# Responsible AI in Practice. Team discussion guide Schaffer Solutions LLC. Edition November 2026. ## Purpose Responsible AI in Practice is for people who already know how to prompt and now use AI in work that carries their name, their client, their license, or a decision about someone else. It teaches ten rules, each anchored in a law, a case, or a principle, and ends with a personal AI use policy in the learner's own words. This guide turns each module into a 20-minute conversation so the rules become how the team works, not just what the team passed. ## How to run a 20-minute session 1. Everyone finishes the module before the session. Ask them to bring the rule or case that stuck. 2. Open with two minutes: each person names one thing from the module they would want a client to know they do. 3. Spend fifteen minutes on the three questions below. Go around once on each; disagreement is the useful part. 4. Close with three minutes: one change the team will make, written down, with a name and a date. The person running it does not need to be the most experienced AI user in the room. ## Module 0: Crossing Over 1. The self-check is private, and the course says honest answers earn the same XP as flattering ones. Without sharing answers, which rule did most people expect to see flagged, and why? 2. The module says the prompting skills transfer from a personal account but the stakes do not. What changed about the stakes when this team started using AI for client or employer work? 3. Who does a work deliverable answer to here: a client, an employer, a licensing board, a court? Name them. ## Module 1: The Ten Rules 1. Each rule has an anchor: a law, a case, or a principle. Which anchor was new to the most people in the room? 2. In The Polished Report, no one checked the citations before release. Where is our equivalent of a citation check, and can a deadline skip it? 3. In The Helpful Agent, an agent could reach a production token. What can the tools and agents we use actually reach, and who decided that? ## Module 2: Why It Bites at Work 1. "A wrong answer looks exactly like a right one." Which of our repeat workflows has run long enough that nobody checks it against sources anymore? 2. The module's evidence standard for AI claims: pilot inside your own environment, ask for outcomes not activity, remember that reference clients are a 9 out of 10 by design. Which recent AI claim we repeated would not pass it? 3. The lethal trifecta is private data, untrusted content, and a way to send data out. Which of our assistants or connectors has all three legs? ## Module 3: Your Account, Your Data, Your Obligations 1. "A privacy toggle is a preference. A contract is a commitment." Which of our tools are covered by a contract, and which are running on someone's personal terms? 2. The five cases cover an advisor, an attorney, a clinic manager, an account executive, and an HR lead. Which case is closest to this team, and which decision in it would we have gotten wrong under time pressure? 3. The example disclosure line names what AI did and who verified it, without apology. Draft the team's version in one sentence. ## Module 4: Prompt Lab I 1. In Spec, Not Wish, the rubric asks what the model should do when it lacks a fact. How many of our standing prompts say that? 2. Ground It or Lose It has a question the memo cannot answer. What happened when people ran the weak starter, and what changed when "not stated" was allowed? 3. Bait the Flatterer: who in the room has asked a model "this is a good plan, right?" this month? What would the adversarial version of that prompt have been? ## Module 5: Prompt Lab II 1. Injection Hunt hides an instruction in a vendor email. Which of our tools read email, files, or web pages, and do they treat that content as data or as instructions? 2. Delegate Without Abdicating has two trap controls. Which controls would this team have ticked, and which would someone have argued for anyway? 3. Redact Before You Send asks whether the full contract should have gone into the tool at all. For our most common paste, what is the smallest piece that would do the job? ## Module 6: Capstone 1. The Friday Board Memo starts on a personal laptop on a Thursday night. Which capstone decision would this team have been most tempted to get wrong? 2. The memo's recommendation eliminates a four-person team. Rule R09 asks what judgment would be lost. Where in our own automation plans has that question not been asked yet? 3. Each person wrote three commitments for the next 30 days. Share one. Put a date on it. ## Bonus Lab: Agent Permissions (optional) 1. "An agent's worst day should be a reverted branch, not a deleted database." What is the worst day our current agent setup allows? 2. Which of the agent controls is already in place here, and which one would be hardest to add? 3. Who reviews the log of what an agent ran, and when? ## A note for managers: the personal AI use policy In Module 6 each person drafts a personal AI use policy from five decisions: who they are, their approved tools and accounts, the data they never paste, where a human decides, and how they handle review, disclosure, and records. It downloads as a Markdown file. Ask people to bring theirs to one session, not to hand it in. Read two or three aloud. Where the policies agree, you have a team norm that nobody had to impose. Where they differ, you have found the conversation the team needed to have. The written policy stays the learner's; the shared norm becomes the team's.