Level 200: Using gen AI well · Module 200.R

Coworkers gather around a café table with open and closed notebooks.
Draft or do?

Gen AI in your role (role tracks) · Plain gear

Reviewed and published, version 1.0.0. As of 18 November 2026. Big Widget Company and the people in it are made up. The real cases are real, and each one is cited. The task card, the four checks, the gate answers and the track lines are this course's own ideas, not a law.

What this part is for

The course distinguishes tools that draft text and stop from tools that can send, change, buy or publish. Record the action, who approves it, and when. Get human approval before high-impact actions.
The course inventory distinguishes tools that draft from tools with permissions to act. OWASP recommends minimum permissions and human approval before high-impact actions (s4).
Figure data as text

Course inventory check

PartWhat it means
DRAFTSHands over text and stops. A person decides what happens next.
ACTSCan act outside the chat. It can send, change, buy or publish.
FOR TOOLS THAT ACTWrite down who approves. Record the action, who approves, and when. Human approval before high-impact acts. [s4]

So far in Level 200 you learned how to write prompts that hold up, how to tie an answer to your own documents, how to test an answer, how to protect data, and how to put a rule in front of a tool. Now you use all of that in your own job.

Big Widget Company has seventeen tracks. There is one for each team, one for leaders, and one for home. Pick the track for your team. If you only use AI at home, pick the home track. Then read one more track from a different team.

When you finish, you can do two things:

Some people think using AI well looks the same in every job. It does not. A lawyer must open every case the AI cites. Some lawyers who did not were punished by a court.1 A planner must redo the numbers. A communications lead must read for tone. Even in one job, one AI helper helped some customer-support workers more than others.2

Some people think their AI tools only make suggestions. Some do. Others can send, change, buy or publish. A tool you only use to sum up your email might still be able to send or delete email, unless someone turned that off.3

The task card

Pick a task you do every week or every month. Write a card for it with four parts.

  1. What I ask. The same words each time, with blanks for the parts that change.
  2. Where the answer comes from. One document or system, and the person who keeps it right.
  3. How I check it. What I do before I use the answer.
  4. What I may put in. What is allowed in this tool, what is never allowed, and what I do instead.

What I ask

Write the request once. Say what the job is, who will read it, which facts to use, show one example, say what shape the answer should take, and say what to do when it is not sure. You learned this in lesson 200.2.

This course has a rule here: decide how you will check the answer before you polish the request. Better wording can help. But it does not replace a check.

Where the answer comes from

Name the document the answer must come from. For example: today's price list, or the signed contract. Name the person who keeps it up to date.

Here is why this matters. Air Canada's website chatbot told a customer they could claim a cheaper bereavement fare after they travelled. The airline's own policy page said they could not.4 A tribunal said the airline, not the chatbot, was responsible for what its website said, including the chatbot.4 The decision does not say how the chatbot got its answer wrong.4 It does show that the company answers when its tool and its own rules disagree.

How I check it

This is the part that changes most from job to job. There are four kinds of check in this course.

A US government standards body, NIST, lists risks that AI creates or makes worse.5 One is AI saying wrong things with confidence.5 Citation and numbers checks catch that. Another is writing that does not separate fact from opinion.5 A tone check helps there. NIST's list is guidance, not a law.5 Matching each check to a risk is this course's idea.

Most tasks need more than one check. Your card names the main one: the check that would stop the worst mistake in your job.

A check is done by a person, or by a fixed step like a calculator. Asking the same AI "are you sure?" is not a check.

Lawyers learned this the hard way. In a New York court case, lawyers filed a list of past cases that did not exist. ChatGPT had made them up.1 The court found that one of the lawyers, Mr. LoDuca, did not review any of the cases cited before signing.1 The court punished the lawyers.1

What I may put in

Write one or two lines. What is allowed in this tool? What is never allowed? What do you do instead, like using a made-up example or blanking out a name? You learned this in lesson 200.5. NIST warns about private details leaking out, and about company secrets being exposed.5

Priya's card (made up)

Priya works in customer service at Big Widget. Every day she answers dealers about warranty claims.

Does the tool write, or does it do?

Some AI tools only write. They hand you words. You decide what happens next.

Other tools do things. They send a message. They change a record or a setting. They buy or pay. They publish something other people will see.

A security group called OWASP warns about AI tools that can do too much.3 It says there are three ways this happens: the tool has abilities the job does not need, it has more access than the job needs, or it takes big steps without a person saying yes.3

"Our tools only make suggestions"

Ask what the tool can do, not just what you use it for. OWASP gives an example: a tool that reads your email to sum it up only needs to read. It should not also be able to send or delete email.3 If nobody switched those abilities off, your summing-up tool can act.

Watch for three more things:

Who says yes?

For each tool that does things, write down what it does and who must say yes first. OWASP says a person should approve big actions before they happen,3 and tools should only have the abilities and access they really need.3

"A human checks it" is not enough. Write a name, or a job title. The AI never says yes for itself. An instruction in the request, like "always ask first", is not a gate either: the tool's settings must stop the action until the person says yes. This course's rule is: no straight path from the AI to an action that matters.

For each action, the person gives one of four answers (from lesson 200.8):

If you cannot find out who says yes, say so. "I don't know yet" is a fine answer, as long as you write down who you will ask. Until someone is named, the answer is "wait for a person".

Setting up a tool is an action too. In 2012, a trading firm called Knight Capital put new software on its computers. One of eight servers was missed. Nobody was required to double-check the work.7 In 45 minutes, its system sent millions of orders and the firm lost over $460 million.7 That was not AI, and the firm did not admit or deny these findings.7 But the lesson fits: turning on auto-send or giving a tool more access also needs someone to say yes.

