THE INTELLIGENT EPC LEADERSHIP PROGRAMME
The contractor that can explain its delivery.
30-Session Programme · 25 Minutes Per Day · Free · For leaders in Nigeria’s engineering, procurement and construction sector
WEEK 5 OF 6 · CLAIMS, TRUST, VALIDATION AND FINANCE · SESSIONS 21–25 · WORK AT YOUR OWN PACE
Part Four · Construction, Commissioning and AcceptanceSESSION 21
Change, Claims and Dispute Evidence
Why This Matters
Every EPC project generates change. If a contractor assembles its position months after the event, memories have hardened and the file may be thin. That is when a fluent assistant becomes dangerous: it can turn an uncertain recollection into something that reads like a contemporaneous fact. This session teaches Chapter 22 for the executive. Preserve the event before arguing the entitlement. Connect it to the consequence that actually followed. Ask the service for a review pack, never a legal conclusion. Protect privileged and client material. Then learn from change honestly, so a recovered omission is not reported as a productivity success. Your commercial credibility depends on all five.
25-Minute Module
0:00 – 5:00
Preserve the event before arguing the entitlement
- A defensible change record starts with what happened, when it happened, who issued or observed it, affected scope and the information available at the time. Preserve original instructions, drawings, correspondence, daily records and relevant schedule updates.
- AI can organise a chronology and identify missing links. It should label inferred dates and associations as inferred.
- THE DOCTRINE
An assistant must not rewrite an uncertain recollection into a contemporaneous fact. Preserve the event first; argue the entitlement later.
- The commercial owner assesses whether the event is a change, which notice provisions apply and how it may be treated. Entitlement, causation and valuation remain separate questions. A supported event does not automatically establish all three.
- This is the Data Foundation Precondition applied to a dispute: an observation is not an estimate.
The best time to write the change record is the day the instruction arrives, not the day the account is due.
5:00 – 11:00
Connect the event to actual consequence
- Plain language first: a delay matters commercially only if it moves a milestone or completion. A late activity with room around it may move nothing.
- For schedule exposure, trace the affected activity and milestone under the relevant analysis and contract. Do not simply multiply every delayed activity by the project's daily cost. Concurrent constraints and available mitigation can change the consequence.
Consequence What to trace What to avoid Schedule The affected activity and milestone under the relevant analysis and contract Multiplying every delayed activity by a daily cost Cost Incremental quantities, resources, indirects and supported rates Counting the same labour in a direct-cost claim and a broad prolongation amount Cost state Which costs were actually incurred, committed or forecast Presenting a forecast as an incurred cost - You met the casebook's simple network in Session 18. Chapter 22 uses it to show why a delayed noncontrolling path may not move final completion. Real delay assessment can require considerably more detailed evidence and specialist analysis.
- I would ask you to treat a claim figure without a traced consequence as a hypothesis, not a number.
Ask which milestone moved, and by how much, before you ask what the delay cost.
11:00 – 17:00
Generate a review pack rather than a legal conclusion
- A suitable service can produce an event summary, source index, chronology, affected packages, notice status, potential cost and unanswered questions. It should include adverse evidence and client responses. Excluding inconvenient documents weakens both decision-making and credibility.
- THE DOCTRINE
The service produces a review pack for competent people to assess. Notices and formal positions require human approval, and the final version and its transmission evidence are retained.
- On the Autonomy Ladder this is drafting for review. A draft incorrectly addressed or based on an obsolete amendment can create practical harm even when its language is polished. Permission to draft is not permission to send.
- AT THE WORKFRONT · THE WORKED CASE
- Workfront. In a fictional EPC contractor, a site engineer's daily record notes a verbal client instruction to resequence a tie-in. The assistant builds the chronology, links the daily record, the later email and the schedule update, and marks the date of the verbal instruction as inferred because only the email is dated.
IN THE COMMITMENT · THE PARALLEL- Commitment. The commercial manager receives the pack, including the client's earlier response disputing the instruction. Whether it is a change, which notice provision applies and what it is worth stay with the commercial owner and the competent professionals. The assistant offers no view on entitlement.
- Its job is to leave a reviewer able to trace each material assertion to evidence and see what competent analysis is still required.
A pack that leaves out the client's best document is a pack your opponent will finish for you.
17:00 – 22:00
Protect privilege and learn without disguising performance
- Apply the authorised treatment to legal advice, negotiation material and client information. A broad project assistant should not retrieve every document merely because it is stored in the project folder.
- Evaluate access at retrieval and at output. A model can disclose restricted information in an answer without displaying the underlying document, so permission tests must examine content as well as file links.
- Analyse causes of change and the contractor's response. Separate client scope growth, engineering correction, procurement substitution and execution rework. The same total cost movement can have very different implications for future bids.
- THE DOCTRINE
Keep the original estimate, authorised changes and current forecast distinguishable. A project that requires a higher contract value to recover an original omission should not be reported as a productivity success.
- This is the Operations Excellence Lens at work: it protects the business from celebrating a metric that does not resolve its obligation.
If your best-performing project is the one with the largest variation account, look again at the estimate.
22:00 – 25:00
Write it down before you close the laptop
- Take one recent change event on your decision chain and note what now exists only in someone's memory.
- Name who in your company holds the change, notice and treatment assessment by role, and where competent legal or specialist analysis is still required.
- Today's assignment produces Section 21 — The event preserved, its consequence connected and the review pack specified.
Write the pack specification as if the other side will read it, because one day they may.
Key Terms
Contemporaneous record
An original instruction, drawing, correspondence, daily record or schedule update made at the time of the event, as distinct from a later recollection or an inferred date.
Entitlement, causation and valuation
Three separate questions about a change event: whether the contractor is entitled, whether the event caused the consequence, and what that consequence is worth. A supported event does not settle all three.
Review pack
A service output for competent reviewers: event summary, source index, chronology, affected packages, notice status, potential cost, unanswered questions, adverse evidence and client responses. It is not a legal conclusion.
