From policy to control
- Lesson 14
- 30 min
- Module 5: Controls, documentation and evidence
Where we are
In Module 4, we wrote Larkfield’s rules for staff and for building and buying AI. People can now read what is allowed. But a policy line alone stops nothing: someone or something must enforce it, and a record must prove that it worked. In this lesson, we turn each policy line into a control with an owner and a proof, starting back at the swimming pool.
Lesson 14 of 19 · AI Governance: Policies, Risk and AccountabilityFrom policy to control
1. What governance is for
2. Laws and standards in plain words
3. Inventory, risk tiers and impact assessment
4. Policies people can follow
5. Controls, documentation and evidence
What you will be able to do
Make each line of an AI policy true in practice. For each line, name the control that enforces it, the person who owns it and the record that proves it worked.
Before you startWhat to know and to have ready
What you need to know
- Larkfield’s rules for staff, from Acceptable use for staff.
- Which controls each risk tier needs, from Risk tiers and proportionate controls.
What you need
- The control map template and its spreadsheet copy, control-map.csv.
- The evidence recorded by earlier courses.
- Your rewritten acceptable-use policy from Module 4. If you do not have it, use the six lines in this lesson.
One rule at a swimming pool
Section titled “One rule at a swimming pool”Go back to the pool from Acceptable use for staff. Its sign says: “No running near the water.” The sign makes the rule clear. But the sign alone stops nobody. So the pool uses three other things.
- The building: a non-slip floor, and a gate that keeps small children out of the deep end. They work even when nobody watches.
- A routine: a lifeguard watches, blows a whistle and writes each problem in a logbook.
- Lessons: swimming teachers tell children why running is dangerous.
When the city’s inspector visits, the sign proves nothing. The inspector reads the lifeguard’s logbook and tests the floor. Those records exist because the work happened, not because somebody wrote them for the visit.
The problem: a line that nothing enforces
Section titled “The problem: a line that nothing enforces”Larkfield’s first AI policy, as Farida found it, has the line “AI output must be checked before use”. Every member of staff agrees with it. Nothing makes it true.
In week 3 of the refund workflow’s pilot, a team lead approved 212 proposals. The team lead opened 0 of them before approving. The median time to approve was 3 seconds. One German-language ticket about a damaged item got a replacement. Grace’s refund policy says that this kind of ticket goes to a person. This situation is incident INC-07, constructed for this course. It is built on a recorded behaviour from AI Agents and Workflow Orchestration. There, one design made 3 changes that nobody asked for in 2 of 70 tickets when every proposal was approved as it was.
Every approval of that week was recorded, with a name and a time. The records looked perfect. Grace found the problem in her weekly sample: she opens ten approved actions and checks each one. The policy line was there, and the platform recorded every click. Only Grace’s routine checked that someone had read the ticket.
The idea: a control, its owner and its evidence
Section titled “The idea: a control, its owner and its evidence”In Risk tiers and proportionate controls, a control was anything that reduces a risk. Here, we look at a control from the policy’s side: it is what makes a policy line true. It comes in three kinds, like the pool:
- A platform control is enforced by a system. For example, the support MCP server refuses an approval from someone without the owner role.
- A process is a routine that people follow, with a record. For example, Grace’s weekly sample.
- Training tells people the rule and why it exists.
Each control has an owner: a named person who answers for it. And each control leaves evidence: a record, made by the work itself, that shows the control worked. A log, a registry entry or a signed decision is evidence. A document written for the audit is only a description.
A control map is a table with one row per policy line: the control, its kind, its owner, its evidence and how often somebody checks it.
Text description of the diagram
The diagram shows one policy line at the top: “A person checks before an AI action goes out.” Three arrows lead to three kinds of control. The platform control: the MCP server refuses an approval by anyone without the owner role. Its record proves that every call was checked, but not that a person read the ticket. The process: Grace’s weekly sample of ten approved actions. Its record proves that someone checked a sample and wrote findings; it costs staff time each week. Training: every member of staff learns the rule. Its record proves who attended, not what they do at the desk.
A tiny example: the refund line, step by step
Section titled “A tiny example: the refund line, step by step”The control map template has one worked row. Here it is, one step at a time, with the value at each step.
Text description of the diagram
The diagram shows six boxes in a row. Box 1, the policy line: “A person approves every refund that an AI system proposes.” Box 2, what must be true: no refund runs without an approval by a person whose limit covers the amount. Box 3, the control: the workflow only proposes, the MCP server refuses approvals without the owner role, and Grace samples approved actions. Box 4, the kind: platform and process. Box 5, the owners: Grace for the policy and Kwame for the platform. Box 6, the evidence and the check: the approval record of each refund, and a weekly sample of ten approvals.
