Level 300: Building with gen AI · Module 300.R

Five coworkers from different jobs talk around a round café table, a yellow hard hat, headset, ledger, lab coat and closed laptop among the coffee cups. The chalkboard reads “Bring your own job”.
Bring your own job

Building in your job (role tracks and prep rails) · Plain gear

Reviewed and published, version 1.0.0. As of 18 November 2026. It is not a credential. Big Widget Company and the people in it are made up. The real cases are real, and each one is cited. The tracks, the gate and the prep-rail rules are this course's own way of working, not a law.

What this part is for

Step 1, DEFINE · COURSE EXERCISE: Name one build. Retrieval or agent? Which documents or tools, and for whom? Step 2, HOLD · COURSE EXERCISE: Name the person and expiry. One person, by name, and the date their approval expires. No name, no run. Step 3, CHECK · COURSE EXERCISE: Check all eight gate parts. Record pass, fail or “don’t know yet”. For each unknown, name who to ask. Step 4, LIMIT · COURSE EXERCISE: Name a risk and a limit. Choose an OWASP risk and a control that holds even if the model is fooled.
Course track exercise using the OWASP risk list (s3). Use made-up examples only; the complete exercise also asks you to read another track’s row.
Figure data as text

Bring a build and a holder

  1. DEFINE · course exercise · Name one build — retrieval or agent? Which documents or tools, and for whom?
  2. HOLD · course exercise · Name the person and expiry — one person, by name, and the date their approval expires. No name, no run.
  3. CHECK · course exercise · Check all eight gate parts — record pass, fail or “don’t know yet”. For each unknown, name who to ask.
  4. LIMIT · course exercise · Name a risk and a limit — choose an OWASP risk and a control that holds even if the model is fooled.

Course track exercise using the OWASP risk list (s3).

Objective: 300.R-a. Text equivalent of figures/bring-a-build-and-a-holder.svg.

In Level 300 you learned to build two kinds of things:

You also learned the operate gate. Before a system runs on its own, one named person holds it.

Now you use all of that in your own job. Big Widget Company has eleven tracks, one for each kind of role. Pick yours. Then read one more, from a different team.

When you finish, you can do two things:

Some people think building with AI is only for programmers. It is not. The person who owns the work often knows best which files are private, which actions can't be undone, and who should check. A US government guide on AI risk, from the standards agency NIST, says people's roles in managing AI risk should be written down and clear across the whole organization.1 Not just in IT (information technology: the team that runs the company's computers and logins).

Where this sits in the three yeses. Lesson 100.9 taught three moves: Get to Yes, Stay at Yes and Recover to Yes.

A build, a holder, a gate

Every track has three parts.

For each action, the gate gives one of four answers: go ahead, only this far, wait for a person, or no.

Here is the gate. It has eight checks.

  1. Holder. Who holds it, and when does their yes run out?
  2. Allowed actions. What may it do alone? What needs a named person to approve first? What may it never do? "Never" means no, every time. Nobody can approve it, not even the holder. Anything not on the list is also a no.
  3. Second channel. For big moves you can't undo, like paying money, check on a second channel the asker does not control. Call back on a number you already have.
  4. Limit. How many actions per hour, per person and in total? What happens past the limit?
  5. Judgment. Where does the AI only help, while a trained person decides?
  6. Stop. What stops it? If something is missing, it stops. It does not carry on.
  7. Receipt. A short record of each decision: what, who held it, which approval, the result, and how to undo it. It holds no private details: no customer names, no document text, no prompts or AI answers.
  8. Count. Everything the build really did must have gone through the gate first. Check one by one: each action in the build's own record should match a gate answer from just before it. Use the gate's latest answer, not an older one. If the gate said yes, then later said no, the no is the answer that counts. The old yes stays written down, but it no longer lets anything run. Then say how many matched, out of how many it did. Two totals that happen to be equal are not enough. Things the gate said no to don't go in the count, because they never happened. If one happened anyway, that is a bigger problem, and the build stops. If you can't check yet, say "can't count yet". That is the honest answer, and it means the build waits.

Each track marks the two or three checks most likely to go wrong. But all eight must pass. "Don't know yet" is always allowed, but it always comes with a name: who will find out. Until then the build waits, and the holder decides what happens next. If nobody holds it, it does not run.

A line in the AI's instructions that says "never pay" does not stop it paying. Not giving it a way to pay does. The gate decides. A person holds the build. The AI decides nothing and signs nothing.

Why one name? Picture an alarm at 2 a.m. It can ring the team's on-call phone. That is fine. But someone has to be the one who decides: stop it, or let it run again. "The team owns it" does not say who. So this course names one holder, with an end date, and writes down who the on-call person calls. NIST's guide says the people who answer for AI risk should be clear to everyone, both people and teams, and that leaders take responsibility for decisions about it.1 It does not say "only one name". That rule is ours. NIST's guide is something companies choose to follow, not a law.2

Why your job? A security group called OWASP lists the top ten risks for AI apps like these. Some of them are: hidden instructions that trick the AI, private data leaking out, and an AI that can do too much.3 All of those depend on what the build may read and do. You know that better than anyone.

Three true stories behind the tracks

A fake video meeting

In 2024, the Hong Kong Government told its lawmakers about a fraud. Someone used deepfakes to make a fake, pre-recorded video meeting.4 The faker pretended to be the company's top finance boss. The worker on the call sent money to five bank accounts. The loss was about 200 million Hong Kong dollars.4 The record does not name the company, and neither do we.

