seven hundred words plus the contract.
YOUR CHARTER
Section 3 — The one data contract we will sign first.
How this is marked — the rubric
| The contract contains all five parts, with the owner named as an individual and a role rather than a function | 30 |
| The standard definition is written out in full and is precise enough to settle a disagreement between two teams | 25 |
| The quality threshold states a level, a named person who is alerted, and the consequence of a breach | 20 |
| The three questions were asked inside the student's own organisation, with answers quoted and elapsed times reported | 15 |
| The honest paragraph names one contested definition and the two functions that use it differently | 10 |
| Total · pass mark 70 | 100 |
Submitting session 3
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 3, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.ASSESSMENT NOW OPEN
Pop Quiz 1 — The Precision Paper opens now. Sit it before you start the next session — it is how you find out whether the doctrine landed, and the practice papers may be resat as often as you like.
Sit Pop Quiz 1 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Data-foundation preconditionWriting a data contractNamed data ownershipDefinition disciplineExecutive questioning
The Doctrine · Framework 3SESSION 4
The Autonomy Ladder: Inform, Advise, Act, Bounded Autonomy
Why This Matters
This is where careful leaders get hurt over the next three years — not by being too cautious, but by letting autonomy outrun their ability to govern it. AI does not arrive as one capability. It arrives as a ladder: systems that inform, systems that draft and suggest, systems that act on named systems with every consequential action approved, and systems granted bounded autonomy inside a narrow, reversible scope. Each rung touches different things and each must be earned separately. Today you learn the four-question test that decides which rung a system has earned, and you place four real candidates from your own business on the ladder.
25-Minute Module
0:00 – 5:00
Autonomy is a ladder, not a switch
- The dangerous assumption is that autonomy is binary — a workflow is either run by people or run by the machine. It is not. Autonomy is a spectrum, traversed gradually and deliberately.
- Four rungs, in plain language first. A system that shows you something. A system that drafts for you to approve. A system that takes action on named systems, with each consequential action approved. And a system that acts inside a narrow, reversible scope you have set, where you review outcomes rather than each step.
- Those are inform, advise (the copilot), act (the supervised agent) and bounded autonomy. The words matter, because a vendor will use the word agent for all four.
Establish which rung is being proposed before anything else. Most disagreements in an AI meeting are two people discussing different rungs.
5:00 – 11:00
The four rungs, and what each may touch
Rung What it may touch Who approves Inform Read-only data Nobody — the human does the acting Advise — copilot Drafts, not systems of record The human, on every output Act — supervised agent Named systems on an allow-list The human, on each consequential action Bounded autonomy A narrow, reversible, instrumented scope The human reviews outcomes, not steps - An agent's power comes from its tools and permissions — what it can read and act upon. So governing an agent means governing its scope, its tools and its approvals, not its intelligence.
- THE DOCTRINE
A system may be trusted to draft a report long before it is trusted to change a setpoint, approve a payment or send a customer a commitment.
Autonomy is earned per workflow, never per vendor. Earning advise on drafting a bid does not earn act on procurement.
11:00 – 17:00
The four-question test
- What does it touch? Name systems and data, not categories. 'The ERP' is not an answer; 'the requisition table, read and write' is. What happens when it is wrong? Not if — when. The worst outcome, and who bears it.
- Who approves? A named role, and whether they approve each action or review outcomes. 'The system is trusted' is not an answer. What is the undo? Tested by whom, taking how long. An untested undo is a hope.
- THE DOCTRINE
Autonomy granted faster than it is governed is not innovation. It is industrialised risk.
Short of time, ask the undo question first. An organisation that cannot describe the undo has rarely thought hard about the other three.
17:00 – 21:00
Where you actually are, not where you believe you are
- Three postures get confused. Human-in-the-loop: a person approves each consequential action. Human-on-the-loop: the system acts within limits and a person reviews outcomes. Human-out-of-the-loop: nobody is in the path. Arriving at one by accident is what is not legitimate.
- ENERGY · THE WORKED CASE
- A reliability model ranks the maintenance backlog by production impact, and most weeks the ranking is accepted without challenge. Nobody signed anything; it simply became the plan. The same model is nowhere near allowed to change a setpoint.
- One workflow has drifted from advise to act. Only the other fact is written down.
OUTSIDE ENERGY · THE PARALLEL- Manufacturing. A scheduling system recommends despatch sequencing and the shift supervisor releases whatever it proposes, because the queue is long and it is usually right. On the line, no model touches a setpoint without an engineer.
