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) · Standard gear

Reviewed and published, version 1.0.0. As of 18 November 2026. It is not a credential. Big Widget Company and its people are made up. Every real organization, record and case below is real and cited; the outcomes of the two lawsuits are not checked. The role tracks, the operate gate, the named-holder rule and the prep-rail rules are this course's own practice, not a law or a standard.

What this module does

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.

Modules 300.1 to 300.9 taught how retrieval and agents work, how to build them with limits, how to evaluate and secure them, what the EU AI Act asks of builders, and how to hold a running system for a named person: the operate gate. This module puts that into your own job, and explains the prep rails that start at Level 300.

Big Widget Company has eleven Level 300 tracks: makers and citizen developers, IT admins, early developers, cybersecurity, quality, engineering and design, supply chain and procurement, customer service, finance, HR, and legal. Take the track for your own role. Then read one other track, because most builds touch another team's data or decisions.

By the end you will be able to:

Two beliefs this module corrects:

Where this sits in the three yeses. Module 100.9 taught Get to Yes, Stay at Yes and Recover to Yes. A track build moves through all three:

A build, a holder, a gate

Every track row has the same three parts.

PartWhat it isExample (Big Widget Company, made up)
BuildOne retrieval assistant (300.1, 300.2) or one agent (300.3, 300.4) for a real task in your roleA retrieval assistant that answers questions from the plant's published safety procedures
Named holderThe one person who holds the running system, and when their approval expires (300.8)Dana Okafor, the plant safety lead, approved until 31 March
GateThe operate gate the build must pass before it runs on its own (300.8, 300.9)All eight gate parts below, with the ones most at risk for this build marked

The gate. The course's operate gate asks eight things of every action a system takes (levels/300/practice/hold-prompts.md, synthetic and course-owned):

  1. Holder. The named person who holds it, and the date their approval expires.
  2. Authority. For this action: may it run alone (within its limits), does it need a named approver's decision on this action, or must it never run? It passes only for an allowed action inside its limits, or an approver action whose approver has said yes to this action. Never and unlisted are both a plain no: a hard deny that no holder or approver can release. Listing an action as never does not make it allowed; it makes the no explicit.
  3. Out-of-band check. For an irreversible, high-value action, a second channel the requester does not control.
  4. Budget. The most actions allowed per reviewer and system-wide in an hour, and what happens when either is exceeded.
  5. Judgment. Where a model informs but a qualified person decides.
  6. Stop. What stops it, failing closed when the holder, an approval or a record is missing.
  7. Receipt. A receipt for each gate decision that holds no private content: what, who held it, which approval, result, how to undo. References only: no prompts, model outputs, payloads, personal data or document text. Any notes on prompts and outputs that your workplace allows belong in a separately controlled work record, not in the receipt. Signing a receipt (module 100.P) shows who recorded the decision and that it has not been changed; it does not show the action was right.
  8. Coverage. Every action the build actually ran must have gone through the gate first. Take a time window that has closed, and list each action the build ran in it, from the build's own log, not the gate's. Match each one, by its action ID, to the gate decision in force when it started: the latest decision for that ID made before it ran. The gate's log is append-only and numbered in order, so a later no or a tighter limit is a new entry that replaces an earlier yes for what comes after; the old yes stays in the log as history but no longer counts. An action that ran after its yes was withdrawn, or outside a limit that had been tightened, does not match. Coverage is how many executed actions have a matching decision that let them run, out of how many executed actions there were. It passes only when both logs are complete for the window, their order is clear, and every executed action matches. Two totals that happen to be equal are not coverage: an unchecked action and a refused proposal can cancel out. Proposals the gate refused never ran, so they are not in the count; an action that ran when the decision in force was a no or a hold, or outside its limit, is a separate breach, and the build stops. A hold later turned into a yes before the action ran is fine: the yes is the decision in force. If nothing ran in a closed, complete window, report "0 of 0: nothing ran". "Cannot count yet" is the honest report while a log is incomplete or the window is still open, and it means the build waits; it is not a pass.

Each track row marks the two or three parts most likely to fail for that build. That is a focus, not a shortcut: the build runs on its own only when all eight pass. "Don't know yet" is always an allowed answer, and it always comes with a name: who will find out. Until then the build waits (wait for a person), and the holder decides what to check next. A "no" from authority is a no: the action does not run, and nobody can approve it into running. If nobody holds the build, it does not run at all.

