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 3 OF 4 · TRUST, CONTROL AND THE BOARD · SESSIONS 11–15 · WORK AT YOUR OWN PACE
Trust and Control · GovernanceSESSION 11
Governance Is Value Protection, Not Paperwork
Why This Matters
You will be asked to approve something you cannot personally verify. That is the governance problem in one sentence. Someone will bring you a model that recommends an intervention, prices a risk or screens a filing, and you cannot check its arithmetic. What you can check is the accountability around it: who owns it, what it may touch, who approved that, what happens when it is wrong, and how you would undo it. This session gives you those controls, and the page you will write to make them a standard in your own organisation.
25-Minute Module
0:00 – 6:00
Trust is engineered, not assumed
- Governance has a poor name in boardrooms because it usually arrives late, as a document, after the programme is running. That version deserves its reputation. Mine is different: governance is what lets you say yes to scaling AI.
- What it produces is operational trust, and no single function produces it. Trust sits where three meet: risk (is the model valid and monitored?), compliance (can we prove we meet our obligations?) and cybersecurity (are the systems secure where they touch anything that moves or pays?).
- THE DOCTRINE
Do not trust AI because it is advanced. Trust it when the accountability around it is strong.
- An ungoverned estate works fine until a model fails consequentially or a number cannot be defended. Until then you are not saving money; you are accumulating an undisclosed liability.
If governance in your business is a folder rather than a control, you do not have governance. You have a filing system.
6:00 – 13:00
What governance must actually cover
- Ask most executives what AI governance covers and they say the models. The models are the easy part. The perimeter also holds the data pipelines, the interfaces to systems that record or act, unvalidated dashboards, the compliance systems, the reports that reach your board, and third-party platforms you did not build. Inside that perimeter, seven questions must have answers, per system.
What governance must cover The question it answers Purpose Which decision does this system exist to improve? Data provenance Where did its data come from, and who owns that data? Approval by consequence Which consequence class is it in, and who approved that? Testing and monitoring How was it validated, and how would we see it drifting? An owner by name Which named person answers for its performance? A documented undo How do we reverse what it did, and who can stop it? A retirement decision When is it revalidated, and when is it retired? - THE DOCTRINE
A policy that is not enforced by a control is a wish. A control without an owner is a gap.
Weak answers give it away: it comes out of the system, IT signed it off. Test the seven on one live model this week.
13:00 – 19:00
Approval calibrated by consequence
- Human approval is the control most often claimed and least often real. A rubber stamp is not a control: ninety-nine per cent approval with almost no rejections means the review is not happening.
- So calibrate approval to consequence rather than one rule everywhere. Low consequence and reversible: the system acts, a human reviews the outcome. Consequential: a qualified human approves each action with the evidence in front of them. Physical, financial or contractual: a human decides every time, and is accountable.
- ENERGY · THE WORKED CASE
- A surveillance model on a Niger Delta flow station flags a well below its expected curve and recommends an intervention. It sits in the inventory, the production engineer named as owner signs the recommendation, the approval is logged, and the model has no path to a setpoint.
- The same model feeding a monthly report is a lighter case and should be governed more lightly. Proportionality is the discipline.
OUTSIDE ENERGY · THE PARALLEL- Public sector. A tax authority’s audit-selection model sits inside the same three lines: the operating directorate owns it, an independent assurance function validates it before and after go-live, internal audit tests whether both did their jobs, and the accounting officer and audit committee set the appetite.
- The mechanism is identical to the flow station. Only the consequence differs, a citizen wrongly selected for audit rather than a deferred barrel, and consequence sets the control.
- The rule beneath both: strictest controls where the consequence is greatest, and nothing consequential ungoverned.
Ask for your recommendation-to-approval rate and your rejection rate. If nobody ever says no, nobody is really reading.
19:00 – 25:00
Three lines of defence, then write it down
- Accountability has to be layered rather than concentrated in one capable individual. The first line, the business that owns and runs the model, must produce the inventory entry, the owner, the validation record and the logs. The second line, risk and compliance and cybersecurity, sets policy, monitors and must have standing to challenge. The third, internal audit, evidences that the controls actually operate. You own the appetite above all three.
- In most organisations the second line for AI does not yet exist. That is itself a priority: you cannot challenge what you do not understand.
- THE DOCTRINE
The first honest question a board should ask is the simplest one: show me the model inventory.
- Before you close the laptop, write the five controls you require before any model goes live, and the role, not the committee, that signs each. Mine are an inventory entry, a named owner, approval calibrated to consequence, a testing-and-monitoring record and a documented undo. That is Section 11 of your charter: our governance standard, five controls and five signatures.
Governance is not the brake on your ambition. It is what lets you press the accelerator and keep your licence.
Key Terms
Model inventory
A live register of every model in use: what it does, what data it uses, what it can touch, who owns it, and when it was last validated.
Approval by consequence
Setting the strength of human approval by what an error would cost rather than one blanket rule. Reversible work gets outcome review; consequential action gets a named decision.
Three lines of defence
Layered accountability: the business owns and runs the model, risk and compliance challenge it, internal audit assures the controls operate. The board owns the risk appetite.
Model drift
The quiet degradation of a model's accuracy as the world it was trained on changes. Monitored by nobody, it becomes confidently wrong rather than obviously broken.