- Both plants have an undocumented rung. The risk is not the model. It is that nobody chose.
- THE DOCTRINE
Human-in-the-loop, human-on-the-loop and human-out-of-the-loop are deliberate choices. If you did not choose, procurement chose for you.
Most organisations sit a rung higher in practice than on paper, and a rung lower than in ambition. Both gaps are yours to close.
21:00 – 25:00
Before you close the laptop
- Above the top rung sits a ceiling. For physical, safety-critical or irreversible decisions a human stays in command — not as a limit awaiting better technology, but because accountability for those consequences must remain human.
- Write down the four rungs and the four questions, then list four real or candidate AI uses in your own business — including anything staff already use without a decision having been taken.
- Section 4 of your charter is Our autonomy ladder — what each system is allowed to touch: the first document in your organisation stating what a system may and may not do.
Draw the line yourself, in writing, before an incident draws it for you. Autonomy is earned — make your organisation earn every rung.
Key Terms
Autonomy ladder
The rungs of increasing AI autonomy — inform, advise, act, bounded autonomy — each allowed to touch different things, and each earned per workflow.
Copilot
A system that drafts, suggests, ranks or summarises for a person who remains the actor and the judge. The lowest-risk rung, where most present value sits.
Bounded autonomy
A system acting within strict, predefined limits on lower-risk, reversible work, with a human reviewing outcomes rather than each step.
Human-in-command ceiling
The deliberate, permanent boundary above which a human must decide: physical, safety-critical and irreversible decisions.
Self-Check — Before You Submit
1. What primarily determines how risky an AI agent is?
- A. The size and sophistication of the underlying model
- B. The tools and permissions it holds — what it can read and act upon
- C. Whether it was built in-house or bought
- D. The volume of data it was trained on
B. The tools and permissions it holds — what it can read and act upon — An agent's power, and therefore its risk, comes from what it can touch. Sophistication is tempting because it is what vendors sell on, but a very capable model that only returns a draft is low risk.
2. A system has drafted compliance submissions under human review for six months with a good record. The team proposes it now file them directly, saving two days a cycle. What is the right response?
- A. Approve it — six months of performance earns the next rung
- B. Refuse permanently; filings must never be automated
- C. Approve it for a quarter and review if something goes wrong
- D. Hold it at advise until you can state what a wrong filing costs, who approves, and what the tested undo is
D. Hold it at advise until you can state what a wrong filing costs, who approves, and what the tested undo is — Performance on the current rung is necessary but not sufficient. The decisive question is the undo: a filed submission is not cheaply reversible. A good record on drafting is evidence about drafting.
3. Human-on-the-loop describes which arrangement?
- A. A person approves every consequential action before it executes
- B. No person is in the decision path at all
- C. The system acts within defined limits and a person reviews outcomes rather than each step
- D. A person is consulted when the system flags low confidence
C. The system acts within defined limits and a person reviews outcomes rather than each step — On-the-loop means outcome review inside set limits, which is what bounded autonomy requires. The first option is human-in-the-loop: per-action approval needs a gate, outcome review needs limits, logging and a tested undo.
Today’s Assignment
Place four real or candidate AI uses from your own business on the autonomy ladder, state the rung each has earned, and name the control that would let one climb safely.
- List four uses that exist or are genuinely proposed in your own organisation, one sentence each. Include at least one that staff already use without a formal decision.
- Run the four questions on each — what it touches (named systems and data, read or write), the worst outcome when it is wrong and who bears it, who approves and on what basis, and what the tested undo is.
- State two rungs per use: the rung it operates on today, and the rung you judge it has earned. Where they differ, say which direction the gap runs and what you will do this month.
- Design one climb. Name the single control that would let one use move up a rung safely — an approval gate, a scope or rate limit, an audit log, a tested undo — and who owns it.
- The honest part. One paragraph on the thing already operating a rung higher than anyone authorised, and how it got there. Send it as Section 4, in four hundred to eight hundred words.
YOUR CHARTER
Section 4 — Our autonomy ladder: what each system is allowed to touch.