The lesson: a face on a screen is not proof. If a build can move money, it needs a second channel.

Cigna and the 1.2 seconds

In 2023, people sued the insurer Cigna. Their court papers say that in two months of 2022, Cigna's doctors turned down over 300,000 payment requests. The papers say that, on average, each one got about 1.2 seconds.5 These are claims in a lawsuit, not a court's findings. This lesson does not say how the case turned out: we have not checked. The papers got the numbers from a news story.5

The lesson: lots of small decisions add up. A build that decides many things needs a limit. A very quick sign-off is a warning sign, but it is not proof on its own. Someone might read for two minutes, then sign in one second. So look at what the person really read, and whether they could say no.

nH Predict

Also in 2023, the estates of two patients sued the insurer UnitedHealth.6 (An estate is what a person leaves behind when they die. The people who manage it can go to court for them.) Their court papers say it used an AI tool called nH Predict to overrule patients' own doctors about care. The papers claim the tool has a "90% error rate".6 These are claims, not findings.6 Lesson 300.8 also shows what the company said back.

The lesson: where a trained person's judgment is the decision, the AI may help. It may not take over.

The eleven tracks

The builds and the people are made up. Each track names its riskiest checks. All eight checks still apply.

Three rules for every track:

A date to know. As of 18 November 2026, the European Commission says EU rules for high-risk uses like hiring tools start on 2 December 2027.9 Dates like this have changed before. Lesson 300.7 shows how to check.

Your track exercise

You will bring this to the project, 300.P. A person reviews it there. An AI does not.

  1. Pick one build from your track, or make up a new one for your job.
  2. Say in one line what it is: retrieval or agent, over which files or with which tools, and for whom.
  3. Name the holder: one person, and the date their yes runs out. If you can't name anyone, it does not run. Write down who you will ask.
  4. Go through the eight checks. For each, write pass, fail or "don't know yet". Circle the two or three riskiest. For each "don't know yet", write who could find out.
  5. Pick the OWASP risk most likely for your build.3 Write one limit that still works even if the AI is tricked.
  6. Read one other team's track. Could their build touch your files or your decisions? Who would you tell? If you can't find a way, say so. Don't make one up.

Use made-up examples only. Never paste real files, logs, prompts or private details. If your workplace lets you keep notes on a real prompt or answer, they stay in your workplace's own records, never in a receipt or anything you share.

Prep rails

From Level 300 on, the course has prep rails. A prep rail is practice that lines up our lessons with the topics of a company's certification exam. The company runs the exam and watches you take it. That is called a proctored exam.

Level 300 has two rails. As of 18 November 2026, we do not name the companies or exams yet. Bob Rapp has to clear each name first.

What a prep rail does. It shows which lessons match which exam topics. It gives you practice questions in our own words. It shows where you are weak before you book.

What a prep rail does not do. These are the course's rules.

A person reviews your Level 300 project, and a host marks it. The AI marks nothing. If you earn the Level 300 badge, it shows you finished our level and project. It is about what you can do, not proof that any system is safe. It is not the company's certificate.

The Foundation already says the same about university courses it lists. Its page says those courses are not degrees, that "none of them gives credit or a certificate through us", and that no university designed or endorses them.10 A prep rail works the same way. It is real learning toward someone else's certificate. Only they can give it.

What this lesson does not say

Words to know

Want to learn more?

Extra reading for this lesson is on hold. A link to an outside course can carry company and exam names, and Bob Rapp clears those first. Until then, the sources below are your reading list.

Sources

  1. s2 National Institute of Standards and Technology (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. Standard. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩ ↩
  2. s1 National Institute of Standards and Technology (2023). AI Risk Management Framework. Official doc. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩
  3. s3 OWASP GenAI Security Project (2025). OWASP Top 10 for LLM Applications 2025. Official doc. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩ ↩ ↩
  4. s8 Government of the Hong Kong Special Administrative Region (2024). LCQ9: Combating frauds involving deepfake. Record. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩ ↩ ↩ ↩ ↩
  5. s9 United States District Court for the Eastern District of California (plaintiffs' pleading) (2023). Kisting-Leung v. Cigna Corp., No. 2:23-cv-01477 (E.D. Cal.), class action complaint, filed 24 July 2023. Record. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩ ↩ ↩ ↩
  6. s10 United States District Court for the District of Minnesota (plaintiffs' pleading) (2023). Estate of Gene B. Lokken v. UnitedHealth Group, Inc., No. 0:23-cv-03514 (D. Minn.), class action complaint, filed 14 November 2023. Record. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩ ↩ ↩ ↩ ↩
  7. s4 OWASP GenAI Security Project (2025). LLM06:2025 Excessive Agency. Official doc. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩
  8. s5 OWASP GenAI Security Project (2025). LLM01:2025 Prompt Injection. Official doc. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩
  9. s7 European Commission (2026). AI Act (regulatory framework for AI). Official doc. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩ ↩ ↩ ↩
  10. s6 AiGovOps Foundation (2026). University AI courses (Foundation learning page). Official doc. Checked 2026-11-19 by source-checker (separate session sc300R-r3, round 3 re-check). ↩

Next: the quiz