Self-Check — Before You Submit
1. Your operations director proposes a model that recommends maintenance interventions. What would you insist on before it goes live?
- A. A vendor assurance that it has been tested at other companies
- B. A named owner, an inventory entry and a documented undo
- C. A six-month period during which no human approval is required
- D. Board approval of the model's technical architecture
B. A named owner, an inventory entry and a documented undo — Governance is accountability, not reputation. A vendor's experience elsewhere is not your control, and a board cannot usefully approve an architecture it cannot assess. The owner, the entry and the undo let you answer for the model when it is wrong.
2. What is the strongest sign that a human-in-the-loop control has become cosmetic?
- A. A very high approval rate with almost no rejections
- B. Approvals taking longer than the operations team would like
- C. The model being retrained more often than planned
- D. Several approvers named against one decision
A. A very high approval rate with almost no rejections — If nobody ever declines, the review is a formality; approval adherence is a metric to measure, not an assumption. Slow approvals and frequent retraining are irritating but show the control is alive.
3. A commercial function has connected a third-party AI platform to your contract repository without a review, and it works well. What do you do first?
- A. Ban the platform and reissue the procurement policy
- B. Leave it running and review it at the next quarterly meeting
- C. Inventory it, assess it, name an owner and define what it may touch
- D. Ask the vendor for its certifications and accept them as assurance
C. Inventory it, assess it, name an owner and define what it may touch — The un-reviewed third-party connection is a common path to cyber exposure and ungoverned model risk, so it cannot wait a quarter. A ban drives the next one underground, and a vendor certifying itself is not your assurance.
Today’s Assignment
Write the one-page AI governance standard for your own organisation: the five controls every model must satisfy before go-live, and the named role that signs each one.
- The five controls, 250 to 350 words. State each as a requirement that can be tested. 'Every model has an inventory entry recording purpose, data source, owner and last validation date' is testable; 'we take governance seriously' is not.
- Five signatures. Name the role in your business that signs each control off, the role and not the committee. Where two controls share a signature, say why.
- Consequence classes, about 150 words. Define two or three consequence classes for your organisation and the approval each requires, in your own units: a barrel, a payment, a patient, a shipment.
- The honest part, about 150 words. Which control can your organisation not evidence today for a single live system, and what would closing it this quarter take? Then one sentence on where AI governance sits now, whether IT, risk, a committee or nowhere, and where you will move it.
YOUR CHARTER
Section 11 — Our governance standard: five controls, five signatures.
How this is marked — the rubric
| Five controls stated as testable requirements, each with a named signing role rather than a committee | 30 |
| Consequence classes defined in the organisation's own units, with the approval each class requires | 25 |
| Honest gap named against a specific control, with a named owner and a dated commitment | 25 |
| Current location of AI governance stated plainly, with the intended change | 20 |
| Total · pass mark 70 | 100 |
Submitting session 11
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: 11, 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
Governance as value protectionModel inventory disciplineApproval by consequenceThree lines of defenceDecision-rights design
Trust and Control · ComplianceSESSION 12
The Compliance Dividend: Numbers You Can Defend
Why This Matters
Somewhere in your business there is a number you report externally that nobody could reconstruct if it were challenged tomorrow. Not because anyone is dishonest, but because the trail runs through a spreadsheet a departed colleague built, an assumption nobody wrote down, and a system since replaced. That was survivable while regulators accepted estimates. It is not survivable now they ask for measurements. The useful part is that the work which makes a number defensible is the same work that makes a model trustworthy, so you pay once and it returns twice.
25-Minute Module
0:00 – 6:00
Reporting became measurement-based
- For most of your career, external reporting was a compilation exercise: assemble a figure near the deadline, file it, defend it in general terms if asked. Regulation in every serious sector is moving towards measurement, meaning a reported number must be traceable to the thing it claims to measure.
- Four questions now sit behind every externally reported figure. Where did the data come from? Who approved it? How was it calculated? Can it be defended?
- THE DOCTRINE
A reported figure that cannot be traced through its data and its calculation is no longer weak management. It is a finding waiting to happen.
- Most organisations discover their answers during an audit, a partner reconciliation or a due diligence. That is the most expensive moment to find out.
Pick one number your business reports outside its own walls and keep it in mind. You will be tracing it tonight.
6:00 – 12:00
The assurance chain
- Treat a defensible number as a chain with four links: measure the thing credibly, assure it by validation and reconciliation, disclose it traceably, and present it to those who finance or regulate you.
Link What it means What breaks it Measure Captured credibly at source, by an instrument or system you can name Manual re-entry, an uncalibrated instrument, a gap nobody notices Assure Validated and reconciled by someone other than the person who produced it One person owning both the number and its own check Disclose Reported in a standard form, with calculation and assumptions alongside A template filled at the deadline from a spreadsheet nobody can open Finance Withstands a regulator, auditor, lender, partner or board Scrutiny arriving before the chain was built - Every link needs an owner. Where two links share one, you have assurance in name only.
If your sector has no measurement-based mandate yet, your lenders and largest customers will impose one first. They usually do.
