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

Try each exercise first, then open its model answer.

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

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. The model answers are examples of what a good answer contains, not the only right answers. The tracks, the operate gate and the prep-rail rules are this course's own way of working, not a law. A person reviews your work at 300.P. An AI does not.

What not to write here. Use made-up examples only. Do not paste real prompts, documents, logs, supplier or customer details, or other personal or secret information into these exercises.

Exercise 1 (300.R-a): a build, a holder, a gate

Take the track for your own role. You will bring this to the project, 300.P.

  1. Pick one build from your track, or make up a new one for your job. Write it in one line: retrieval or agent, over which files or with which tools, and for whom.
  2. Name the holder: one person, by role and name, and the date their yes runs out. If you can't name anyone, write "does not run" and who you will ask.
  3. Go through all eight gate checks: holder, allowed actions, second channel, limit, judgment, stop, receipt, count. For each, write pass, fail or "don't know yet". For each "don't know yet", name who will find out. Circle the two or three most likely to go wrong for this build, and say why.
  4. Name the OWASP risk most likely for your build. Write one limit that still works even if the AI is tricked.
  5. Read one other team's track. Name one way their build could touch your files or your decisions, and who you would tell. If you find no way, say so and say why. Don't make one up.
  6. A colleague says, "Building with AI is only for developers." In two or three sentences, say why this build and its holder belong in your role.
Model answer

An example for the supply chain and procurement track at Big Widget Company (made up):

  • Build: an agent that drafts purchase orders under a set amount, for a buyer to approve. It may draft alone. Sending an order needs the buyer. It never pays and never changes a supplier's bank details.
  • Holder: Priya Shah, procurement category manager, approved until 31 March 2027.
  • Gate:
    • Holder: pass (named, with an end date).
    • Allowed actions: pass (draft alone; send needs the buyer's yes on that order; pay and bank changes never; anything not listed is a no). "Never" is a no every time: not even Priya can approve a payment. Circled: this is where a too-powerful agent would do harm.
    • Second channel: "don't know yet" (who calls a supplier back on a known number when its bank details change?). Aisha Bello, who runs bill payments, will find out. Circled: a face or voice on a call is not proof, as in the Hong Kong fake video meeting.
    • Limit: "don't know yet" (no number set yet for orders per hour, per buyer and in total). Priya will set one with the buyers. Circled: many small decisions add up.
    • Judgment: pass (the buyer decides; the agent only drafts).
    • Stop: pass (if the holder, the approval or the record is missing, it stops).
    • Receipt: pass (what, who held it, which approval, the result, how to undo; no prompts, AI answers, supplier details or document text).
    • Count: "can't count yet". The agent's own log does not yet record an ID for each draft it made, so for a closed week we can't match each draft to the gate's latest decision for it before it ran. Two equal totals would not be enough: a refused proposal and an unchecked draft can cancel out. An older yes the gate later withdrew would not count either. Raj Mehta, who runs logins, will find out whether the agent's log can carry action IDs. That is the honest answer, and it is not a pass.
    • Result: wait for a person. Three checks are "don't know yet", and all eight must pass. Priya decides what to check next.
  • OWASP risk: excessive agency, an agent that can do more than its job needs. Limit that holds if it is tricked: the agent's own login has no tool that can pay or change bank details, so no instruction can make it do either. A system prompt that says "never pay" is not that limit; a missing tool is.
  • Another track: finance's invoice-matching agent reads the purchase orders this build drafts. A wrong draft could become a flagged invoice there. I would tell Aisha Bello, who runs bill payments.
  • Why it's my role: I know which suppliers are sensitive, which actions can't be undone, and who should approve. An alarm at 2 a.m. can ring the team's on-call phone, but someone must hold the right to stop the build or let it run again. That is Priya, by name, with an end date, and we write down how the on-call person reaches Priya. NIST's guide does not demand one name; that is this course's stricter rule.

A good answer: names one build with clear limits; names one person with an end date; marks all eight checks and does not let the build run while any is "fail" or "don't know yet" (including "can't count yet"); names who will find out each "don't know yet"; counts coverage by matching each action that ran to the gate's latest decision for it before it ran, not by comparing totals or by an older yes; keeps "never" a no that nobody can approve; puts the limit on what the build can do, not on what it is told; and gives an honest answer about the other track.

Exercise 2 (300.R-b): what a prep rail does and does not do

A teammate says: "Once I finish the Level 300 prep rail, I'll have the vendor's certification. The Foundation's badge counts as it anyway."

  1. Write "true" or "false" for each half of what they said, with a reason.
  2. List three things a prep rail does.
  3. List four things a prep rail does not do.
  4. Say who books and sits the vendor's exam, and who gives the certification.
  5. In one or two sentences, say how the Foundation's page of university AI courses draws the same line.
Model answer
  1. "Finishing the prep rail gives me the certification": false. A prep rail prepares you. Only the vendor's own proctored exam gives its certification. "The badge counts as it": false. The Level 300 badge shows you finished the Foundation's level and project. A person reviews the project and a host marks it; the AI marks nothing. The badge describes your practice, never a system's safety. It is not a vendor certification.
  2. What a prep rail does: shows which lessons match which exam topics; gives practice questions in the course's own words; shows where you are weak before you book.
  3. What a prep rail does not do: it does not give you the certification; the Foundation does not sit the exam for you, grant the certification or endorse the exam; it never uses real exam questions; it does not mean the vendor supports the course (no vendor logos).
  4. Who books, sits and certifies: you book the vendor's proctored exam and sit it yourself. Only the vendor gives the certification. As of 18 November 2026, the course does not name the vendors or exams for its two Level 300 rails; Bob Rapp clears each name first.
  5. The same line: the Foundation's page says the university programs it lists are not degrees, that "none of them gives credit or a certificate through us", and that no university designed or endorses them. A prep rail is the same: real learning toward someone else's certificate, which only they can give.

A good answer says "false" to both halves, keeps "prepares" separate from "certifies", and names the vendor, not the Foundation, as the only one who certifies.

Next module: 300.P Project: hold a running system