Output-level access control
Testing what a model discloses in its answer as well as which files it retrieves, because restricted content can appear without the underlying document being displayed.
Self-Check — Before You Submit
1. A project director proposes valuing a delay claim by multiplying every delayed activity by the project's daily cost. What does Chapter 22 require instead?
- A. Use the daily cost but add a contingency for concurrency
- B. Trace the affected activity and milestone under the relevant analysis and contract, considering concurrent constraints and mitigation
- C. Let the assistant select the highest supportable figure
- D. Value only activities on the client's latest schedule
B. Trace the affected activity and milestone under the relevant analysis and contract, considering concurrent constraints and mitigation — The chapter says do not simply multiply every delayed activity by the daily cost. A contingency for concurrency is the tempting answer, but it still skips the trace of actual consequence, and a delayed noncontrolling path may not move completion at all.
2. Your commercial team wants the claims assistant to state whether the client is liable for an event. What is the right response?
- A. Allow it if the assistant cites every source
- B. Allow it for events under a set value
- C. Allow it once the legal team has trained the prompt
- D. Decline: the service produces a review pack, and entitlement, causation and valuation stay with the accountable professionals
D. Decline: the service produces a review pack, and entitlement, causation and valuation stay with the accountable professionals — Chapter 22 limits the service to a review pack and requires human approval for notices and formal positions. Citations make a pack traceable; they do not turn a model into the holder of a legal conclusion.
3. A project recovered an original estimating omission through a higher contract value and now shows an improved forecast margin. How should it be reported?
- A. Keep the original estimate, authorised changes and current forecast distinguishable, and do not report it as a productivity success
- B. As a productivity success, because the forecast improved
- C. As client scope growth, because the contract value rose
- D. As execution rework, because cost moved
A. Keep the original estimate, authorised changes and current forecast distinguishable, and do not report it as a productivity success — The chapter says exactly this. Calling it client scope growth is tempting because the value came from the client, but the cause was the contractor's own omission, and future bids need that distinction.
Today’s Assignment
Write Section 21 of your plan: one change event on your decision chain preserved, its consequence connected and the review pack your company should require specified.
- Choose one change event from your company's recent work on your decision chain. List what was preserved at the time (instructions, drawings, correspondence, daily records, schedule updates) and what was not. Be honest about the gaps.
- Separate the facts you can evidence from the dates and associations you can only infer, and label the inferred ones.
- State how you would connect the event to actual consequence: the milestone it moved or did not move, incremental cost by incurred, committed and forecast, and where labour could be double counted.
- Specify the review pack your business should receive, including adverse evidence and client responses, and name by role who approves notices and formal positions. Do not reach any legal conclusion yourself.
- State who may access privileged and negotiation material, and how you would test the assistant's answers as well as its file links.
- Classify your last few material changes as client scope growth, engineering correction, procurement substitution or execution rework, and say whether any recovery has been reported as performance.
YOUR 90-DAY PLAN
Section 21 — The event preserved, its consequence connected and the review pack specified.
How this is marked — the rubric
| Separates preserved contemporaneous evidence from inferred dates and recollection, and names the gaps honestly | 25 |
| Traces the event to an actual milestone and cost consequence without daily-rate multiplication or double counting | 25 |
| Specifies a review pack with adverse evidence, human approval of notices and no legal conclusion | 25 |
| Sets privilege and output access tests and classifies causes of change without disguising performance | 25 |
| Total · pass mark 70 | 100 |
Submitting session 21
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: 21, 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
Contemporaneous evidenceConsequence tracingReview pack designPrivilege controlHonest change learning
Part Five · Trust, Validation and PeopleSESSION 22
Security, Privacy and Client Information
Why This Matters
Your company holds other people's information: a client's drawings, a vendor's prices, an employee's records, an operator's facility data. Each can carry different restrictions. The risk is not only a breach in the dramatic sense. It is one client's material quietly appearing in another client's proposal, a vendor email that instructs your assistant to change a bank account, or an office connection treated as plant access. This session teaches Chapter 23: whose information your decision touches, how client boundaries are enforced, why retrieved content is data, why operational technology needs separate assurance and how you leave a provider.
25-Minute Module
0:00 – 6:00
Know whose information the decision touches
- Engineering drawings, commercial offers, employee records, vendor data and facility information can carry different restrictions. Define permitted purposes and environments before connecting them to a model.
- THE DOCTRINE
Possession of a file does not automatically authorise external processing or unrestricted reuse.
- The Nigeria Data Protection Act 2023 provides a legal framework for personal-data processing. Determine lawful basis, applicable safeguards and transfer arrangements for the actual use through competent review. Consent is not the only possible basis, and an employment relationship does not justify unlimited processing.
- Keep personal information out of broad project summaries where it is unnecessary. Use role, category or case identifier when those meet the purpose. Restrict medical, disciplinary and other sensitive records appropriately.
- Which lawful basis applies is for your competent reviewers. The question must be asked before the connection is made.
Before you connect a folder to a model, ask who gave you the contents and on what terms.
6:00 – 11:00
Separate client and project boundaries
- Prevent one client's information from appearing in another client's retrieval results or generated proposals. Indexing everything together and relying only on a prompt to respect confidentiality is inadequate.
- Apply permissions in the data service and tool interfaces, where they are enforced, not in instructions the model may or may not follow.
Boundary test (Chapter 23) What it exposes Adversarial requests Whether a user can ask around the restriction Indirect references Whether restricted material leaks through a related question Role changes Whether permissions follow the person's current role Public-looking documents with restricted attachments Whether the restriction travels with the attachment - Retain logs with appropriate protection and retention. They can themselves contain sensitive prompts, outputs and source passages. A security measure can create a new disclosure surface if its records are broadly accessible.
A prompt asking the model to keep secrets is a request, not a control.