12:00 – 19:00
Trace one number, properly
- Here is the exercise I run with executive teams. Take one externally reported figure and walk it backwards, writing each step down: the filing it appears in, the calculation and where its assumptions are recorded, every data source with its system and owner by role, the approvals and the evidence they happened, and the point at which the trail goes cold.
- ENERGY · THE WORKED CASE
- The worked case. Nigeria's upstream regulator, NUPRC, has directed operators towards standardised templates and measurement-based methane and greenhouse-gas reporting. Filing a number is no longer enough: it must be traceable from source measurement, through calculations and assumptions, to the submission.
- That is a lineage problem more than an emissions problem. The operator who can prove its methane figures gains in every audit, reconciliation and lender conversation.
OUTSIDE ENERGY · THE PARALLEL- Other sectors have lived here for years. Reporting under IFRS lives or dies on the audit trail behind each balance. A pharmaceutical plant must produce a batch record of what was made, from what, by whom. A bank must evidence its model-risk work and its consumer-duty decisions.
- If you run a hospital, a factory or a lender, you are already inside this. The only question is whether your evidence is engineered or assembled.
- THE DOCTRINE
Credibility cannot be assembled at the moment capital asks for it. It has to be built into the operating model before.
Do the walk with the person who compiles the number, not the one who signs it. The compiler knows where the joins are.
19:00 – 25:00
The compliance dividend
- This is not a compliance session, and here is why. The controls that make a number defensible, being owned data, standard definitions, validation rules, lineage and logged approvals, are the same controls that make a model trustworthy.
- THE DOCTRINE
The same governance that makes a model trustworthy makes a filing defensible. The investment pays twice: once in performance, once in assurance.
- Then the commercial half, which executives under-argue. Capital prices credibility as well as returns, through a flywheel: trusted data produces defensible numbers, which build lender confidence, which lowers the cost of capital, which funds growth that strengthens the data.
- Honestly, none of this manufactures a sound business. Lineage on a badly measured figure only proves faster that it was badly measured. The dividend goes to those willing to fix the measurement.
- Write the number you chose and the place its trail goes cold. That is Section 12 of your charter: one reported number, traced, and the gaps we must close. No number without a baseline, no gap without a named owner.
When you next argue for a data investment, argue it as a financing investment too. That is the version a chief financial officer funds.
Key Terms
Data lineage
The documented path of a number from its source measurement, through every transformation and assumption, to the report it appears in. Coverage is the share of critical flows you can follow.
Assurance chain
The four links behind a defensible figure: measure, assure, disclose, finance. Engineered end to end so a number can be reconstructed on demand rather than assembled at the deadline.
Measurement-based reporting
A standard in which a reported figure must derive traceably from actual measurement rather than estimation. Batch records and audit trails are older versions of the same demand.
Credibility flywheel
The compounding loop in which trusted data produces defensible numbers, which build lender confidence, which lowers the cost of capital, which funds growth that strengthens the data.
Self-Check — Before You Submit
1. What makes an externally reported figure defensible?
- A. It was produced by an advanced model rather than a spreadsheet
- B. It was approved by a sufficiently senior executive
- C. It is consistent with the figure reported last year
- D. It can be reconstructed on demand from its data, calculation and approvals
D. It can be reconstructed on demand from its data, calculation and approvals — Defensibility is reconstructibility. Seniority of the approver and sophistication of the tool are both compatible with a number nobody can trace, and consistency with a wrong prior figure compounds the error rather than correcting it.
2. Your finance director argues the traceability work can wait a year, because this figure has never once been queried. What is the strongest counter?
- A. Scrutiny is tested suddenly but credibility is built continuously, so the trail must exist before it is demanded
- B. Regulators impose penalties where lineage cannot be produced
- C. Competitors in the sector are already doing the work
- D. The data team happens to have capacity this quarter
A. Scrutiny is tested suddenly but credibility is built continuously, so the trail must exist before it is demanded — The honest argument is about timing, not threat: the trail either exists when the question arrives or it does not. Penalties may not apply, peer activity is not evidence, and spare capacity evaporates when something urgent lands.
3. Which of these best describes the compliance dividend?
- A. Compliance work can be automated so that headcount falls
- B. Regulators reward early adopters with lighter oversight
- C. The governance that makes a model trustworthy also makes a filing defensible
- D. Reporting more frequently improves the quality of the underlying data
C. The governance that makes a model trustworthy also makes a filing defensible — The dividend is the shared foundation: owned data, standard definitions, validation and lineage serve the model and the filing at once. Regulators promise nothing of the sort, and reporting more often on weak data simply repeats the error.
Today’s Assignment
Take one number your own organisation reports externally, trace its lineage as far as you honestly can, and write what is missing to make it reconstructible on demand.
- Name the number, about 100 words. One figure reported outside your own walls, to a regulator, lender, partner, auditor or board. What it is, where it is filed, how often, and who signs it today.
- Trace it backwards, 300 to 400 words. The calculation and where its assumptions are written down; every data source with its system and owner by role; the approvals and the evidence they happened; and the point at which the trail goes cold.
- The gaps, about 150 words. What is missing before this number could be reconstructed on demand. Give each gap a named owner by role and a date.
- The honest part. If you were asked tomorrow to reproduce this figure from source, how long would it take and who would do it? If the answer depends on one person, say so; that is a finding, not a staffing note. Close with what defensibility here would be worth commercially.