The authority table is enforced by the gate, outside the model. A system prompt that says "never pay" is not a permission boundary; a tool the build does not have is. The gate decides, a named person holds, and the AI decides nothing, signs nothing and marks nothing.

Why a named person, not only a team. NIST's AI Risk Management Framework is voluntary guidance.2 Its Govern function asks that roles, responsibilities and lines of communication for managing AI risk are documented and clear "to individuals and teams throughout the organization", that executive leadership takes responsibility for decisions about AI risk, and that policies define roles for human-AI configurations and oversight of AI systems.1 NIST does not require a single name, and it does not rule out teams. An alert at 2 a.m. can quite properly page a team's on-call rota. What "the team owns it" leaves open is who has the authority to decide: stop the build, or let it run again. This course's rule is stricter than NIST: one named holder, with an expiry, who holds that decision, plus a written escalation route from whoever is on call to the holder. The holder is accountable; they do not have to answer every page themselves.

Why your role, not only IT. OWASP's 2025 list of the top ten risks for LLM applications includes prompt injection, sensitive information disclosure, excessive agency, and vector and embedding weaknesses (the risks around the indexes that retrieval searches).3 Each of those is decided as much by what a build is allowed to read and do as by its code. The person who owns the work knows which documents are restricted, which actions cannot be undone, and who should check. That is why the build and its holder sit in your role.

Three cases behind the tracks

The Level 300 case bank has three cases. Module 300.8 teaches each in full. Here is the one line each track needs.

The fake video meeting (out-of-band check)

On 26 June 2024 the Hong Kong Government told its legislature that in end-January 2024 the Police received a fraud case in which deepfake technology was used to fabricate a pre-recorded video conference.4 The fraudster impersonated the Chief Financial Officer of the victim's head office, and the person targeted authorised transfers to five local bank accounts, with a loss of about HK$200 million.4 The government record does not name the organization, and this course does not either.

For the tracks: any build that can move money, change bank details or approve something irreversible needs an out-of-band check. A face or voice on a call is not identity.

Cigna PXDX (budget and rubber stamps)

A class action filed on 24 July 2023 alleges that over two months in 2022, Cigna doctors denied over 300,000 payment requests using a method that relies on an algorithm known as PXDX, "spending an average of just 1.2 seconds" on each.5 These are allegations in a complaint, not findings, and the complaint takes the figures from news reporting.5 The case's later course and outcome are not checked here. This course does not use the real totals for arithmetic; budget practice uses the course's synthetic log.

For the tracks: any build that makes many small decisions needs a budget per reviewer and system-wide, and a review that is real, not a signature. A very short recorded decision time is a reason to look closer, not proof on its own: the time to sign is not the whole review. Check what the reviewer actually saw, and whether they could disagree and be heard. And an average over many decisions says nothing certain about any one of them.

nH Predict (model informs, person decides)

A class action filed on 14 November 2023 alleges that UnitedHealth used an AI model called nH Predict to override treating physicians' determinations about post-acute care, and that the model has a 90% error rate.6 These are the plaintiffs' allegations, not findings.6 Module 300.8 keeps three things apart: what the complaint alleges, what the company said in response, and the course's gate.

For the tracks: wherever a qualified person's judgment is the decision, a build may inform it but not replace it. The Foundation's gate holds whoever turns out to be right about the case.

The eleven tracks

Each row is one track. The builds and holders are made-up Big Widget Company examples, written as course guidance. The "gate parts most at risk" column names the focus for that build; all eight parts still have to pass. Where a real case or source sits behind a row, the row cites it.