How this is marked — the rubric
| Four real uses from the student's own organisation, including one in informal use, each with what it touches named as specific systems | 25 |
| All four test questions answered for every use, with the wrong-answer consequence and the party who bears it stated concretely | 30 |
| Both the current rung and the earned rung stated for each use, with the direction of any gap identified | 25 |
| One named control, with a named owner, that would let one use climb exactly one rung safely | 20 |
| Total · pass mark 70 | 100 |
Submitting session 4
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 4, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.Capabilities Covered
Autonomy-ladder placementConsequence and reversibilityApproval designUndo disciplineGoverning scope, not intelligence
The Doctrine · Framework 4SESSION 5
The Value–Risk Map: Fund What Pays, Govern What Bites
Why This Matters
This is the framework you will use most often, and the one that makes you useful in a room full of enthusiasm. Every AI candidate sits on a map with two axes: how much value it creates, and how much damage it does when it is wrong. Read both and the funding decision very nearly makes itself. Read only the first and you will approve the exciting thing that bites you. Today you plot six real candidates from your own business, and defend a decision to park one of them. That second half is the harder one.
25-Minute Module
0:00 – 5:00
Telling candidates apart
- Not all AI applications are equal, and the most important executive act is to tell them apart — to distinguish between the dozen candidates on the list in front of you.
- Two axes, and only two. How much value does this create, measured in something your business already counts? And how much damage does it do when it is wrong?
- The second axis is the one left off the slide. It also decides how much governance the candidate needs, and therefore what it really costs.
A proposal that arrives with one axis on it is a pitch. With two axes, it is a decision.
5:00 – 11:00
Asymmetry of consequence
- A wrong recommendation in a consumer app produces a poor playlist. One at a wellhead, in a credit decision, in a dosage or on a payment run produces a safety event, a loss or a regulatory breach. That gap has a name: asymmetry of consequence.
- THE DOCTRINE
Asymmetry of consequence demands asymmetry of control. Identical technology in two parts of your business warrants entirely different governance.
- ENERGY · THE WORKED CASE
- A model that drafts the weekly production narrative and a model that ranks wells for intervention use similar techniques. One wastes an editor's hour. The other sends a rig to the wrong candidate, and rig time is the scarcest thing on the asset.
- Fund both. Govern one of them like a capital decision.
OUTSIDE ENERGY · THE PARALLEL- Telecoms. A churn model that mis-scores a retention offer costs a discount. A fraud model with authority to bar lines mis-scores an enterprise account and takes a hospital off the network on a Friday night.
- The second needs a human gate and a tested undo. The first needs a baseline and a marketing manager.
The question hype avoids is not 'can we?' but: what does it cost us when it fails, and is the value worth that exposure?
11:00 – 17:00
Four quadrants, used as a funding instrument
- Plot the candidates and four groups appear. Treat them as four funding behaviours, not four labels.
Quadrant What it means What you do Quick wins Real value, low consequence when wrong Fund now; bank the credibility Foundations Modest value, others depend on them Fund to unlock what they enable Big bets High value, high consequence when wrong Govern heavily and stage-gate the money Park Low value, high consequence, however exciting Do not start here. Record why you parked it - Foundations get cut because they have no business case of their own, and the big bets that needed them then fail for reasons nobody traces back. Park items get started anyway, because they are exciting.
- THE DOCTRINE
Start in quick wins. Fund foundations to unlock big bets. Never start in park, and nothing scales on a promise.
A portfolio with nothing in park is not a portfolio. It is a wish list with an approval stamp on it.
17:00 – 21:00
Defending a park decision
- Parking something is the hardest act here, because you are refusing a thing with a champion, a demonstration and often a board member's interest.
- Do it on the record, in four lines. What it is. The value, in your own units. What it costs when it is wrong, and who bears that cost. And the condition for revisiting it — a control that does not yet exist, a data domain that is not contracted, a rung it has not earned.
- That fourth line turns a refusal into a decision. No invites a rematch. Not until we have a tested undo and a named data owner gives the champion something to build.
Sometimes the exciting thing is genuinely a big bet rather than a park item. Value is the axis you must not flatter, in either direction.
21:00 – 25:00
Week 1 closes
- Four frameworks, one reflex. The loop shows where value leaks. The data foundation shows what must be trustworthy. The autonomy ladder shows how much a system may touch. The value–risk map shows which candidates get money and which get governance.
- Run them in that order on whatever anyone brings you. Which stage of the loop does this close? What data does it depend on, and who owns it? What rung does it need, and has it earned it? Which quadrant is it in? A proposal that survives all four deserves your time.
- Section 5 of your charter is Our value–risk map and the thing we will deliberately not do. That second clause is the point. Practice Exam 1 — The Inflection then opens: fifteen questions across Sessions 1 to 5, pass mark 70.
Twenty-five minutes a day, about your own business. The people who finish are not the cleverest; they are the ones who keep the appointment.
Key Terms
Value–risk map
A two-axis map placing every AI candidate by the value it creates and the damage it does when it is wrong. What turns a list of ideas into a decision.