YOUR CHARTER
Section 12 — One reported number, traced: the gaps we must close.
How this is marked — the rubric
| Number identified precisely, with the filing it appears in, its frequency and its current signatory | 20 |
| Lineage traced through calculation, assumptions, sources by system and owner, and approvals, with the break point named | 35 |
| Gaps listed with a named owner by role and a date for each, not an intention | 25 |
| Honest assessment of reproduction time, single-person dependency and commercial worth | 20 |
| Total · pass mark 70 | 100 |
Submitting session 12
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: 12, 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
Lineage disciplineDefensible reportingAssurance-chain designCapital credibilityAudit readiness
Trust and Control · SecuritySESSION 13
Cybersecurity, the Operational Boundary and the Shadow-AI Problem
Why This Matters
There are two conversations about AI and security, and boards get them wrong in opposite directions. The first is the boundary: AI reaching out of your analysis and into the systems that move, pay or record things, connected by whoever needed it working that week. The second is the leakage your own people are doing right now, with good intentions, because the tool on their phone is better than the one you gave them. The first is a design problem, the second a management problem, and the answer to both is a governed environment rather than a prohibition. Banning it does not stop it. It makes it invisible.
25-Minute Module
0:00 – 7:00
The boundary, generalised
- Your organisation holds commercial terms, client documents, regulatory records, engineering or clinical files, personal data, and in industrial businesses the systems that control physical equipment. AI touches more of it, faster, through more interfaces than anything you have deployed.
- In oil and gas we call it the operational-technology boundary: the line between enterprise systems and the control systems that move physical things. The rule is absolute. AI must never create uncontrolled bridges across it; connectivity must be intentional, segmented, monitored and governed.
- If you run no physical plant, the general form still applies: any bridge between an analytical system and a system of record or a system of action.
Kind of system What it does What a bridge to it requires System of insight Analyses, ranks, drafts, recommends. Changes nothing itself Governed data and access. Read-only by default System of record Holds the authoritative version: the ledger, the patient file, the customer master A named interface, write permissions justified case by case, every change logged and reversible System of action Moves, pays, despatches, commits or controls Segmentation, human approval on consequential actions, hard limits, monitoring, a tested stop
Most organisations have no inventory of the third row. Ask your security lead which of your AI systems can write, not just read.
7:00 – 13:00
Shadow AI: the exposure nobody logged
- Now the everyday risk, among the most common and most underestimated exposures in business today: staff pasting client data, contracts, drawings, patient or customer records and regulatory submissions into uncontrolled public AI tools. Under deadline, trying to do the job well.
- ENERGY · THE WORKED CASE
- A proposals engineer, forty hours from a tender deadline, pastes a client's scope of work and the commercial section of an old bid into a public tool to tighten the summary. The bid goes out cleaner. The scope and your pricing logic have left your control, and nothing recorded it.
- The same pattern turns up with drawings, regulatory submissions and joint-venture correspondence: the documents hardest to explain losing.
OUTSIDE ENERGY · THE PARALLEL- Healthcare. A registrar pastes a discharge summary into a public assistant to shorten it before the clinic closes. The patient is served faster, and identifiable clinical information has gone to a third party with no agreement and no recall.
- Every sector has its version: a case file, a customer statement, an unannounced result. The intent is good; the exposure is identical.
- THE DOCTRINE
AI without cybersecurity discipline is not innovation. It is exposure with a friendly interface.
- Nobody will volunteer any of this if the standing instruction is a ban, which is exactly why a ban leaves you blind to your own exposure.
The question is which document you would least like to find in a public tool, and how you would ever know.
13:00 – 19:00
The answer is a governed environment
- THE DOCTRINE
Give your people a safe, governed AI environment so the answer to 'can I use AI for this?' is yes within controls. A ban does not end the usage; it moves it out of your sight.
- A governed environment is a decision rather than a purchase: a sanctioned tool with your own data agreement behind it, classified information, written rules on what may go in, logging you can inspect, and training that explains the reasoning.
- The control set runs to eight items, and you should be able to name the owner of each: data classification; access and identity; model and prompt governance; vendor risk; data-loss prevention; audit logs; incident response; and training.
- Then rehearse the incident. The rehearsal usually reveals that nobody knows who declares an AI-related disclosure.
Write the rule as a sentence your busiest manager can repeat from memory. If it needs a policy document, it will be ignored by Thursday.
19:00 – 25:00
Before you close the laptop
- Write two things down: the bridge your organisation will not build, stated as a rule somebody else could apply, and the environment you will give your people instead.
- That is Section 13 of your charter: our safe-use standard and the boundary we will not cross. The shortest section in the charter, and the one most likely to be read by everybody.
- The assignment asks for two pages and an honest number. Sizing the exposure is itself the first piece of security work, because you cannot control what you have never measured.
- Pop Quiz 3 opens after this session and covers Sessions 11 to 13: governance and defensibility.
Security written as prohibition produces evasion. Security written as permission within limits produces compliance. Choose the second, and mean it.
Key Terms
Operational-technology boundary
The line between enterprise systems and the control systems that move physical things. Generally: any bridge between an analytical system and a system of record or of action.