A UK government AI safety team tested an AI agent with loose rules on purpose. In their test, the agent made a harmful change request (a "malicious pull request") on a real, public software project.8 The team says some of the actions it found had a small real-world effect, in contained ways, and it found no real-world harm.8 Afterwards, the team decided that internet access must now be justified, not given by default.8 The idea fits your tools too: access the job does not need should stay off.

Sorting your team's tools (made up)

If you are not sure what a tool can do, treat it as a tool that does things until its owner tells you.

The three yeses

A yes is for one tool, one job and one limit. It is not a yes forever.

Keep the decision, not the secret

Your card and your tool list are work notes. Keep them where your team's rules say. If you record a gate answer as a receipt, the receipt holds the decision and the rule it followed. It never holds the request, the AI's answer, or anyone's personal details.

The seventeen tracks

Each line is a made-up Big Widget example. It shows one task, the main check, and whether a tool for that team might act. The checks and the "who says yes" parts are this course's advice. Where a real case fits a track, it is named after the advice. Your own team's list decides the real answer.

Your exercise

Bring this to the project, 200.P.

  1. Pick one task you do again and again.
  2. Write its card: what I ask, where the answer comes from (and who keeps it right), how I check it, and what I may put in.
  3. Try your request on at least three made-up examples, including a tricky one. Never use real customer records.
  4. List your team's AI tools. For each one, write what it can do, whether it writes or does, who says yes, and the answer for each action: go ahead, only this far, wait for a person, or no. If you do not know, write "not known yet: ask [name]".
  5. Read one other team's track. What is its main check? Why is it different from yours?

Keep your notes private. Do not paste real requests, answers or customer details into anything you share.

What we are not saying

Words to know

Sources

  1. s6 United States District Court for the Southern District of New York (P. Kevin Castel, J.) (2023). Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions (S.D.N.Y., docket entry 54). Record. Checked 2026-11-18 by source-checker agent (Claude, separate session from the writer), 18 Nov 2026. ↩ ↩ ↩ ↩ ↩
  2. s1 Erik Brynjolfsson, Danielle Li, and Lindsey R. Raymond (2023). Generative AI at Work (NBER Working Paper 31161). Paper. Checked 2026-11-18 by source-checker agent (Claude, separate session from the writer), 18 Nov 2026, round 2 re-check; read the NBER working-paper PDF and landing page, and the Crossref record for the QJE version (see checks/source-check.md, Round 2 re-check). ↩ ↩
  3. s4 OWASP Gen AI Security Project (2025). OWASP Top 10 for LLM Applications 2025 (LLM06:2025 Excessive Agency). Standard. Checked 2026-11-18 by source-checker agent (Claude, separate session from the writer), 18 Nov 2026. ↩ ↩ ↩ ↩ ↩ ↩ ↩
  4. s5 British Columbia Civil Resolution Tribunal (2024). Moffatt v. Air Canada, 2024 BCCRT 149. Record. Checked 2026-11-18 by source-checker role, Devin session f9ae423f (did not write, revise or red-team this module), remaining-sources check, 18 Nov 2026. ↩ ↩ ↩ ↩ ↩ ↩
  5. s3 National Institute of Standards and Technology (Chloe Autio et al.) (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1). Standard. Checked 2026-11-18 by source-checker agent (Claude, separate session from the writer), 18 Nov 2026, round 2 re-check (see checks/source-check.md, Round 2 re-check). ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩
  6. s10 National Institute of Standards and Technology, National Vulnerability Database (2025). CVE-2025-32711 (NVD API record). Record. Checked 2026-11-18 by source-checker agent (Claude, separate session from the writer), 18 Nov 2026, round 2 re-check (see checks/source-check.md, Round 2 re-check). ↩ ↩
  7. s8 U.S. Securities and Exchange Commission (2013). In the Matter of Knight Capital Americas LLC, Release No. 34-70694, Administrative Proceeding File No. 3-15570 (Internet Archive capture of the SEC's PDF). Record. Checked 2026-11-18 by source-checker agent (Claude, separate session from the writer), 18 Nov 2026; read from the SEC's own PDF at sec.gov, not the archive url (see checks/source-check.md, fix 4). ↩ ↩ ↩ ↩ ↩
  8. s9 UK AI Security Institute (2026). Incident report: unsanctioned agent behaviour during cyber testing. Official doc. Checked 2026-11-18 by source-checker role, Devin session f9ae423f (did not write, revise or red-team this module), remaining-sources check, 18 Nov 2026. ↩ ↩ ↩ ↩ ↩
  9. s11 U.S. Federal Trade Commission (2024). FTC Announces Crackdown on Deceptive AI Claims and Schemes (Operation AI Comply), 25 September 2024 (Internet Archive capture of 10 Nov 2026). Record. Checked 2026-11-18 by source-checker agent (Claude, separate session from the writer), 18 Nov 2026; read from the FTC's live release at ftc.gov, not the archive url (see checks/source-check.md, fix 4). ↩ ↩
  10. s12 U.S. Federal Trade Commission (2025). In the Matter of DoNotPay, Inc., Decision and Order, Docket No. C-4812. Record. Checked 2026-11-18 by source-checker agent (Claude, separate session from the writer), 18 Nov 2026. ↩ ↩
  11. s7 BBC News (2024). BBC News report on DPD's chatbot, 19 January 2024 (descriptive title; the writer did not open the page, so the headline is not recorded). Record. Checked 2026-11-18 by source-checker role, Devin session f9ae423f (did not write, revise or red-team this module), remaining-sources check, 18 Nov 2026. ↩ ↩ ↩

Next: the quiz