Asymmetry of consequence
An error in one part of a business costs far more, and is harder to reverse, than the same error elsewhere — so identical technology needs different control.
Stage gate
A funding checkpoint at which a candidate must show evidence against a baseline before the next money is released.
Park
A deliberate, recorded decision not to start a candidate, because its value is low relative to what it costs when it is wrong.
Self-Check — Before You Submit
1. A candidate creates high value and does serious damage when it is wrong. Which quadrant, and what does that imply?
- A. Quick win — fund it immediately to build momentum
- B. Park — the exposure outweighs any value
- C. Foundation — fund it to unlock the portfolio
- D. Big bet — fund it, but govern it heavily and stage-gate the money against evidence
D. Big bet — fund it, but govern it heavily and stage-gate the money against evidence — High value plus high consequence is a big bet: worth doing, and not worth doing carelessly, which is what stage-gating is for. Park is tempting for anyone who has been burned, but it hands the advantage to a competitor willing to govern it properly.
2. Your team has fallen for a candidate with a superb demonstration. It saves a little manual effort, and a wrong output would commit the company contractually to a customer. One of its advocates is a board member. What do you do?
- A. Approve a small pilot to keep goodwill, and add governance later
- B. Park it on the record, in the four lines, with the condition for revisiting it
- C. Refuse it verbally and move to the next item
- D. Reclassify it as a foundation, to build confidence with AI
B. Park it on the record, in the four lines, with the condition for revisiting it — Low value against a high, hard-to-reverse consequence is a park item, and the written record with a revisit condition is what makes the decision survive the next enthusiastic meeting. Governance added later is how a demonstration becomes a commitment you must honour.
3. Why are foundation candidates the ones most often cut from a portfolio?
- A. Their direct return is modest, so they lose to candidates carrying their own business case
- B. They are the most expensive items in the portfolio
- C. They carry the highest consequence when wrong
- D. Regulators rarely accept them as compliant investments
A. Their direct return is modest, so they lose to candidates carrying their own business case — Foundations are judged by what they make possible, not by their own return, so a review scoring every candidate on direct payback cuts them every time. Cost is the tempting answer; the problem is the scoring basis.
Today’s Assignment
Plot at least six AI candidates from your own organisation on the value–risk map, name each quadrant, and defend one decision to park something the organisation is excited about.
- List at least six real candidates — proposed, piloted, purchased or wished for in your own business. Not a generic industry list. One sentence each.
- Score both axes. Value: in your own units — barrels, cases, claims, shipments, customers, hours, naira — against the baseline you would measure from, naming its source or stating that it does not exist. Consequence: the worst outcome when it is wrong, who bears it, and whether it is reversible.
- Assign one quadrant each — quick win, foundation, big bet or park — with a one-line reason. A candidate you want in two quadrants means the axes are not yet scored honestly.
- Name your first quick win and your first foundation, with the owner by role for each and what you would measure at ninety days.
- Defend one park decision in four lines, on something your organisation is genuinely excited about: what it is, the value, the cost when it is wrong and who bears it, and the condition for revisiting it. Send it as Section 5, in five hundred to nine hundred words.
YOUR CHARTER
Section 5 — Our value–risk map and the thing we will deliberately not do.
How this is marked — the rubric
| At least six real candidates from the student's own organisation, each with a one-sentence description | 20 |
| Value scored in the organisation's own units for every candidate, with the baseline source named or its absence stated | 25 |
| Consequence scored for every candidate, with the worst outcome, who bears it, and reversibility stated | 25 |
| One quadrant assigned per candidate with a one-line reason, and no candidate placed in two | 15 |
| A park decision defended in all four lines, on a candidate the organisation is currently enthusiastic about | 15 |
| Total · pass mark 70 | 100 |
Submitting session 5
Write your answer in Word or Google Docs. The first line must read
Write your answer in Word or Google Docs. The first line must read
Session: 5, or the submission is refused. Upload the Word file (.docx), or switch to the Google Doc tab and paste your document’s share link.Marked against the rubric above and returned as a written letter, saved to your progress page. Do not submit until you have passed the self-check above.ASSESSMENT NOW OPEN
Practice Exam 1 — The Inflection opens now. Sit it before you start the next session — it is how you find out whether the doctrine landed, and the practice papers may be resat as often as you like.
Sit Practice Exam 1 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Value–risk mappingAsymmetry of controlPortfolio quadrantsDefending a park decisionWeek 1 framework recall