Shadow AI
Unsanctioned use of public AI tools with company, client or personal information. It is driven by a capability gap, and grows fastest where the official answer is a blanket ban.
Data classification
Deciding in advance, and in writing, which categories of information may be put into which systems. It turns 'use your judgement' into a rule staff can follow.
Data-loss prevention
Controls that detect and block sensitive information leaving your environment by email, upload or an AI interface. It makes a safe-use rule enforced rather than published.
Self-Check — Before You Submit
1. Your risk committee proposes blocking all public AI tools on the corporate network. What is the strongest executive response?
- A. Agree, because the leakage risk outweighs the productivity benefit
- B. Restrict the uncontrolled tools only as part of providing a sanctioned environment with written rules
- C. Allow everything, but require staff to report afterwards what they used
- D. Ask legal to draft a policy and revisit next quarter
B. Restrict the uncontrolled tools only as part of providing a sanctioned environment with written rules — A block alone does not remove the demand; it moves it to personal devices where you have no visibility, no logging and no agreement with the provider. Self-reporting is not a control, and a policy with no environment behind it is not a defence.
2. Outside industrial operations, what is the general form of the operational-technology boundary risk?
- A. Staff installing unapproved software on company machines
- B. Storing company data outside your own country
- C. Connecting two enterprise applications to each other
- D. Any bridge between an analytical system and a system of record or a system of action
D. Any bridge between an analytical system and a system of record or a system of action — The risk is not the analysis; it is the path from analysis into something that records or does. Unapproved software and data location are real but different issues, and integration is routine. What matters is whether the bridge can write, pay or control.
3. Which of these is evidence that a safe-use standard is operating rather than merely existing?
- A. Audit logs showing what was submitted, by whom, inside the sanctioned environment
- B. A signed acknowledgement from every member of staff
- C. The standard published on the company intranet
- D. A clause added to the staff handbook
A. Audit logs showing what was submitted, by whom, inside the sanctioned environment — Enforcement is evidence; attestation is not. Signatures, publication and handbook clauses prove the rule exists. Only the logs prove it is followed and let you size the exposure honestly.
Today’s Assignment
Write the two-page safe-use standard for your own organisation, and add an honest estimate of the shadow-AI usage in it today.
- Page one, what may and may not go in, about 300 words. Classify your own information into three or four categories and state which AI environment each may be used in. Use your real categories: client confidential, patient identifiable, price-sensitive, personal data, regulatory submission, drawings.
- The environment. Name the sanctioned environment your people must use, or state plainly that you have none and what you will do this quarter. It is the half most standards omit, and the half that decides whether the rule holds.
- Page two, the boundary, about 200 words. State the bridge your organisation will not build, in one testable sentence, and name the systems on each side: your operational-technology line, or your system of record or of action.
- Consequences and the route back. What happens when the rule is broken, distinguishing an honest mistake reported promptly from concealment. Name who decides, and how someone declares a breach without ending their career.
- The honest number, about 150 words. Estimate what share of your staff put company or client information into uncontrolled public tools, how you reached the estimate, and the three document types most exposed. Zero will be marked as an answer you have not looked for.
YOUR CHARTER
Section 13 — Our safe-use standard and the boundary we will not cross.
How this is marked — the rubric
| Information classified into the organisation's own categories, each mapped to a permitted environment | 25 |
| Sanctioned environment named, or its absence stated with a dated commitment and an owner | 20 |
| Boundary stated as one testable rule, with the systems on each side named | 25 |
| Breach consequences distinguish honest reporting from concealment, with a named decision-maker | 15 |
| Shadow-usage estimate given with its method and the three most exposed document types | 15 |
| Total · pass mark 70 | 100 |
Submitting session 13
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: 13, 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 3 — Governance and Defensibility 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 3 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Boundary designShadow-AI managementData classificationIncident readinessSafe-use standards
Trust and Control · The FrontierSESSION 14
Agentic AI and Controlled Autonomy: The 2027 Question
Why This Matters
You have already placed your systems on the autonomy ladder. This session is about the rung most vendors are now selling and most boards cannot yet interrogate: software that does not answer a question but carries out a task. It is genuinely useful and genuinely risky, and the difference between those outcomes is not the model. It is what you allowed it to touch. By the end you will be able to sit in a vendor meeting and ask the four questions that separate a present opportunity from a premature risk, and you will have written the line your own organisation will not cross.
25-Minute Module
0:00 – 6:00
What an agent actually is
- Strip the marketing away. An AI agent pursues a goal in a loop: it reads its environment, decides what to do next, acts by calling tools and systems, then checks the result and adjusts. What separates it from the copilot your team uses is that it takes multi-step actions through tools. It does things in systems rather than producing text for a person to act on.
- That property is the whole of its value and the whole of its risk. An agent that reads a document library and drafts a response is a productivity tool. An agent that can execute a transaction or touch an operational system is a different class of thing.
- THE DOCTRINE
An agent's power comes from its tools and permissions. Controlling an agent is controlling its tools, its scope and its approvals, not choosing a better model.
When someone says agentic, ask what it can do without asking anyone. That answer, not the demonstration, is the product.
6:00 – 12:00
A dial, not a switch, and a permanent ceiling
- Session 4 gave you the ladder. In practice autonomy is a dial you turn gradually, workflow by workflow, and one position on it you never reach.