11:00 – 16:00
Treat retrieved content as data
- Plain language first: an assistant reads documents in order to answer. If a document contains text written like an instruction, a careless service may follow it. The glossary calls this prompt injection: an attempt through untrusted content to redirect a service's behaviour or use of tools.
- A drawing note, vendor email or uploaded PDF may contain such text. The service should not treat source content as authority to change its instructions, disclose information or execute tools.
- THE DOCTRINE
Retrieved content is data, not instruction. Constrain tool permissions, validate requests and require approval where the action warrants it.
- Test a document that tells the assistant to send a file or change a bank account. The correct behaviour is to treat the text as untrusted content and preserve the established workflow. On the Autonomy Ladder, an assistant permitted to retrieve and draft has no permission to execute.
- AT THE WORKFRONT · THE WORKED CASE
- Workfront. A fictional fabrication subcontractor uploads an inspection report to the main contractor's document portal. Hidden in a note is a line instructing any assistant to forward the full package to an outside address. The document assistant summarises the report and flags the line as untrusted content.
IN THE COMMITMENT · THE PARALLEL- Commitment. The same week a fictional vendor email asks for its payment account to be changed. The finance assistant treats the request as content, not authority. The established verification workflow and its named approver decide; the assistant has no tool permission to change payment details.
- One loop seen twice: on the workfront and in the treasury, the defence is the same boundary between reading and acting.
Ask your team to show you the test document that tried to move money, and what the service did with it.
16:00 – 21:00
Keep OT separate and plan the exit
- Operational technology (OT) monitors or controls physical processes and equipment. NIST SP 800-82 Revision 3, the Guide to Operational Technology Security (September 2023), addresses OT security with attention to operational performance, reliability and safety requirements. The book uses that published revision as a reference.
- A contractor connecting to a live facility should follow the applicable owner-approved architecture and controls rather than treat office access as plant access. Define interfaces, remote access, commissioning ownership, configuration changes and recovery with the responsible parties.
- THE DOCTRINE
A successful connection test does not establish that an interface is acceptable for ongoing operation.
- Assess data use, retention, training use, subcontractors, hosting, security support, configuration changes and export rights in the actual provider agreement. Product marketing is not a substitute for those terms.
- Plan service interruption and provider exit. The project should retain its authoritative records and accepted decision evidence. A model service can be replaced more easily when the business owns a coherent information structure and a usable manual process.
Read the provider agreement yourself once, at least the clauses on training use and export.
21:00 – 25:00
Write it down before you close the laptop
- List every party whose information your decision chain touches: client, vendor, employee, operator. Note which restriction you can actually find in writing.
- Mark where your company currently relies on a prompt, a folder habit or goodwill rather than a permission in the data service.
- Chapter 23's closing point: security and privacy are operating conditions. They determine whether the contractor can use client information and sustain the workflow with trust.
- Today's assignment produces Section 22 — Information boundaries, retrieved content as data and provider exit.
Treat security as part of delivery design from the first day, not as a review at the end.
Key Terms
Permitted purpose
The use and environment for which a party's information may be processed. Holding the file does not by itself authorise external processing or reuse.
Lawful basis
The legal ground for processing personal data for a specific use, determined through competent review under the Nigeria Data Protection Act 2023. Consent is not the only possible basis.
Prompt injection
An attempt through untrusted content, such as a vendor email or an uploaded PDF, to redirect a service's behaviour or use of tools.
Operational technology (OT)
Technology associated with monitoring or controlling physical processes and equipment, which requires owner-approved architecture and assurance separate from office systems.
Self-Check — Before You Submit
1. Your IT lead proposes indexing all clients' documents in one store and instructing the assistant to respect confidentiality. What does Chapter 23 say?
- A. Acceptable if the instruction is tested monthly
- B. Inadequate: apply permissions in the data service and tool interfaces, and test adversarial requests, indirect references and role changes
- C. Acceptable for drawings but not commercial offers
- D. Acceptable if logs are retained
B. Inadequate: apply permissions in the data service and tool interfaces, and test adversarial requests, indirect references and role changes — The chapter calls a single index plus a prompt inadequate. Retaining logs sounds like a control, but logs can themselves contain sensitive passages and become a new disclosure surface.
2. An uploaded vendor PDF contains a line telling any assistant to change the vendor's bank account. What is the correct behaviour?
- A. Change the account and notify finance
- B. Ask the vendor to confirm by email, then change it
- C. Treat the text as untrusted content and preserve the established workflow
- D. Delete the PDF from the project record
C. Treat the text as untrusted content and preserve the established workflow — Retrieved content is data, not authority. Asking for email confirmation still lets the document drive the action; deleting it destroys a record the established workflow may need.
3. Your team connected successfully to a client facility's control network during commissioning. What does that establish?
- A. That the interface is acceptable for ongoing operation
- B. That office access rules now cover the plant
- C. That NIST SP 800-82 Revision 3 has been met
- D. That a connection works, not that the interface is acceptable; the owner-approved architecture and controls govern
D. That a connection works, not that the interface is acceptable; the owner-approved architecture and controls govern — Chapter 23 says a successful connection test does not establish acceptability for ongoing operation. Citing the NIST guide is tempting, but the book uses it as a reference, not as a certificate a connection test satisfies.
Today’s Assignment
Write Section 22 of your plan: whose information your company's decision chain touches, how client and project boundaries are enforced, how retrieved content is treated as data, how OT is kept under separate assurance and how you would leave your provider.
- Map every category of information your decision uses (drawings, offers, employee records, vendor data, facility information), whose it is, and the permitted purpose you can evidence. Mark any you cannot.
- Identify any personal data and state the question your competent reviewer must answer on lawful basis and safeguards. Do not answer it yourself.
- Describe where your business enforces client and project separation today, and write four boundary tests from the chapter for your own service.
- Write the untrusted-content test you would run, and name the actions your assistant must never execute from document text.