- Policy line. “A person approves every refund that an AI system proposes.”
- What must be true. No refund runs without an approval. The approver’s limit covers the amount.
- Control. The workflow can only propose. The support MCP server checks every approval itself. In MCP: Connect AI Applications to Tools and Data, it received 72 calls and refused 54 (recorded). Of those, 7 were refused because of the caller’s role.
- Kind. Platform. But step 6 shows that the platform alone is not enough, so a process is added.
- Owner. Grace owns the policy line. Kwame owns the platform control.
- Evidence. Each refund leaves an approval record: who, when and the exact action. A recorded example: Sam proposed OP-0001, a refund of 25.00 dollars for ticket T-30002. When Sam tried to approve it, the server answered “only an owner of larkfield can decide”. Grace then approved it, and both answers were recorded. How often checked: a weekly sample of ten approvals, because a record of a click does not show that the ticket was read.
The recorded results show why the approval matters. In the agents course, the design that made 3 unwanted changes made 0 with a strict approver. The approval step is the control that stops them. A person who approves everything switches it off.
Which kind of control, and when
Section titled “Which kind of control, and when”| Kind | Good for | Cannot do | Its evidence | Cost |
|---|---|---|---|---|
| Platform | facts a system can check: who, how much, which data, which tool | judge meaning, or see AI that goes around the platform | logs and refusals, for every call | build once (platform team), then little |
| Process | judgment: did the person read it, is the answer right? | run on every case; it samples | a sample sheet with dates and findings | staff time, every week or month |
| Training | telling people the rule and why | prove behaviour | attendance and test results | a few hours per person each year |
Two rules of thumb follow. First, use a platform control wherever a machine can check a fact, because it works on every case. Second, never let training be the only control for a line that protects customers. At Larkfield, 42% of staff have done the AI training (constructed). That record proves who attended. It does not show what anyone does with a long queue.
The risk tier also sets the strength. A low-tier tool can live with a rule and training. A medium or high system needs a platform control and a process on its riskiest lines.
Hand the technical controls to the platform team
Section titled “Hand the technical controls to the platform team”Farida does not build these controls. Kwame’s platform team does. They include the registry, the gateway, budgets, tool permissions and kill switches. The gateway is the one route through which registered systems call a model, so it can check and record each call. A kill switch stops a system at once. A course on governing AI agents teaches them.
Text description of the diagram
The diagram shows two sides. On the left, governance: Farida and the accountable owners. They own the policy line, what must be true, the owner of each control, the evidence needed and how often it is checked. On the right, the platform team: Kwame. It owns the registry, the gateway, budgets, tool permissions and kill switches. An arrow goes from governance to the platform: a control request that says what must be true and what evidence is needed. An arrow comes back: records made by the work, such as registry entries, gateway refusals and approval logs.
So Farida’s request to Kwame is short. It says what must be true, the evidence she needs and how often. It does not say how to build it. If Kwame cannot produce the evidence, the control does not exist yet, and the row says so.
Discuss: “We trained everyone”
Section titled “Discuss: “We trained everyone””Farida shows the empty control map at the team meeting.
- Tomás: “We sent the policy to everyone and ran a training. So every line is enforced.”
- Kwame: “I can block every AI website on company laptops. Then nobody pastes customer data anywhere.”
- Omar: “A weekly sample of approvals is slow. The platform records every approval anyway.”
- Sam: “How do we choose the kind of control for each line?”
Before you read on, answer these:
- What does the training record prove, and what does it not prove?
- What would Kwame’s block stop, and what would it miss?
- Is Omar right that the approval records are enough?
The course’s answer. Training is needed, but it proves attendance, not behaviour. Kwame’s block is a good platform control for the best-known free chatbots. It misses tools on phones and AI inside the tools Larkfield buys. And blocking everything pushes people to work around it. So block the free chatbots and keep a list of approved tools. Omar is wrong about the records: pilot week 3 had a perfect record of 212 clicks. The sample is the only control that checks reading.
What would change the answer: suppose the workflow showed each ticket to the approver and recorded that they opened it. Then the platform could prove one more fact, and the sample could be smaller. It could not disappear: opening a ticket is still not reading it.
When to build a control map, and what it costs
Section titled “When to build a control map, and what it costs”Build the control map right after the policy is approved, and before the first audit or board review. For a one-page policy, it takes an afternoon with each line’s owner. Update the row when the policy line or the system changes.
Do not map every sentence of every document. Map the lines that protect someone: customers, staff or the company’s data. A row with no evidence is still useful. It shows a gap that you know about, instead of one that an auditor finds.