Position on the dial Who acts, who decides Where it belongs Realistic timing Copilot The person acts and judges; AI assists Drafting, search, summarising, analysis Now, where most value sits Supervised agent The agent executes; a human approves each consequential action Document-heavy back-office and commercial work Now to two years Bounded autonomy The agent acts within limits; humans review outcomes Lower-risk operational workflows, once controls are proven Two to three years, governance-gated Full autonomy on physical or safety decisions No human in command Nowhere. A permanent boundary Not on any horizon - The placement test is two questions, and it is the central governance decision: what does an error cost, and how hard is it to reverse? Low consequence and reversible is where agents act first; high consequence and hard to reverse stays human-gated, however capable the technology becomes.
Write the ceiling down before anyone asks you to move it. It is a statement about where accountability lives, not about today's technology.
12:00 – 19:00
Guardrails, and the worked cases
- Safety here is engineered, not purchased. Six guardrails, from the bottom up: bounded scope, so it touches only what it is permitted to; evaluation and testing before and during deployment; audit and observability, so every action can be reconstructed; approval gates on consequential actions; kill switches and limits on spend, scope and rate; and human oversight that can inspect, override and stop.
- THE DOCTRINE
Autonomy granted faster than it is governed is not innovation. It is industrialised risk.
- ENERGY · THE WORKED CASE
- Tender assembly is the strongest first candidate in an oil and gas service business: multi-step, document-heavy, deadline-driven and reversible. A supervised agent assembles the compliance matrix, drafts responses from a governed library of past bids, checks completeness and prepares the local-content pack. A human approves before submission.
- What it must never hold is a path to the price, the signature or any operational system. The prize is a faster, cleaner, more complete bid.
OUTSIDE ENERGY · THE PARALLEL- Logistics and shipping. An agent reconciles delivery exceptions across carrier records, assembles the evidence pack for a claim and drafts the customer notification. A human approves the claim and the notification.
- The same lines hold. Rerouting a vessel, releasing a payment or committing to a customer stays with a person who can be held to it.
- An agent inherits data and permissions, and every weakness in them. On ungoverned data it produces confident error at speed.
Give an agent the minimum scope its task requires and nothing more. Scope is the guardrail people grant generously and regret specifically.
19:00 – 25:00
Reliability, the vendor filter, and the line
- Judgement stays human for three reasons. These systems can be confidently wrong, producing fluent output that is incorrect and looks authoritative. They can be manipulated through their inputs, and that risk grows with every tool you connect. And they pattern-match within their training; they do not carry consequence and accountability as an experienced engineer does.
- In a vendor meeting, four questions cut through the hype. Where does this sit on consequence and reversibility? What bounds its scope and gates its actions? What is logged, and what stops it? What data and permissions does it inherit?
- For anything already running, the board asks: what agents do we have, what can each touch, who owns each, where is our ceiling, and can we supervise what we deployed.
- Before you close the laptop, name one process a supervised agent could carry within two years, and write the line you will never let an agent cross, with the reason. That is Section 14 of your charter: where we will allow agency, and the line we will never cross.
Be ambitious about the drudgery and unflinching about the ceiling. Holding both in the same head is the whole skill.
Key Terms
AI agent
A system that pursues a goal in a loop: reading its environment, planning, acting through tools and observing the result. It differs from a copilot in taking multi-step action in systems.
Bounded autonomy
Permission to act within strict, predefined limits on lower-risk work, with humans reviewing outcomes rather than each step. Earned through reliability under supervision, never granted by default.
Guardrail stack
The six controls that contain an agent: bounded scope, evaluation and testing, audit and observability, approval gates, kill switches and limits, and human oversight.
Human-in-command ceiling
The deliberate, permanent line beyond which a human must decide, covering anything physical, financial, contractual or reputational. Not a limitation to be lifted as technology improves.
Self-Check — Before You Submit
1. What distinguishes an AI agent from the copilot your team already uses?
- A. It uses a larger and more recent model
- B. It can be asked questions in plain language
- C. It takes multi-step actions through tools rather than producing output for a person to act on
- D. It runs continuously rather than only when opened
C. It takes multi-step actions through tools rather than producing output for a person to act on — The defining property is action through tools, and it is the source of both the value and the risk. A larger model that still only drafts text is a copilot, and a scheduled report that changes nothing is not an agent.
2. A vendor offers an agent that will monitor your plant and adjust equipment settings automatically, with a monthly human review of what it did. What is your position?
- A. Approve a limited trial on non-critical equipment with the monthly review as proposed
- B. Approve it if the vendor accepts contractual liability for any incident
- C. Commission a technical evaluation and decide on its results
- D. Decline the acting part: physical settings stay human-decided, and take the monitoring instead
D. Decline the acting part: physical settings stay human-decided, and take the monitoring instead — This crosses the permanent ceiling. The consequences are physical and hard to reverse, so a monthly review is assurance arriving after the harm, a liability clause does not undo an incident, and an evaluation answers capability when the question is consequence.