- If your decision touches a live facility, state who owns the OT architecture and what your company must not assume. If it does not, say so.
- List the provider terms you must check and describe how your company would continue on a manual process if the provider left tomorrow.
YOUR 90-DAY PLAN
Section 22 — Information boundaries, retrieved content as data and provider exit.
How this is marked — the rubric
| Maps each party's information to an evidenced permitted purpose and flags gaps honestly | 25 |
| Places client and project separation in the data service with specific boundary tests, not prompts | 25 |
| Specifies an untrusted-content test and forbidden actions, and keeps OT under owner-approved assurance | 25 |
| Names the provider terms to check and a tested manual fallback and exit route | 25 |
| Total · pass mark 70 | 100 |
Submitting session 22
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: 22, 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
Information mappingClient boundary controlUntrusted content handlingOT assuranceProvider exit planning
Part Five · Trust, Validation and PeopleSESSION 23
Model Validation and Operational Release
Why This Matters
This session decides whether any AI service in your company deserves trust with a real decision. Vendors will show you a percentage. Your team will show you a demonstration that worked. Neither tells you whether the service supports your claim, on your population, through the whole chain from evidence to reviewer, at a threshold that reflects what a missed problem costs your business. Chapter 24 gives you the discipline: define the claim before the test, build the dataset around the cases you will actually see, evaluate the complete response chain, make the commercial threshold explicit and compare the service with simpler alternatives. Bring your own service, or the proposal in front of you.
25-Minute Module
0:00 – 4:00
Define the claim before selecting the test
- State task, population, evidence, users, permitted authority and decision horizon. A test of document extraction cannot establish competence for contract interpretation or safe work release.
- Create an error taxonomy based on consequence: unsupported factual assertion, wrong revision, missed material condition, incorrect calculation, permission failure and unauthorised execution. Separate these from cosmetic defects.
- THE DOCTRINE
Agree acceptance conditions before reviewing results, or a team may redefine success around whichever metric looks strongest. Keep mandatory gates distinct from comparative performance targets.
If the acceptance condition was written after the results came in, it is a description, not a test.
4:00 – 9:00
Design the dataset around the deployment population
- Use representative and challenge cases: scans, tables, missing pages, conflicting revisions and absent answers. Keep a protected test set separate from development and prompt tuning.
- Plain language first on leakage: the test lets the model see something it could not have known at decision time, so it looks clever in testing and fails in use. If successive updates from the same package appear on both sides of a split, the model can look accurate while failing on a new package. Group related records, preserve time order and exclude information unavailable at prediction time.
- The Technical Annex adds a precise case: a notice dated Monday but published Wednesday was unavailable to Tuesday's monitoring run. Store both event time and the time the organisation could access the information.
- Document sampling limitations. A pilot on one discipline and one client's clean files does not establish performance across all clients and document types.
Ask how many packages and projects sit behind the test, not how many rows.
9:00 – 15:00
Evaluate the chain and make the threshold explicit
- Evaluate the complete response chain: evidence found, source support and completeness, calculations against an independently checked result, calibration and warning time, and for execution, permissions, approval and recovery. The human reviewer is part of the chain; measure time, corrections, missed defects and queue capacity.
- The Annex's hypothetical quote test: ten required evidence units, eight retrieved, six used correctly, one unsupported commitment added. Retrieval recall is 8/10 = 80%; correct use of retrieved units 6/8 = 75%; coverage in the final answer 6/10 = 60%. The invented commitment stays a separate defect.
Chapter 24 hypothetical: 1,000 quote requirements, 50 material omissions Calculation Result Accuracy of the service (40 true positives, 60 false positives, 10 false negatives, 890 true negatives) 930/1,000 93% Accuracy of a service that flags nothing 950/1,000 95% Precision / recall 40/100 and 40/50 40% / 80% Specificity 890/950 93.68% Assumed loss: NGN 2,000 per unnecessary review, NGN 200,000 per missed omission 60 × 2,000 + 10 × 200,000 NGN 2.12 million Flag-nothing assumed loss 50 × 200,000 NGN 10 million - THE DOCTRINE
A classification score is useful only when it changes a defined decision. Accuracy alone would reward the useless comparator.
- Under the chapter's two-action loss model, flag when a calibrated probability exceeds 2,000/202,000, approximately 0.99%. The book is explicit: this is an algebraic consequence of assumed costs, not a general EPC release threshold. Reviewers examine 100/40 = 2.5 flagged requirements per detected omission.
Never accept an accuracy figure without the prevalence and the cost of a miss beside it.
15:00 – 21:00
Calibration, abstention and simpler alternatives
- Plain language first on calibration: when the service says 70%, events like that should happen about seven times in ten across comparable cases. Casebook Case 6 (assumed): forecasts 0.1, 0.3, 0.7 and 0.9; outcomes 0, 1, 0, 1. The Brier score is (0.01 + 0.49 + 0.49 + 0.01)/4 = 0.25. A constant 0.5 also scores 0.25. Four cases cannot establish calibration.
- Case 6 then ties probability to response: an assumed ₦2 million intervention preventing an eligible ₦10 million loss breaks even at p = 0.2; if it prevents only half, at p = 0.4.
- Plain language on abstention: the service declines cases it should not answer and routes them to people. In the Annex's hypothetical, it accepts 120 of 200 cases with three material errors: coverage 60%, selective error 3/120 = 2.5%. The remaining 80 need an owned route; they have not disappeared from the workload.
- AT THE WORKFRONT · THE WORKED CASE
- Workfront. A fictional EPC contractor tests a quote-review assistant against a requirement checklist and spreadsheet controls, holding cases and cut-off constant. Removing the generative drafting stage leaves omissions unchanged; exact identifier matching fixes wrong-package retrieval.