Common mistake
A control with no evidence
Symptom: the control map says “Training” or “Managers check” for a line, and the evidence column says “policy sent to all staff”.
Cause: the row describes an intention. Nothing in the work leaves a record that the control happened.
Fix: for each row, ask: “If an auditor asks for the last three months, which record do we show?” If the answer is a document written today, the control leaves no record. Add a process with a record, such as a sample sheet, or a platform control with a log.
Guided practice: the data line
Section titled “Guided practice: the data line”Use line 3 of the rewritten policy: “Never put customer or sensitive data into a tool that is not on the list.”
-
Open your copy of
control-map.csv. The refund row is already filled in. -
Write the line in the first column. Write what must be true: no customer data reaches an AI service that is not on the list.
-
Add a platform control: Kwame’s gateway checks the data class of each call from a registered system.
Check: the gateway sees only the 4 of 8 systems in the registry. The weekly digest script uses its own key, so the gateway never sees it.
-
Add a second control for what the gateway misses: a monthly check of provider bills and company cards against the registry.
-
Add training, and block the best-known free chatbots on company laptops.
-
Write the owner: Mei, who owns the data records.
-
Write the evidence: the gateway’s refusals, the monthly bill check with its findings, and training completion.
The bill check is what found the digest script. Without it, the line would look enforced while the biggest gap stayed open.
Your turn: the control map for Larkfield’s policy
Section titled “Your turn: the control map for Larkfield’s policy”Map every line of the rewritten policy. If you rewrote it yourself in Module 4, use your own lines. Otherwise, use this version (constructed):
- Use the approved tools on the list.
- Know the four data classes: public, internal, customer and sensitive.
- Never put customer or sensitive data into a tool that is not on the list.
- A person checks before anything goes out, and never approves what they have not opened.
- A new tool or a new AI feature: ask the risk lead first.
- Report a mistake to your team lead and the risk lead on the same day.
For each line, fill in the control, the kind, the owner, the evidence and how often it is checked. Mark any row where the evidence does not exist yet.
One possible answer
- 1. Approved tools: the gateway accepts only registered systems, and the list is published (platform and process). Owners: Kwame and Farida. Evidence: registry entries, the gateway’s refusals of unregistered keys, and bills checked against the registry. Monthly.
- 2 and 3. Data: a data-class check at the gateway, a block of the best-known free chatbots and training (platform and training). Owner: Mei. Evidence: the gateway’s refused calls, the bill check and training completion (42% today, constructed). Monthly.
- 4. A person checks: the approval step in the workflow and the MCP server, and the weekly sample (platform and process). Owner: Grace. Evidence: approval records and the sample sheet with findings. Weekly.
- 5. Ask first: the “New AI use” form; contract owners report AI features that arrive in updates (process). Owner: Farida. Evidence: the request log with dates and decisions, and the inventory. Each request.
- 6. Report mistakes: an incident form, read first at each board meeting (process). Owner: Farida. Evidence: incident records with times. Each incident.
Knowledge check
Section titled “Knowledge check”Check your understanding
In pilot week 3, every approval was recorded with a name and a time. Why did these records not show the problem, and which control found it?
Show the answer
The approval record proves that a person clicked, at a time. It does not prove that the person opened or read the ticket: 212 approvals were opened 0 times. Grace’s weekly sample of ten approved actions found it, because a person opened each action and checked it against the policy.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
For each control, name its kind (platform, process or training): (a) the MCP server refuses an approval without the owner role; (b) a monthly check of provider bills against the registry; (c) a course on the data classes for all staff.
Show the answer
(a) platform: a system enforces it on every call. (b) process: people follow a routine and record what they find. (c) training: it tells people the rule and why, and its record shows only who attended.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Farida asks Kwame for a control on the line 'use the approved tools'. What should her request say, and what should it leave to Kwame?
Show the answer
It should say what must be true: only registered systems call a model through the platform. It should name the evidence: registry entries and the gateway’s refusals of unregistered keys. And it should say how often she checks it: monthly. How to build the gateway rule is Kwame’s job. She should also say what the gateway cannot see, such as a script with its own key, so that a second control covers it.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Recap
- A control makes a policy line true. It is a platform control, a process or training, and it has a named owner.
- Evidence is a record made by the work: a log, a registry entry, a sample sheet. A document written for the audit is a description.
- Use a platform control where a system can check a fact, and a process where judgment is needed. Training alone never protects customers.
- Governance says what must be true and what proves it. The platform team builds how.
You can now say who enforces each policy line and what proves it. But many of those proofs are documents, and a document that nobody updates soon says something false. Next, you see which documents an AI system needs and how to keep them current without extra work.
Next: Documentation