Lesson 6 · AI Foundations

Practical AI scenario and knowledge check

Length: 29 minutes across 7 sections
You will be able to apply: The Six Blanks
You will produce: Six Blanks Decision Dossier · Commit Statement
You will work: 2 gated questions

Reading

Which line is still blank?
Core question

Graduate practice

Signature experience outcomes

Synthesize the course frameworks into one defensible AI-use decision.
Distinguish verified evidence from assumptions, unknowns, and fluent claims.
Defend a delegation boundary against a credible counterargument.
Specify an observable stop condition and a named owner.

Four instruments so far. The Reversibility Gate decides what may be delegated. The Envelope Test finds where the capability holds. The Six Slots make the instruction transferable. The Evidence Contract settles the claims. The Three-Way Match protects the data. Each is useful alone. Run in sequence against one real task, they become a single discipline you can complete in under fifteen minutes.

This lesson assembles them. It also does the thing that matters more than assembly: it teaches you to read your own incomplete answer as a location, not as an inconvenience.

Next: The pathology: Selective Rigour

Section 2 of 7

The pathology: Selective Rigour

Watch what people actually do with a discipline they have accepted. They apply it to the tasks where applying it is easy, where the answers are obvious, where nothing is urgent and nobody is waiting. Then they reach the task that is complicated, contested and late, and they skip it. Not consciously. It simply does not occur to them, because the task is already hard enough.

PrincipleSelective Rigour is the reflex of applying a control most thoroughly where it matters least. It produces an organisation with impeccable documentation of its trivial AI uses and nothing at all on the one that will make the news. The tell is comfort: if running the discipline felt straightforward, you probably ran it on something that did not need it.

The counter is to treat difficulty as a signal to slow down rather than to skip, and to treat the blank line as the finding.

Next: The Six Blanks

Section 3 of 7Framework

The Six Blanks

Definition

For any AI-assisted task, six lines must be complete before it runs: the task and its delegation verdict, the data boundary, the prompt contract, the named reviewer, the settling evidence, and the stop condition. Six lines, one page, one system. A completed page means the task is ready. An incomplete page is not a failure of the exercise: the blank line is the precise location of the risk, and it is the only output of the exercise that tells you anything you did not already know.

The stop condition is the line most often left empty and the one that matters most under pressure. It states, in advance, the observation that ends the use.

When to Use It

Use it before a task moves from experiment to routine, before another person starts running it, and before an existing use expands in scope, volume or audience. Fifteen minutes at those three moments is worth more than any amount of review afterwards.

Use it hardest on the task you least want to spend fifteen minutes on. That reluctance is diagnostic.

How to Apply It

Fill the six lines in order, and leave a line visibly blank rather than writing something vague to complete the page.
Write the reviewer as a person, not a role, and confirm they know the task exists.
Write the stop condition as an observation somebody could make on a normal Tuesday, not as a catastrophe.
Treat every blank as a task with an owner and a date, and do not run the use until the blanks are closed.

Interactive modelThe Six Blanksstack · 6 elements

Stop condition

An observation that ends the use, written before it starts.

Each line draws on a specific instrument and is complete only when it meets a concrete standard.

Stop condition: An observation that ends the use, written before it starts.
Settling evidence: Every claim class has an artifact, or is deleted.
Named reviewer: A person, not a role, who is not the generator.
Prompt contract: A colleague can run it without asking you anything.
Data boundary: All three legs hold for the account actually in use.
Task and verdict: Undo cost, payer and verdict are all written down.

Line	Instrument	Complete when
Task and verdict	The Reversibility Gate	Undo cost, payer and verdict are all written down
Data boundary	The Three-Way Match	All three legs hold for the account actually in use
Prompt contract	The Six Slots	A colleague can run it without asking you anything
Named reviewer	The Evidence Contract	A person, not a role, who is not the generator
Settling evidence	The Evidence Contract	Every claim class has an artifact, or is deleted
Stop condition	The Envelope Test	An observation that ends the use, written before it starts

Worked example 1 of 3

A marketing team at OmniCorp Retail proposes drafting campaign variants from approved product facts. Trevor Okafor runs the six blanks with them in a room, on one page, in twelve minutes.

