Agentic AI Workshop 2026 · Capstone Project 2 of 5

Business
Performance
Analyst Agent

Ask the numbers a question in plain English. Get the answer, the chart and the reason why.

CP‑02 · Business Performance Analyst Agent

02

“How did we do last month?”

“Which products are losing margin?”

CP‑02 · Business Performance Analyst Agent

03

01

Map the question

to the right tables and metrics: revenue, margin, volume, cost

02

Write and run SQL or pandas

and fix its own errors from the hint each tool returns

03

Draw the right chart

bar, line or breakdown, chosen to suit the question

04

Explain what drove it

volume, price, mix, channel and category, before giving a reason

05

Flag anomalies

months or segments well off their own recent trend

06

Show its working

the exact queries, taken from the tool log

CP‑02 · Business Performance Analyst Agent

04

1

Plain-English question

2

Plan the analysis

3

Write query or code

4

Run and self-correct

5

Chart and explanation

6

Weekly briefing

Same engine, real data

We kept the course notebook’s agent loop and pointed it at Microsoft’s AdventureWorks sample: 31,465 orders, May 2022 to June 2025, ten territories.

Keeps going

When Groq’s free daily allowance runs out, it switches to OpenAI automatically. Shorter prompts cut the cost per round by 32%.

CP‑02 · Business Performance Analyst Agent

05

CP‑02 · Business Performance Analyst Agent

06

Read-only database

The file is opened read-only, and only SELECT is allowed. “Delete the orders” is refused.

Always show the query

The queries under each answer come from the tool log, never from the AI’s own words.

Say when the data can’t answer

Returns, competitor prices, satisfaction: not in the data, so it says so. No stand-ins.

Never a forecast as fact

It shows the past trend. Any projection is labelled “ESTIMATE, not a fact”.

CP‑02 · Business Performance Analyst Agent

07

−$150,905

Margin went from $110,755 in Q1 to −$40,151 in Q2.

• Margin per unit: −$256,848
• Higher volume: +$240,225
• Reseller channel: −$174,193
• Mountain Bikes: −$166,669

Every figure has its query underneath.

CP‑02 · Business Performance Analyst Agent

08

−$96,403

Northwest margin, May 2024

We planted a 25% bike discount in a copy of the data, then asked a leadership question that leads to it.

The agent traced the loss to the reseller discount, which jumped from 0.02% to 21.69%.

A rehearsal on a separate copy of the database. The real data is never changed.

CP‑02 · Business Performance Analyst Agent

09

KPIs straight from SQL

So the AI can’t mistype a number.

AI writes the story

Commentary and charts on what moved.

One page

With every query it used listed.

CP‑02 · Business Performance Analyst Agent

10

7 / 7

The brief’s four questions plus three guardrail traps, marked against answers worked out from the database.

It was 6 of 7 until we fixed how it answered “return rate”. Plus 27 automatic checks that need no AI key.

CP‑02 · Business Performance Analyst Agent

11

01

Keep the engine, change the inputs

We added a new section instead of rewriting, so nothing that worked broke.

02

Never trust a model with numbers it can look up

KPI tables and query lists come from code and the tool log.

03

Measure the tokens

On a free plan, the fixed text sent every round is the real cost.

04

Plan for a live demo to fail

Labelled replays of real runs kept the demo going.

05

Check against the brief, not just the tests

Reading it line by line found three gaps the tests missed.

CP‑02 · Business Performance Analyst Agent

12

A course demonstration. AdventureWorks is Microsoft’s public sample of a made-up bicycle company; no real business data is used.