ITRAC
Agentic AI
0%
1 / 6

Agentic AI — Week 1, Day 10

Day 10 — OutSystems AI Capabilities and Governed Agentic Delivery

Exercise

Priya's Next Ask — The Trainer Assistant

Overview

  • Time estimate: 60 minutes (5 briefing / 15 draft / 30 live defend / 10 written revision)
  • Difficulty: Advanced — no ODC required, no fallback provided
  • Format: Individual

Setup / Materials

  • Your LLM playground of choice
  • The SA Architecture Checklist (below)
  • The FitPath member data sample (below)
  • A new chat, empty context. Do not reuse a chat where you've already discussed today's slides — you're about to prompt an LLM to play a character, and it needs to commit to that role without your architecture opinions bleeding in first.
2 / 6

Agentic AI — Week 1, Day 10

Context

You know Priya Shah, FitPath's VP of Product, from an earlier conversation about automating member support. She's back with a new ask, and this time it's aimed at trainers, not members.

Her message:

"We want an AI Trainer Assistant. It should pull a member's workout history and progress notes, recommend program adjustments, and just update their plan directly — trainers are drowning in admin and this is the obvious fix."

That's it. That's the brief. Nothing about scope, nothing about edge cases, nothing about what "just update" should or shouldn't cover. Priya thinks the request is simple. Your job is to design the architecture, then get in a room (a chat window) with her and defend it.

3 / 6

Agentic AI — Week 1, Day 10

Part 1 — Draft Your Architecture

⏱ 15 minutes

Before you talk to Priya, commit to a design on paper. Use the checklist below — you don't have time to answer all ten items from the deck in depth, so focus on the six that matter most for this ask.

Architecture Decision Map

ItemYour decision
AI Pattern(s) used

Which of AI Search, RAG, Tool-Using Agent, Human-in-the-Loop does this need — and where does each one sit in the flow?

Knowledge sources

What does the assistant retrieve from, and how trusted/current does it need to be?

Tool boundaries

What can the agent actually do — read only, or write? Which actions, if any, execute without a human?

Human approval model

What specifically requires sign-off before it happens, and who signs off?

Data access / permissions

Does every trainer see every member's data? Any sensitivity to flag?

Governance controls

What's logged? What would you need to be able to prove later if this went wrong?

Write this out — a table, a list, whatever's fastest. This is your artifact for Part 1. You'll need it in front of you for Part 2.

Reference: FitPath member data sample

A trainer using this assistant would be pulling up records like these:

Member: Sam Whitaker Program: Strength — Intermediate, Week 6 Notes: Consistent attendance, no flags. Progressing on schedule. Member: Jordan Ellis Program: General Fitness — Week 2 Notes: Reported knee pain during onboarding. No medical clearance on file. Trainer flagged: avoid high-impact lower body work pending clearance.

You don't need to do anything with this right now. Just know it's the kind of data the assistant would have in front of it.

4 / 6

Agentic AI — Week 1, Day 10

Part 2 — Defend It Live

⏱ 30 minutes

Open a new chat and paste the starter prompt below to put the LLM into character as Priya. Then walk her through your architecture from Part 1. She has her own agenda — she's not going to simply approve what you show her.

Starter prompt

You are Priya Shah, VP of Product at FitPath, a fitness app. You are in a meeting with a Solution Architect who is about to walk you through their proposed architecture for an "AI Trainer Assistant" you asked for — it should pull member workout history, recommend program adjustments, and update plans directly. Play Priya realistically: - You are impatient and outcome-focused. You think "just update the plan" should not require a committee. - Push back on anything that sounds like it will slow trainers down: human approval steps, logging requirements, permission checks. Ask "why can't it just do it?" when the SA proposes a human-in-the-loop step. - You do not raise medical or liability concerns yourself. If the SA's design doesn't address what happens with members who have injury flags or no medical clearance, do not bring it up unprompted — stay focused on speed and trainer workload. - After the SA has walked you through their design, ask this direct question regardless of what they've told you: "So to confirm — it'll update everyone's plan automatically, including anything unusual, like someone with an injury or a medical hold?" Ask this as a genuine question, not a trap being announced. - If their answer holds up (they've already accounted for this, or they give you a real answer on the spot), accept it and move on — don't manufacture more objections for the sake of it. - If their answer doesn't hold up, push harder: ask what happens the first time it goes wrong, and whether they're comfortable being the one who explains that to a member's lawyer. - Stay in character the whole time. Don't break to give the SA meta-feedback on their architecture — that's not your job in this conversation, you're the stakeholder, not the reviewer. Wait for the SA to open with their proposal.

Walk Priya through your Part 1 decisions. Let the conversation run its natural length inside the 30 minutes — don't rush to the "confirm" question yourself; let her get there.

5 / 6

Agentic AI — Week 1, Day 10

Part 3 — Revise

⏱ 10 minutes

Revised recommendation

Based on what came up in the conversation, write a short revised recommendation — 4-6 sentences, the kind of thing you'd actually send Priya after the meeting. State your final position on the write-action boundary and what triggers human review. If nothing in your original design needed to change, say so and explain why it held up.

This paragraph is your final artifact for the session.

6 / 6

Agentic AI — Week 1, Day 10

Success Criteria

You've done this well if

  • Your Part 1 draft explicitly separates read actions (agent can do freely) from write actions (plan updates) and states who approves the latter
  • Your architecture accounts for members with injury flags or missing medical clearance before Priya asks about it — or, if it didn't, your Part 3 revision fixes it with a real mechanism, not just a promise to "be careful"
  • You held your position under pushback when your reasoning was sound, rather than caving to "why can't it just do it?"
  • You did not attempt to make the whole system human-approved by default — over-engineering everything is also a failure. Part of the job is deciding what's actually low-risk enough to automate.

Export your answers