IN THE COMMITMENT · THE PARALLEL- Commitment. The release argument says so plainly: generation improves phrasing, the identifier fix reduces material errors, the threshold rests on the assumed cost of a missed omission, and refused cases go to a named estimator.
- Test against simpler alternatives: source subscriptions, keyword rules or a curated digest for portal monitoring; a mandatory-field validator and supervisor review for PSRs; a checklist and spreadsheet controls for quotes.
An AI service must earn its additional complexity against a useful baseline, and some will not.
21:00 – 25:00
Shadow, release, the record, then write
- Improvement after implementation does not by itself show AI caused it. You met the zero-failure bound in Session 19: after 100 clean tests it is still approximately 2.95%, a teaching calculation, not a release standard.
- Operate in shadow alongside the existing method, preserving both outputs and the information cut-off. At release record evidence, limitations, owners, configuration, authority, monitoring and fallback, and show the team can stop the service and use the manual process. Keep each test and incident in the Appendix E evaluation and incident record: a model restart does not correct a wrong purchase, notice or certificate already issued.
- THE DOCTRINE
An accepted service is a maintained capability, not a permanently approved model.
- Pop Quiz 5 opens now: five questions on Sessions 21 to 23, closed book. Today's assignment produces Section 23 — The claim, the dataset, the response chain and the explicit threshold.
Write the claim in one sentence before you look at a single result.
Key Terms
Leakage
Information improperly available during evaluation, such as future status used to predict an earlier outcome, which makes a model look more accurate than it will be in use.
Calibration
Agreement between predicted probabilities and observed frequencies across an appropriate population, assessed with uncertainty and sufficient data.
Abstention (selective prediction)
A service declining some cases and reporting performance on those it accepts. Coverage and selective error are reported together, and refused cases need an owned fallback route.
Protected test set
Evaluation cases kept separate from development, prompt tuning and threshold selection, so they assess the fixed service rather than the team's adjustments.
Self-Check — Before You Submit
1. In the Chapter 24 hypothetical, a service scores 93% accuracy on 1,000 quote requirements. Why is that alone insufficient?
- A. A service that flags nothing scores 95% on the same sample while missing every omission
- B. 93% is below the book's release standard
- C. Accuracy cannot be computed for quotes
- D. The sample is too large to be representative
A. A service that flags nothing scores 95% on the same sample while missing every omission — With 50 omissions in 1,000, flagging nothing gives 950/1,000. The book sets no general release standard, so the second option is a trap: the decision needs precision, recall, the population and the cost of a miss.
2. A vendor's test shows strong forecast accuracy, but successive monthly updates from the same packages appear in both training and test data. What is the main concern?
- A. Calibration
- B. Abstention
- C. Leakage: the model can look accurate while failing on a new package
- D. Prompt injection
C. Leakage: the model can look accurate while failing on a new package — Chapter 24 names this exact pattern. Calibration matters too, but it cannot be trusted until related records are grouped and time order is preserved.
3. A quote-review service accepts 120 of 200 eligible cases and makes three material errors among them. What must leadership decide?
- A. Nothing: 2.5% selective error is acceptable
- B. The owned route, response time and capacity for the 80 refused cases
- C. Whether to raise coverage to 100% immediately
- D. Whether to drop the material-error taxonomy
B. The owned route, response time and capacity for the 80 refused cases — Coverage is 60% and selective error 2.5%, but the 80 refused cases remain workload. The Annex says leadership owns that capacity decision; treating the error rate as the whole answer hides the queue.
Today’s Assignment
Write Section 23 of your plan: the claim your company's service must support, the dataset, the complete response chain, the explicit commercial threshold and the simpler alternatives it must beat.
- State the claim in one sentence: task, population, evidence, users, permitted authority and decision horizon. If your company has no AI in use yet, write it for the proposal in front of you.
- Write the error taxonomy by consequence and the acceptance conditions, separating mandatory gates from comparative targets, before any results.
- Describe the dataset: representative and challenge cases, the protected set, how related records are grouped and how time order and access time are preserved to prevent leakage.
- Map the complete response chain, including the human reviewer's time and queue capacity, and the route for refused cases.
- Make the commercial threshold explicit with your own assumed costs of an unnecessary review and a missed problem, labelled as assumptions to be validated.
- Name at least two simpler alternatives your service must beat, and say honestly whether you expect it to.
YOUR 90-DAY PLAN
Section 23 — The claim, the dataset, the response chain and the explicit threshold.
How this is marked — the rubric
| States a specific claim with population, authority and horizon, and acceptance conditions set before results | 25 |
| Designs a dataset that prevents leakage and dependence, with a protected test set and stated sampling limits | 25 |
| Evaluates the complete chain including reviewer capacity and refused cases | 20 |
| Sets an explicit, assumption-labelled threshold and compares against named simpler alternatives | 30 |
| Total · pass mark 70 | 100 |
Submitting session 23
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: 23, 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 5 — Claims, Security and Validation 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 5 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Claim definitionLeakage preventionResponse chain evaluationCommercial thresholdsCalibration judgementRelease discipline
Part Five · Trust, Validation and PeopleSESSION 24
People, Capability and Organisational Learning
Why This Matters
Every service we have discussed depends on people who are rarely in the business case: the person who maintains source relationships, the reviewer who catches the unsupported answer, the administrator who removes a leaver's access, the engineer who explains why a lesson does not transfer. If you do not name them and fund their time, the work lands on your best technical reviewers as an invisible addition, and the net capacity effect may be negative. This session teaches Chapter 25: competent owners for the workflow, competence shown on real tasks, experience captured without invented certainty, incentives that do not corrupt the data, and succession that survives an individual's absence.
25-Minute Module
0:00 – 6:00
The workflow needs competent owners
- AI changes some tasks and creates others. Someone must maintain source relationships, review unsupported answers, administer permissions, resolve data defects and investigate performance. Name these responsibilities and fund the capacity.
- THE DOCTRINE
If a service saves drafting time while doubling specialist review effort, the net capacity effect may be negative. Measure the full workflow.