3. Where should an organisation deploy its first supervised agent?
- A. Wherever the largest cost saving is available
- B. In high-volume, document-heavy work where errors are cheap and reversible
- C. In the function with the strongest technical team
- D. In the process the chief executive finds most frustrating
B. In high-volume, document-heavy work where errors are cheap and reversible — Consequence and reversibility set the entry point, not the size of the prize or the enthusiasm of the sponsor. The largest saving often sits in exactly the consequential work that should stay gated.
Today’s Assignment
Choose one process in your own organisation that a supervised agent could run within two years, write the guardrails it would require, and state the one thing you will never let an agent do in your business.
- The process, about 200 words. Name it, who does it now, how often it runs each month and how long it takes. Then justify the choice against consequence and reversibility: what does an error cost, and how is it reversed?
- The guardrails, about 300 words. All six for this process: the scope and tools it may touch; how you would test it before and monitor it after; what is logged; which actions need approval and by which role; the limits and the stop; and who oversees it daily.
- The permissions it inherits. List the data and system access it would need, and what you would refuse it. Minimum scope is the discipline, so write down what you withhold.
- The line, about 150 words. One thing you will never let an agent do in your organisation, with the reason, stated as a rule somebody could apply in your absence.
- The honest part. Which foundation does your organisation not yet have, whether governed data, an access model, an inventory, audit or supervising capability? Name your first step and who takes it.
YOUR CHARTER
Section 14 — Where we will allow agency, and the line we will never cross.
How this is marked — the rubric
| Process chosen and defended against consequence and reversibility, with its volume and effort stated | 25 |
| All six guardrails specified for that process, with approval roles named rather than described | 30 |
| Permissions listed, including what is deliberately withheld | 15 |
| The line stated as one applicable rule, with a defended reason | 15 |
| Missing foundation named honestly, with a first step and a named owner | 15 |
| Total · pass mark 70 | 100 |
Submitting session 14
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: 14, 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
Agentic literacyAutonomy calibrationGuardrail designVendor interrogationFrontier governance
Trust and Control · The BoardSESSION 15
The Board Scoreboard: Sixteen Metrics Across Four Themes
Why This Matters
This session decides whether anything you have written in the past fortnight survives contact with your own board. AI in your organisation will be judged on numbers, and the numbers most programmes offer are the wrong ones: models built, dashboards launched, people trained. None of those is value. What follows is a scoreboard of sixteen measures across four themes, taught in energy language and then translated, because a bank and a hospital need the same four themes with different nouns. Then the discipline that makes it real: every metric gets a baseline, an owner and a cadence, or it does not go on the page.
25-Minute Module
0:00 – 5:00
Activity is not value
- There is a standard disguise for a failing AI programme, and you have seen the slide. Eleven models in production. Four dashboards live. Two hundred people trained. Every one of those can rise while the business performs as it did last year.
- THE DOCTRINE
Activity metrics are not value. Outcome metrics moved against a baseline are.
- The baseline matters for a reason that is not administrative. Without a defined starting point, no improvement can be credibly claimed and every claim becomes opinion, which means the most confident function wins the argument.
If a metric on your board pack has no baseline, it is not a metric. It is a sentence with a number in it.
5:00 – 12:00
Sixteen measures, four themes
- Sixteen measures across four themes give a board the whole picture on one page. Read the middle column for the discipline and the right-hand column for the translation into your own business.
Theme The measures, in energy wording The same theme elsewhere Reliability and operations Downtime reduction, maintenance compliance, equipment availability with mean time between failures and to repair beneath it, deferment reduction, report timeliness, procurement-delay reduction Unplanned outage, planned-work compliance, capacity availability, throughput lost, reporting timeliness, supply delay. A hospital reads theatre availability and cancelled procedures Commercial Bid quality, win-rate improvement, contractor-performance score, cost saved, revenue protected or recovered Proposal quality, win rate, supplier performance, cost removed, revenue protected Compliance and data Submission accuracy, data-quality score, safety-observation closure rate Submission accuracy, data-quality score, issue closure rate Capability Adoption rate, training completion The same two, meaningful only beside an outcome measure - That last row deserves a moment. Adoption and training are the two activity measures that earn a place on a board page, and only because strong outcomes with no adoption are fragile while adoption with no outcomes is theatre.
Eight to twelve of these is a board scoreboard. All sixteen is a management report. Choose deliberately, and say which you are presenting.
12:00 – 18:00
Translating it into your own business
- ENERGY · THE WORKED CASE
- On a mature onshore asset the measures are concrete: deferment reduction against last year's monthly average, mean time between failures on the critical compressors, maintenance compliance, and the time from a field signal to a validated action. Each carries a figure, a date and a named engineer.
- The unit of value is the barrel protected. Every measure converts into it eventually, which is what makes the page arguable.
OUTSIDE ENERGY · THE PARALLEL- Retail and distribution. The same themes, different nouns: on-shelf availability and vehicle uptime for reliability, delivery timeliness for latency, promotion performance and supplier scorecards for commercial, stock-record accuracy for data, till-level adoption for capability.
- The unit is the sale protected or the case delivered. A hospital would use the procedure, a bank the approved facility. Choose it once and hold every function to it.
- The translation rule is short: keep the four themes, change the nouns, never relax the baseline requirement. Where a measure has no genuine equivalent, drop it rather than inventing an analogue nobody will use.
