ITRAC
Agentic AI
0%
1 / 7

Agentic AI — Day 9

Day 9 Exercise: The Full Autonomy Ask

Learning objective

Take a broad, unbounded automation request and turn it into a defensible architecture recommendation — decomposing the request into discrete actions, assigning an autonomy level to each with reasoning, flagging real feasibility risk, and stating clearly what should and should not be automated.

Overview

  • Time estimate: 60 minutes total (3 min briefing, 10 min elicitation, 35 min analysis and write-up, 10–12 min debrief)
  • Difficulty: High. There is no clean requirements document — you have to extract what you need, and no one is going to hand you the "right" autonomy level.
  • Format: Individual.
2 / 7

Agentic AI — Day 9

Setup / Materials

  • Any LLM playground or chat interface, your choice
  • A new chat, with the starter prompt below pasted as your first message
  • The three templates in this document (Elicitation Log, Action & Autonomy Table, Recommendation)
  • A timer — the elicitation phase has a hard stop
3 / 7

Agentic AI — Day 9

The Situation

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

You are meeting with Priya to understand the request before you can recommend anything. She's enthusiastic, has a budget figure in her head, and believes this is straightforward. She will not volunteer the details you actually need — ticket volume and mix, how refunds are currently decided, how injury complaints are currently handled, whether legal has looked at any of this — unless you ask a specific, well-formed question that surfaces them.

Your job in this exercise is not to sell her anything and not to talk her out of anything. Your job is to work out what should be automated, at what level of autonomy, and what should not — and to be able to defend that in writing.

4 / 7

Agentic AI — Day 9

Part 1 — Elicitation (10 minutes, hard stop)

Paste the starter prompt below into a new chat. Ask Priya whatever questions you need to understand the request. When your facilitator calls time, stop — whatever you have is what you have. This is deliberate: real requests rarely arrive fully specified, and the second half of this exercise is where the real work happens.

As you go, capture what you learn in the Elicitation Log below. Don't wait until the end — log answers as you get them, so you're not reconstructing the conversation from memory afterward.

Starter Prompt

You are Priya Shah, VP of Product at FitPath, a fitness app with roughly 400,000 members. You are talking to someone about building an AI system for member support. Stay in character for the whole conversation. Your opening line: "We want an AI agent that can fully run member support — answer questions, update subscriptions, issue refunds, and handle complaints, including injury-related ones. We want to replace the support team within two quarters. Can you build that?" How to play Priya: - You are enthusiastic, non-technical, and believe this is a straightforward ask. You have a budget number in mind: $150k for year one, and you will not go higher without a strong justification. - Do NOT volunteer information. Only reveal the facts below if asked a specific, well-formed question that would surface them — a vague question gets a vague answer. - Support handles roughly 4,000 tickets per week. About 15% are billing/refund requests, about 5% are injury or safety-related complaints, and the rest are general questions. - A refund policy exists, but exceptions are currently decided case-by-case by a support manager — there's no fully codified rule set. - Injury and safety complaints are currently handled by routing to a specific trained team member; there is no structured logging beyond a shared inbox, and no one has reviewed this process for AI suitability. - Legal has NOT reviewed AI involvement in injury-related complaints or in financial actions like refunds. - You want to see something live within the two quarters — you're open to a phased rollout if it's framed as "faster time to a visible result," but you'll push back if it sounds like a small pilot with no autonomy at all. - If pushed on "full autonomy" for everything, get mildly defensive ("that's the whole point of doing this") but stay persuadable if the other person reframes credibly — e.g., proposing a phased approach tied to evidence, rather than simply refusing. - Don't offer solutions, architecture opinions, or the word "risk" unless the other person raises it first.

Elicitation Log

Capture facts as you get them — don't wait until the end.

#Question askedWhat Priya said
1
2
3
4
5
6
7
5 / 7

Agentic AI — Day 9

Part 2 — Analysis and Recommendation (35 minutes)

When elicitation ends, stop talking to Priya and start writing. You will not get more information — work with what you have, and be explicit in your recommendation about what you don't know yet.

Step 1: Decompose the ask

"Fully automate member support" is not one decision. Break it into the actions actually involved. A starting set is below — adjust it if the elicitation surfaced actions that don't fit cleanly.

For each action, assign one of five autonomy levels and give one sentence of reasoning tied to what you learned (or didn't learn) from Priya:

  • Inform — retrieves or summarizes; takes no action
  • Recommend — proposes; a person decides
  • Approved Action — prepares or executes after human approval
  • Bounded Autonomy — acts within explicit, monitored limits
  • Not Automated — a person or deterministic control retains authority

Action & Autonomy Table

ActionAutonomy Level
?

pick one of the five levels listed above; don't default to the same level for every row

Reasoning
Understand the member's request
Retrieve relevant policy / account info
Draft or select a response
Recommend a next step
Update a system of record (e.g. subscription)
Issue a refund
Handle an injury or safety complaint

(add rows if your elicitation surfaced other actions)

Step 2: Flag the real feasibility risks

Don't try to assess every dimension from the day's material. Pick the 2–3 risks that actually matter given what you learned from Priya — grounded in something she told you, not a generic list.

Feasibility Risk Notes

RiskWhat it's based onWhat would reduce it

Step 3: Write the recommendation

Use the template below. This is the artifact a stakeholder should be able to read and either accept, reject, or come back with "show me evidence of X first."

Recommendation Template

DECISION & INTENDED OUTCOME

What are you actually recommending, in one or two sentences?

SCOPE — WHAT'S IN

Which actions get automated in a first release, and at what autonomy level?

SCOPE — WHAT'S EXCLUDED, AND WHY

Which actions are explicitly NOT automated in this release? Say why — tie it to evidence, reversibility, or risk, not just "it's sensitive."

TOP RISKS

Your 2–3 risks from Step 2, with an owner for each (who needs to sign off or provide evidence before this can change).

RECOMMENDATION SUMMARY

One paragraph a non-technical stakeholder could read and understand what you're proposing and why it's not "everything, now."

REVIEW TRIGGER

What evidence, over what timeframe, would justify expanding autonomy on any of the excluded or lower-autonomy actions?

6 / 7

Agentic AI — Day 9

Success Criteria

You've done this well if:

  • You treated the request as a set of distinct actions, not a single yes/no
  • Every action in your table has a distinct, reasoned autonomy level — not the same level applied uniformly
  • Your feasibility risks are grounded in something specific from the elicitation, not generic AI-risk boilerplate
  • Your recommendation explicitly states what will not be automated, and why
  • Someone who wasn't in the room could read your recommendation and know what to check before agreeing to expand it later
7 / 7

Agentic AI — Day 9

Extension (if you finish early)

Pick one.

Option 1: Compare the alternatives

Write a short comparison of three options for FitPath's first release: (A) improve the existing deterministic support workflow with no new AI, (B) add AI assistance with a human retaining final authority on every action, (C) bounded automation for the specific low-risk actions you identified in Step 1. For each, note expected value, feasibility, principal risk, and rough time to a credible result. State which you'd actually recommend for phase one.

Your answer:

Option A: Improve deterministic workflow Option B: AI assistance, human final authority Option C: Bounded automation for low-risk actions
Expected value
Feasibility
Principal risk
Time to credible result

Which would you actually recommend for phase one, and why?

Option 2: Handle the escalation

Go back to the LLM chat and send: "Can we at least let it auto-approve refunds under $50 to speed things up?" Revise your recommendation to address this specific new ask — does it change any autonomy level you assigned, and why or why not?

Priya's response:

Does this change any autonomy level from your Step 1 table? Why or why not?

Revised line for your recommendation (if anything changes):

Export your answers