Role (Chapter 25) What training must enable Executives Evaluate decisions, authority and economics Engineers Assess technical inputs and limitations Procurement Interpret exceptions and preserve vendor commitments Document control Revision and access competence Information technology Supported operation and recovery - In Sequencing Discipline terms, a bounded workflow needs an owner before it needs a model. A service with no funded owner is not yet a workflow.
If the review work is not in someone's job description, it is in nobody's.
6:00 – 11:00
Demonstrate competence on real tasks
- Use an accepted practical assessment: find the applicable revision, identify an unsupported claim, reconstruct a calculation, challenge a wrong tag match, handle a missing answer and recover from interruption.
- Certificates can document training; successful task performance establishes usable capability.
- Assess teams as well as individuals. A capable engineer cannot sustain a service if document updates are uncontrolled. A capable analyst cannot resolve an approval ambiguity. The operating model depends on their interfaces.
- AT THE WORKFRONT · THE WORKED CASE
- Workfront. A fictional EPC contractor's document controller is given a live package and asked to find the applicable revision, then to challenge a tag match the assistant got wrong. She finds the revision but accepts the wrong tag because the assistant's citation looked complete.
IN THE COMMITMENT · THE PARALLEL- Commitment. The project director had planned to report the team as trained on the strength of attendance. The practical result changes the decision: the service stays advisory, and the tag-challenge exercise is repeated with the engineer and the document controller together, because the failure sat at their interface.
- One loop seen twice: the same practical evidence that coaches the workfront tells leadership what authority the service may hold.
Ask for a demonstration on a real package, not a certificate on a wall.
11:00 – 16:00
Capture experience without inventing certainty
- Project learning should preserve scope, conditions, evidence, action, outcome and limits. An assistant can retrieve related cases and draft lessons.
- The responsible owners should distinguish a demonstrated cause from an interpretation. A generated lesson that reads with confidence has not thereby been demonstrated.
- THE DOCTRINE
A previous successful response may be unsuitable for a different project. Retain the conditions that made it work.
- The book's own example: a lesson reduced to a slogan such as expedite early can conceal whether the original intervention had usable float, funding and technical approval.
- This links back to the change analysis of Session 21: a lesson is only as honest as the cause recorded beneath it.
When a lesson fits on a poster, ask what conditions were left off it.
16:00 – 21:00
Incentives shape the data; succession keeps the capability
If people are rewarded only for The data may become Progress reported Optimistic reports Deployment of a pilot A record that minimises validation failures Supplier scores that determine exclusion without review Supplier reporting optimised rather than delivery - Align incentives with accepted work, early supported exceptions, quality of decisions and verified outcomes. Provide a route to challenge incorrect records and model outputs. Learning improves when people can reveal problems without having to defend themselves against an automated conclusion.
- At least the essential workflow knowledge should survive an individual's absence. Retain definitions, evaluation cases, configuration, operating procedures and recovery instructions. Test continuity through a practical exercise rather than a handover presentation alone.
- THE DOCTRINE
Permanent dependence on external support can be acceptable if it is explicit, funded and supported. It should not be disguised as complete capability transfer.
- For external support, specify what local teams will maintain and what requires specialist assistance.
Look at what your bonus scheme rewards, then look at your dashboards; they usually agree.
21:00 – 25:00
Write it down before you close the laptop
- Name every task your decision chain's service creates, and who holds each today. Mark the ones that have fallen on an already loaded reviewer.
- Write down the one incentive in your company most likely to make the data on your decision chain optimistic.
- Chapter 25's closing line: leadership must create the time, authority and incentives for competence to become part of everyday work.
- Today's assignment produces Section 24 — Competent owners, demonstrated competence, honest incentives and succession.
Fund the reviewer before you fund the next feature.
Key Terms
Competent owner
A named person with funded capacity for a task the service creates, such as maintaining sources, reviewing unsupported answers, administering permissions or investigating performance.
Practical competence assessment
An accepted test on real tasks, such as finding the applicable revision or recovering from interruption, which establishes usable capability where a certificate only documents training.
Net capacity effect
The change in total workflow effort once drafting savings, specialist review, correction and support are all counted.
Continuity exercise
A practical test that essential workflow knowledge survives an individual's absence, using retained definitions, evaluation cases, configuration, procedures and recovery instructions.
Self-Check — Before You Submit
1. A drafting assistant halves engineers' drafting time but doubles the senior reviewer's checking effort. What should leadership conclude first?
- A. The service is a success because drafting time fell
- B. The service should be withdrawn
- C. The reviewer needs more training
- D. Measure the full workflow: the net capacity effect may be negative
D. Measure the full workflow: the net capacity effect may be negative — Chapter 25 uses this pattern. Celebrating the drafting saving counts only half the workflow; withdrawing the service jumps to a decision before the full effect is measured.
2. Which evidence best establishes that a document control team can operate the service?
- A. Successful performance on practical tasks such as finding the applicable revision and recovering from interruption
- B. Completion certificates for every team member
- C. A handover presentation from the implementation partner
- D. The vendor's statement of ease of use
A. Successful performance on practical tasks such as finding the applicable revision and recovering from interruption — Certificates document training; task performance establishes usable capability. A handover presentation is what the chapter says is insufficient for continuity on its own.
3. A specialised service will always need an external partner for model maintenance. How should this be presented to the board?
- A. As complete capability transfer, to support the business case
- B. As explicit, funded and supported dependence, with what local teams maintain stated
- C. As a temporary arrangement without a date
- D. It should not be presented, as it is operational
B. As explicit, funded and supported dependence, with what local teams maintain stated — Permanent dependence can be acceptable if explicit, funded and supported. Calling it capability transfer is the tempting presentation, and the chapter says it should not be disguised that way.