Trevor Okafor: Task and verdict.
Campaign manager: Draft five variants per campaign. Undo cost is a wrong claim reaching the public. Payer is the customer, then legal. Assist.
Trevor Okafor: Data boundary.
Campaign manager: Approved product facts only, enterprise account, marketing purpose. All three hold.
Trevor Okafor: Reviewer.
Campaign manager: Whoever is on rota that week.
Trevor Okafor: That is a blank, not an answer. And stop condition?
Campaign manager: We had not thought about that one.

Four lines complete, two blank, twelve minutes spent. The team went away with two tasks rather than an approval: name the reviewer, and write the stop condition. They came back with both the following morning. The stop condition they wrote was specific and low: two variants in one week containing a product claim not present in the approved facts pauses the workflow.

Why This Works

A blank is unarguable in a way that a concern is not. Saying a workflow feels risky invites a debate about risk appetite. Pointing at an empty line labelled reviewer invites a name. The page converts judgment into completion, which is the only form of rigour that survives a busy team.

Writing the stop condition before starting works because it is written by a person who is not yet invested. The same person, three months in, with the workflow embedded and a team depending on it, will never set that threshold as low as they set it on day one.

Worked example 2 of 3Optional depth

Alan Brixmoor at OmniCorp Public ran the six blanks against an AI query-classification workflow that had been in production for seven months. Five lines completed easily from existing practice. The stop condition line was blank, and filling it exposed something worse: nobody was measuring anything that could have triggered a stop. The workflow had been running unobserved since launch. Closing that one blank took a fortnight and changed the workflow from unmonitored to governed, without changing a line of the prompt.

Worked example 3 of 3Optional depth

Jo Halvorsen at OmniCorp Studio applied the page to six AI uses across her eleven-person practice in a single afternoon. Five completed. The sixth, drafting client-facing scope language, failed on settling evidence, because the only artifact that could settle a scope claim was the pricing worksheet and nobody had said so. That was the same finding she had reached through the Reversibility Gate months earlier and never written down. The page caught what her memory had not retained, which is the entire argument for the page.

Edge Cases and NuancesOptional depth

A page can be complete and wrong, if the reviewer is named but has no time, or the stop condition is set so high it can never be observed. Completeness is necessary, not sufficient. Some tasks genuinely have no stop condition because they cannot be stopped without stopping the business, and that is a finding to escalate rather than a line to fill. And a page written for one task does not cover the adjacent one, however similar it looks: adjacency is where every one of these instruments fails, and it fails there because adjacency is invisible from inside the task.

Interactive modelBlank line versus vague entrycontrast · 2 elements

Blank line

Identifies the precise location of the risk and becomes a task with an owner and a date.

A blank is a finding; a vague entry is a false certification that hides the risk.

Blank line: Identifies the precise location of the risk and becomes a task with an owner and a date.
Vague entry: Certifies readiness falsely and misleads the next person who reads the page.

Knowledge check

A team completes the Six Blanks page for an AI workflow but writes the team on the reviewer line and will be monitored on the stop-condition line. Is the page complete?

Answer first, then check.

Next: Common Failure Modes

Section 4 of 7

Common Failure Modes

Failure modeCompleting the page rather than the thinking. What it looks like in the moment: you reach the reviewer line, you do not have a name, and you write the team rather than leave it empty. It takes half a second and it feels like progress. You do the same on the stop condition, writing that the use will be monitored. The cost when this happens: the page now certifies that the task is ready, and the certification is false in exactly the two places it mattered, so the next person to look at it is misled rather than informed. The correction: a blank line is a finding. Writing something into it to avoid the blank destroys the only value the exercise has.

Failure modeRunning it where it is easy. What it looks like in the moment: you complete pages for four internal drafting tasks in an hour, feel the discipline working, and leave the customer-facing workflow for when there is more time. There never is more time, and the workflow was the reason you started. The cost when this happens: the organisation acquires documentation that is inversely correlated with risk, and the gap is invisible because the folder looks full. The correction: order the queue by who pays when it goes wrong, and start at the top. If the first page you complete was comfortable, you started in the wrong place.