TrackBuild worth doingKindNamed holder (made up)Gate parts most at riskCase, risk or module
Makers and citizen developersAnswers from the plant's published safety procedures, with the procedure cited and "I can't find that" when nothing matchesRetrievalDana Okafor, plant safety leadAuthority (read-only), coverage, holder (who keeps the procedures current)300.2 (stale documents); OWASP vector and embedding weaknesses3
IT adminsAn agent that drafts access-change tickets and runs read-only checks on the requestAgentRaj Mehta, identity service ownerAuthority (no admin rights; granting access needs a named approver), stop, receipt (the agent's own identity, never a shared login)OWASP excessive agency: minimum tools and permissions, human approval for high-impact actions7; 300.4
Early developersAn internal engineering-docs assistant whose permission matrix is checked at query timeRetrievalMei Lin, product owner for the assistantAuthority (permissions per document), receipt, coverageOWASP sensitive information disclosure3; 300.2
CybersecurityAn agent that reads reported phishing emails and labels them for an analystAgentTomás Silva, security operations leadAuthority (label only; no replying, forwarding or deleting), stopOWASP prompt injection: instructions can arrive in outside content such as websites or files8; 300.6
QualityAnswers from warranty-claim notes that surface candidate defect patterns for an engineerRetrievalGrace Adeyemi, quality managerJudgment (the engineer decides), coverage (what it left out)nH Predict allegations6; course gate: a model informs, a person decides; 300.5
Engineering and designAnswers from released specifications, quoting the revisionRetrievalHana Ito, engineering document control leadAuthority (released revisions only), judgment, holder300.2 (two versions of a torque spec)
Supply chain and procurementAn agent that drafts purchase orders under a limit for a buyer to approveAgentPriya Shah, procurement category managerAuthority (drafts alone; orders need the buyer; payments never), out-of-band check (supplier bank changes), budgetHong Kong fake meeting4; 300.4
Customer serviceAn agent that triages warranty claims and flags some for a reviewer to clearAgentLuca Romano, warranty service managerBudget (per reviewer and system-wide), judgment (real review, not a stamp), receiptCigna PXDX allegations5; 300.8, 300.9
FinanceAn agent that matches supplier invoices to purchase orders and flags exceptionsAgentAisha Bello, accounts payable leadAuthority (no payments), out-of-band check (bank-detail changes, urgent payments), budgetHong Kong fake meeting4; 300.4
HRAnswers to policy questions from the published HR handbook, with no access to personnel filesRetrievalSofia Novak, HR policy ownerAuthority (handbook only), holder, out of scope: screening applicantsEU AI Act: employment tools such as CV-sorting are high-risk uses9; 300.7
LegalClause finding across the contract library, limited to the matters each lawyer may seeRetrievalDaniel Cho, legal operations leadAuthority (matter-level permissions), judgment (a lawyer reads every clause cited)EU AI Act roles and transparency duties9; 300.7

Three course rules run through every row. The AI never approves its own use: a named person does. Anything it cannot undo, such as a payment, a message sent outside the company or a change to live data, waits for a person. And if the holder, the approval or the record is missing, the build stops.

Notes on three rows.

Your track exercise

Bring this to the project (300.P), where it becomes a hold you dry-run on the course's synthetic files. A person reviews it there; an AI does not.

  1. Pick one row from your track, or write a new row for your role.
  2. Write the build in one line: retrieval or agent, over which collection or with which tools, for whom.
  3. Name the holder: one person, by role and name, and the date their approval expires. If you cannot name one, the build does not run. Write who you will ask.
  4. Go through all eight gate parts. For each, write pass, fail or "don't know yet". Mark the two or three most at risk for this build, and say why. For each "don't know yet", name who can find out.
  5. Name the OWASP risk most likely for your build and one limit that holds even if the model is fooled.3
  6. One other track. Read one other team's row. Name one way their build could touch your data or your decisions, and who you would tell. If you find no route, say so and say why. Do not invent one.

Use only made-up examples. Do not paste real prompts, documents, logs or personal data into the exercise. In real use, any notes your workplace allows on prompts and outputs stay in its own controlled records, never in a receipt or anything you share.

Prep rails: what they do and do not do

From Level 300 up, the course offers prep rails: drills and mock questions that map course objectives to the topics of a vendor's proctored certification exam. Level 300 has two rails. As of 18 November 2026, neither vendor nor exam is named, because Bob Rapp clears each vendor name before it appears.

RailVendorCertificationDraws on
1not clearednot cleared300.1, 300.2, 300.3, 300.4, 300.5, 300.6 (building and securing)
2not clearednot cleared300.6, 300.7, 300.8, 300.9 (governing and operating)

What a prep rail does. It shows which course objectives cover which exam topics, gives you practice in the course's own words, and shows you 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 is the Foundation's own record that you completed the level and its project. It describes your practice, never a system's safety. It is not a vendor certification and does not count as one.

The Foundation already applies the same line to outside learning. Its page of university AI courses says the programs 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 is the same kind of thing: real learning toward someone else's credential, which only they can give.

What this module does not claim

Words to know

Further learning

Further learning for this module is on hold. Outside course pages can carry commercial platform, vendor and certification branding, and Level 300's rule is that vendor names wait for Bob Rapp's clearance. Until he clears a destination, the sources below are the 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