Today’s Assignment
Write Section 24 of your plan: the competent owners your company's workflow needs, how their competence is shown on real tasks, how experience is captured, how incentives shape the data and how succession is maintained.
- List the tasks the service on your decision chain creates and name an owner by role for each, with the time you will fund. If your business has no AI in use yet, work from the proposal in front of you.
- Say honestly where the work currently falls on an overloaded reviewer, and estimate the net capacity effect as far as you can evidence it.
- Write three practical competence tasks for your own teams, including one that tests an interface between two roles.
- Describe how lessons on your decision chain will retain scope, conditions, evidence, action, outcome and limits, and who distinguishes cause from interpretation.
- Name one incentive in your company that could make the data optimistic and what you would change.
- State what must survive an individual's absence, how you will test continuity, and what external dependence remains.
YOUR 90-DAY PLAN
Section 24 — Competent owners, demonstrated competence, honest incentives and succession.
How this is marked — the rubric
| Names each created task with an owner by role and funded capacity, and assesses net capacity honestly | 30 |
| Specifies practical competence tasks for individuals and team interfaces | 20 |
| Captures experience with context and limits, separating demonstrated cause from interpretation | 20 |
| Identifies a real distorting incentive and a tested succession and dependence arrangement | 30 |
| Total · pass mark 70 | 100 |
Submitting session 24
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: 24, 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
Workflow ownershipPractical competenceHonest lessonsIncentive designSuccession planning
Part Six · Capital and the Leadership ExecutionSESSION 25
Finance, Treasury and Portfolio Exposure
Why This Matters
A profitable contract can still break a contractor that cannot fund it. Chapter 26 puts the question plainly: cash goes out to suppliers and payroll before accepted milestones are collected, and a late invoice can raise your funding need without changing a single naira of the contract value. This session teaches Chapter 26 with the book's two worked funding cases: Casebook Case 5 and the walkthrough's delayed receipt. Both use assumed numbers, and both show the same lesson: timing alone can move your peak funding need. You will also follow currency exposure to obligations, test portfolio dependence and start the benefits ledger your board will rely on.
25-Minute Module
0:00 – 5:00
Profitability and funding are different questions
- Plain language first on peak funding: add up cash in and cash out day by day; the deepest point the running balance reaches is the money you must find from somewhere else before collections catch up.
- Map deposits, progress payments, payroll, subcontract payments, retention, taxes for competent review and financing costs to actual dates and currencies, separate from accounting recognition.
- An advance can improve liquidity while carrying security, deduction and performance obligations under its actual terms. It is not additional revenue simply because cash arrives early. A guarantee can create contingent exposure and use banking capacity even before it is called.
- THE DOCTRINE
AI can identify likely invoice delays and retrieve missing acceptance evidence. It cannot make an unaccepted milestone eligible for payment or establish that a forecast receipt will occur.
Ask your finance director for the deepest point of the cash curve, not the margin at the end of it.
5:00 – 11:00
Casebook Case 5: peak funding and the delayed invoice
- Assumed inputs, ignoring taxes, retention, financing and other movements solely for teaching: ₦300 million advance on day 0; ₦500 million to suppliers on day 5; ₦200 million labour and other eligible cash costs on day 15; ₦450 million milestone on day 30; ₦100 million paid on day 35; ₦200 million collected on day 60.
Day Movement Cumulative, as planned Cumulative, ₦450m received day 45 0 +₦300m ₦300m ₦300m 5 −₦500m Negative ₦200m Negative ₦200m 15 −₦200m Negative ₦400m Negative ₦400m 30 +₦450m (as planned) ₦50m Negative ₦400m 35 −₦100m Negative ₦50m Negative ₦500m 45 +₦450m (delayed) — Negative ₦50m 60 +₦200m ₦150m ₦150m - Peak external funding need is ₦400 million as planned and ₦500 million with the delay: an increase of ₦100 million, although the receipt amount is unchanged. The final surplus stays ₦150 million in both.
- With an assumed simple annual rate of 24% on a 365-day basis, the extra ₦450 million is funded for fifteen days: ₦450 million × 0.24 × 15/365, approximately ₦4.438 million. Actual interest depends on the funding arrangement, drawdowns, repayment and charges.
- The delay does not itself reduce the receipt. If collectability is also in doubt, model that as a separate scenario.
A payment prediction has value only when it improves acceptance evidence, invoicing, follow-up or funding.
11:00 – 16:00
The walkthrough: funding becomes the controlling constraint
- Return to the walkthrough's assumed NGN 160 million price and NGN 125 million provision from Session 7, excluding the later change. Receipts are assumed eligible, collectible net cash; no taxes, retention or security cash.
Week Movement (NGN m) Cumulative, as planned Cumulative, week-six receipt at week nine 0 Receive 32, pay 20 12 12 4 Pay 45 Negative 33 Negative 33 6 Receive 48 (as planned) 15 Negative 33 8 Pay 35 Negative 20 Negative 68 9 Receive 48 (delayed) — Negative 20 10 Receive 48 28 28 12 Pay 15 13 13 16 Receive 32, pay 10 35 35 - Peak funding need rises by NGN 35 million, from NGN 33 million to NGN 68 million, before any liquidity buffer. Receipts total 160 and payments 125, so the end position is 35 either way.
- AT THE WORKFRONT · THE WORKED CASE
- Workfront. In the walkthrough, every receipt is assumed eligible and collectible solely for the teaching calculation. The AI finance brief shows the receipt assumption, the eligible evidence and the actual collection state, and does not treat an expected invoice date as received cash.
IN THE COMMITMENT · THE PARALLEL- Commitment. Treasury sees the funded obligations at week eight against a deficit of NGN 68 million, not 33. The brief may not remove a payment to make the model easier to fund. The later change needs its own dated cash schedule.
The same milestone that commissioning chases is the one treasury is waiting for; manage it once.