Next: The page end to end

Section 5 of 7

The page end to end

Priya Raghunathan at OmniCorp Logistics applied the six blanks to the workflow she least wanted to examine: AI-drafted customer notifications for delayed shipments, running for five months, popular with the regional teams, and the single highest-volume AI use in the operation.

Task and verdict: Assist, undo cost a customer told the wrong delay reason, payer the customer. Complete. Data boundary: shipment records and customer contact details on the enterprise account, in region. Complete, after she discovered and removed an integration to a shared folder nobody had approved. Prompt contract: the prompt existed but lived in one person's saved messages, and three regions had each modified their own copy. Blank. Named reviewer: none, because the notifications sent automatically. Blank. Settling evidence: identifiers and reason codes against the dispatch record, which nobody was checking because there was no reviewer. Blank by consequence. Stop condition: none.

Two lines complete, four blank, on the workflow everyone regarded as the success story. Priya did not shut it down. She published the page, assigned each blank an owner and a date, and set a temporary volume cap while they were closed. Four weeks later all six lines were complete, the three regional prompt variants had become one, and the stop condition was live: three notifications in a week with a reason code that does not match the dispatch record pauses automatic sending. Nothing about the workflow's value had changed. Everything about its defensibility had.

Decision point

Dr. Naomi Ellery at OmniCorp Health has fifteen minutes and two AI uses to review. The first is an internal meeting-summary tool used by her administrative team, easy to assess and certain to complete cleanly. The second is a clinical letter drafting pilot with contested scope, an unclear reviewer arrangement and a consultant who is impatient with the question. Which do you take, and what do you expect to produce?

Write your answer before you continue

Confidence before seeing the analysis

What evidence or assumption could change your decision?

Commit, calibrate, and name contrary evidence first.

Next: Self-check

Section 6 of 7

Self-check

Knowledge checkList every AI use you are responsible for, ordered by who pays when it goes wrong, worst first. Complete the six blanks for the one at the top. Count your blanks. That number, not the number of completed pages, is what you actually learned today.

Mark the level that describes you today. Nothing is submitted.

Behaviour	Ready	Developing	Not yet
Choosing what to examine	Given the OmniCorp Health pair, takes the contested clinical pilot and accepts an incomplete page as the output.	Given the same pair, takes the clinical pilot but stops when the consultant pushes back.	Given the same pair, takes the meeting-summary tool to demonstrate the method.
Handling a blank line	Leaves it blank, assigns an owner and a date, and caps the use until it is closed.	Leaves it blank and raises it, without an owner or a constraint on the use.	Writes a plausible entry so the page reads as complete.
Writing a stop condition	States an observation a person could make in a normal week, with a threshold low enough to trigger before harm.	States a condition that would trigger only after visible harm.	States that the use will be monitored.

Signature learning object

Six Blanks Decision Dossier

0/6 complete

Build a one-page decision record for a real or fictional AI use. Write for a skeptical reviewer who was not present. Preserve uncertainty rather than filling it with confidence. Your completed dossier can be copied or printed as a portfolio object.

Decision and bounded use

Affected party and undo cost

Evidence that supports the decision

Material unknowns and assumptions

Strongest counterargument and your response

Stop condition, owner, and review date

Next: Commit

Section 7 of 7

Commit

Commit Statement

Complete every line in your own words, then sign and date it. Attach the page for your highest-stakes AI use, blanks included.

I commit to

I will stop

I will start

I will know I succeeded when

Signed Date

Window	Field application
Days 1 to 7	Order every AI use you touch by who pays when it fails. Complete the six blanks for the one at the top, blanks and all.
Days 8 to 21	Assign an owner and a date to every blank, and cap scope or volume until they close.
Days 22 to 30	Complete the page for the second and third uses on the list, and check whether the first one's stop condition has ever been observable.

Six blanks per system is governance you can run alone, on one page, for the uses you personally control. It does not scale to a portfolio: it has no register, no decision rights, no independent assurance, no evidence retention and no way to answer a regulator asking who approved what and when. Turning a page you can complete into an operating model an organisation can be held to is the capability the paid programs develop next.

DisclaimerGeneral guidance only. All organisations and people named in this lesson are fictional. Regulated organisations should confirm requirements with a qualified professional before relying on this material.

Required practice must be complete.