- THE DOCTRINE
No number without a baseline. No insight without a named owner.
Your finance and operations functions already hold four or five of your baselines. Look there before commissioning a study.
18:00 – 25:00
The board page, and Week 3 wrap
- One page, five columns: the metric, the baseline with its date and source, the current value, the trend, and the owner. Put the review cadence beside each, or a monthly metric sits on a quarterly page and nobody acts.
Metric Baseline, with date and source Current Owner Cadence Unplanned downtime, critical equipment Illustrative: 46 hours a month, 2026 average, maintenance system Illustrative: 31 hours Maintenance manager Monthly Accuracy of regulatory submissions Not yet established Not available Compliance manager Quarterly - Both rows are illustrative, and the second is the instructive one. Not yet established is an honest entry and better than a plausible figure nobody can source.
- The questions that expose a programme in trouble, in order: what was the baseline and where did it come from; who owns this number; what has moved since we last met; and what have we stopped doing? A programme that has never stopped anything is accumulating, not prioritising.
- Three weeks in you hold a governance standard, one number traced, a safe-use standard, and the line you will not let an agent cross. This session adds Section 15: our board scoreboard, metric, baseline, owner, cadence. Practice Exam 3 opens after this session.
Come to Week 4 with the scoreboard finished. It becomes the spine of everything that follows, and retrofitting it later is twice the work.
Key Terms
Baseline
The measured starting point of a metric, carrying the date it was taken and the system it came from. Without it, improvement is opinion rather than evidence.
Outcome metric
A measure of something the business already cares about, such as output, downtime, cost or accuracy, moved against a baseline. An activity metric counts effort instead.
Deferment
Output the business was capable of producing and did not achieve. Deferment is the energy term; elsewhere it is cancelled procedures, missed deliveries or idle lines.
Review cadence
The rhythm at which a metric is reviewed and its owner answers for it: daily on exceptions, weekly on control health, monthly on performance and investment.
Self-Check — Before You Submit
1. Which of these belongs on a board scoreboard?
- A. The number of AI models in production
- B. Unplanned downtime against a dated baseline, with a named owner
- C. The percentage of staff who have attended AI training
- D. The number of dashboards delivered this quarter
B. Unplanned downtime against a dated baseline, with a named owner — Only one option measures something the business already cares about against a starting point. Models, dashboards and attendance are activity and can all rise while performance is unchanged. Training belongs beside an outcome measure, never instead of one.
2. Your programme office presents a metric showing a 22 per cent improvement but cannot say where the starting figure came from. What do you do?
- A. Accept it and ask for the source at the next review
- B. Halve the claimed improvement as a prudence adjustment
- C. Ask finance to recalculate the improvement
- D. Take it off the page until the baseline has a date and a source, and commission that baseline first
D. Take it off the page until the baseline has a date and a source, and commission that baseline first — An unsourced baseline makes the improvement unarguable in either direction, so it cannot stay on a board page. Nor should it be quietly discounted, because an invented adjustment is still invented, and finance cannot recompute from a source that never existed.
3. Why do adoption and training completion appear on the scoreboard at all, given that both are activity measures?
- A. Because strong outcomes with no adoption are fragile, and adoption with no outcomes is theatre, so each is read against the other
- B. Because a board needs one metric that is straightforward to move
- C. Because human-resources reporting requires them
- D. Because they are the most reliable predictor of financial return
A. Because strong outcomes with no adoption are fragile, and adoption with no outcomes is theatre, so each is read against the other — They sit there as a cross-check on durability, not as evidence of value. They are certainly easier to move than an outcome measure, which is precisely why they must never be shown alone.
Today’s Assignment
Build the board scoreboard for your own organisation: eight to twelve metrics, each with a baseline value, the source of that baseline, an owner by role and a review cadence.
- The scoreboard, eight to twelve metrics. Cover all four themes, in your own nouns. For each metric give the measure, the baseline value, the date and system it came from, the owner by role, and the cadence.
- The unit of value. One sentence naming the single unit your business converts improvement into, whether the barrel, the case, the claim, the shipment, the bed or the naira, and how at least three metrics convert into it.
- Missing baselines, about 150 words. Mark every metric whose baseline does not yet exist and, for each, name who will establish it and by what date. In most organisations this is a third of the page.
- The two you would defend, about 200 words. Choose the two metrics you would put in front of your own board first, and explain why those two, in your business, would expose a programme in trouble earliest.
- The honest part. Name one metric your organisation reports today that you now believe is activity dressed as outcome, and what you would replace it with.
YOUR CHARTER
Section 15 — Our board scoreboard: metric, baseline, owner, cadence.
How this is marked — the rubric
| Eight to twelve metrics across all four themes, each with a baseline value and its dated source | 30 |
| An owner named by role and a review cadence stated for every metric, with no blanks | 25 |
| Missing baselines marked honestly, each with a named owner and a date to establish it | 20 |
| Unit of value named, with at least three metrics shown converting into it | 15 |
| One current activity metric identified, with an outcome metric proposed to replace it | 10 |
| Total · pass mark 70 | 100 |
Submitting session 15
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: 15, 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 3 — Trust, Control and the Board 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 3 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Board reportingBaseline disciplineOutcome measurementMetric ownershipOperating cadence