Overview

• Time estimate: ~45 min core (20–25 min discovery + objection handling, 15–20 min memo) + 10–15 min debrief.
• Difficulty: High. No fallback script — the simulated customer will not hand you the answer.
• Format — your choice:
  • Solo: you run the whole conversation and write the memo alone.
  • Paired: one partner drives the chat (asks questions, handles objections) while the other listens and tracks discovery gaps in real time; both co-author the memo. Same driver for the whole exercise — no swapping mid-way.

Setup / Materials

• Any LLM playground (Anthropic Console, OpenAI Playground, Google AI Studio — your choice)
• A brand new chat
• The starter prompt below, pasted as your first message
• The blank memo template below

What We're Going to Cover

Today's framework has seven steps: Business Outcome → Workflow Reality → AI Fit → Architecture Option → Risk & Governance → Delivery Path → Value Case. In this exercise you'll work through the first five live, in conversation, then write them up as a recommendation.

FitPath's VP of Product, Priya Shah, wants an autonomous AI agent to run member support — including modifying subscriptions, issuing refunds, and handling injury-related complaints — with the goal of eliminating the human support team within two quarters.

You're the solutions architect on the call. Priya is enthusiastic, non-technical, has a budget number in mind, and will not volunteer process detail, refund policy, injury-escalation history, or data readiness unless you ask a specific, well-formed question. If you push back on "full autonomy," she'll get a little defensive — but she's persuadable if you reframe credibly instead of just saying no.

Paste this as your first message to the LLM:

Part 1 — Discovery & Objection Handling ⏱ 20–25 minutes

Run the conversation. Your job is to leave with clear answers to:

• What's the actual business outcome Priya wants?
• What does the current support workflow look like — volume, exceptions, ownership?
• Where does judgment/risk enter (refunds, injury complaints)?
• What's the data/process readiness?

Along the way, Priya will raise or imply at least one real objection (cost, autonomy, "why not just use ChatGPT," data readiness). Handle it live, in the chat — don't just note it for later. Use the validate → reframe → architecture pattern from the slides, not a flat "no."

Do not propose an architecture yet. If you catch yourself naming a tool or pattern before you've asked about volume, risk, and ownership — stop and back up.

Part 2 — Recommendation Memo ⏱ 15–20 minutes

Step away from the chat. Using what you learned, fill out the memo template below. Write it like Priya will actually read it — not like lecture notes.

FitPath — AI-Enabled Support: Recommendation

1. Business outcome we're solving for:

2. What the workflow actually looks like (not the happy path):

3. Where AI genuinely fits vs. where it doesn't

AI fits:

Traditional logic / no AI needed:

4. Recommended architecture (name the pattern, justify the choice):

5. Risk & governance — what needs a human, what can be autonomous, and why:

6. Phased delivery path (first 90 days):

7. What I would NOT recommend doing yet, and why:

• Memo names a hybrid architecture, not a single pattern — the scenario has at least three genuinely different risk tiers (general Q&A, refunds, injury complaints) that don't belong in the same autonomy bucket.
• At least one part of the original ask is explicitly pushed back on or phased out, not accepted at face value ("full autonomy," "replace the team in two quarters").
• Section 5 draws a direct line to injury-complaint handling requiring human review.
• The objection you handled in Part 1 shows up in the memo's reasoning, not just the transcript.

Triage Sprint — ask the LLM to generate 5 more FitPath AI feature requests of mixed quality (some strong, some weak-value, some risk-heavy). Triage each with the 🟢🟡🔴 lens in one sentence apiece. No memo — just the call and the one-line justification.