16:00 – 21:00
Currency, portfolio dependence, the benefits ledger and capital
- Currency exposure follows obligations. The book's assumed example: a $1 million payable valued at ₦1,500/$ is ₦1.5 billion; at an assumed settlement rate of ₦1,800/$ it is ₦1.8 billion, a ₦300 million difference. These are teaching rates, not current quotations. A dollar receivable and payable offset only where timing, amount and accessibility make it usable. Report actual and scenario exposure separately.
- Projects can share a client, supplier, bank line, specialist team or transport route. Compute the aggregate dated cash position rather than summing each project's peak, and test coincident peaks as an explicit downside.
- THE DOCTRINE
A working-capital release is not automatically annual profit. A saved labour hour is not automatically a cash saving. An earlier receipt changes timing unless the amount or collectability also changes.
- The benefits ledger separates additional contribution, cost avoided, capacity created, working capital released and financing-cost reduction, recording baseline, counterfactual, intervention, outcome, costs, finance acceptance and which party receives the value.
- For capital allocation, compare full investment cost (data repair, integration, evaluation, review, training, support, response cost and exit) with supported benefits over a defined horizon, with downside and delayed-start cases. A board can approve a service for assurance without a precise monetary benefit; say so rather than invent a return.
If two applications claim the same schedule recovery, examine the overlap before you count it twice.
21:00 – 25:00
Write it down before you close the laptop
- Mark the deepest point of the cash curve for your decision chain's next milestone, even roughly.
- Chapter 26's close: the finance view shows whether the contractor can fund the response and whether the claimed benefit belongs to the business.
- Practice Exam 5 opens now: fifteen questions on Sessions 21 to 25, closing Week 5. Today's assignment produces Section 25 — Profit and funding separated, currency exposure and the benefits ledger.
A business case that cannot show its funding curve is not finished.
Key Terms
Peak funding need
The deepest negative point of the cumulative dated cash position, before any minimum cash buffer or financing cost; the money that must be found before collections catch up.
Advance
Early cash under the contract's actual terms, which may carry security, deduction and performance obligations. It improves liquidity but is not additional revenue because it arrives early.
Benefits ledger
A record separating additional contribution, cost avoided, capacity created, working capital released and financing-cost reduction, with baseline, counterfactual, costs, finance acceptance and the party receiving the value.
Working capital release
A reduction in cash tied to the operating cycle under a supported basis. It is not automatically recurring profit.
Self-Check — Before You Submit
1. In Casebook Case 5, the ₦450 million milestone moves from day 30 to day 45 with no change in amount. What happens to peak funding need?
- A. It is unchanged, because the receipt amount is unchanged
- B. It rises by ₦100 million, from ₦400 million to ₦500 million
- C. It rises by ₦450 million
- D. It falls, because the day-35 payment is now covered
B. It rises by ₦100 million, from ₦400 million to ₦500 million — Without the day-30 receipt, the day-35 payment takes the balance from negative ₦400 million to negative ₦500 million. Assuming no change because the amount is unchanged is exactly the error the case is built to expose.
2. A finance assistant forecasts that a milestone invoice will be paid next month, but the acceptance certificate is not yet signed. How should the funding plan treat it?
- A. As received cash, since the forecast is confident
- B. As revenue earned, since the work is physically complete
- C. As a forecast receipt with its eligible evidence and collection state shown; AI cannot make an unaccepted milestone eligible
- D. Remove the related supplier payment to balance the plan
C. As a forecast receipt with its eligible evidence and collection state shown; AI cannot make an unaccepted milestone eligible — Chapter 26 says AI cannot make an unaccepted milestone eligible or establish that a forecast receipt will occur. The walkthrough forbids removing a payment to make the model easier to fund.
3. Three projects each show a peak funding need, at different dates, and two share one bank line and one client. What does the chapter require?
- A. Compute the aggregate dated cash position, model shared events and test coincident peaks as an explicit downside
- B. Sum the three peaks as the portfolio need
- C. Use the largest single project peak
- D. Average the three peaks
A. Compute the aggregate dated cash position, model shared events and test coincident peaks as an explicit downside — Summing peaks that occur at different times overstates the base case, while ignoring the shared bank line and client hides the downside. The chapter asks for both the dated aggregate and the coincident-peak scenario.
Today’s Assignment
Write Section 25 of your plan: separate profitability from funding for your company's decision, follow currency exposure to obligations, state portfolio dependence and start the benefits ledger.
- Build a dated cash schedule for one contract or milestone on your decision chain, using actual or clearly labelled assumed figures, and identify the peak funding need. Then delay the key receipt and recompute it.
- State which receipts depend on acceptance evidence your company does not yet hold, and which assumptions your finance team would challenge.
- List currency-specific receipts and payments, the conversion basis and date, and whether any apparent offset is usable in timing, amount and accessibility.
- Name the clients, suppliers, bank lines, specialist teams or routes your business shares across projects, and describe one coincident-peak downside scenario.
- Start the benefits ledger for your decision: baseline, counterfactual, intervention, costs, finance acceptance and which category each claimed benefit belongs to. If the honest rationale is assurance rather than money, say so.
YOUR 90-DAY PLAN
Section 25 — Profit and funding separated, currency exposure and the benefits ledger.
How this is marked — the rubric
| Separates profitability from funding with a dated cash schedule, peak need and a recomputed delay scenario | 30 |
| Follows currency exposure to obligations and tests whether offsets are usable | 20 |
| Identifies shared portfolio dependencies and a coincident-peak downside | 20 |
| Starts a benefits ledger with baseline, counterfactual, category, costs and finance acceptance, without inventing a return | 30 |
| Total · pass mark 70 | 100 |
Submitting session 25
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: 25, 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 5 — Trust, Validation, People and Finance 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 5 →Closed book, no notes, and Coach Ritchie switches himself off while you sit it.Capabilities Covered
Funding versus profitPeak funding analysisCurrency exposurePortfolio dependenceBenefits ledgerHonest capital allocation