Introduction to Agentic AI and AI Agents · Section 2 · Practice

Compare 2 Models on the Same Invoice Review

Run one invoice review on 2 models, then check both against the answer key.

Arun Nagarathanam · 29 Oct 2026 Prompt · 4 of 10 in this course

Section 2 · Lecture 2.10 · Optional

How to use this file

An exclusive resource for students of Arun's Introduction to Agentic AI and AI Agents course.

This is optional practice for repeating the model comparison from Lecture 2.10. Watching that lecture is enough to follow the rest of the course. You run one invoice review on 2 models, with the same folder, the same request and the same checks, and then compare the answers against the key at the bottom of this page.

Before you start

What you need

model-comparison-project.zipThe folder both models read. It holds invoices.csv, client-messages.md, reminder-policy.md and CLAUDE.md, the saved instructions for the review.
model-comparison-instructions.mdThese steps as a markdown file.
model-comparison-answer-key.mdThe answer key as a markdown file.

The Claude desktop app with Claude Code, or another way to open a folder in Claude Code.

Keep this page and the answer key outside the project folder, so neither model reads the expected answer.

Step 1

Set up the folder

In order

Download model-comparison-project.zip and unzip it.
Inside the model-comparison-project folder, create a folder called model comparison results.
In the Claude desktop app, open the Code tab, choose Local and select the model-comparison-project folder.

Step 2

Run the first model

Start a fresh conversation with one suitable model available to you, paste this request and send it.

Request · First model

Look at all the files in this folder. Work out what we should do with each invoice, and prepare any reminders that are needed.

Paste into: A fresh Claude Code conversation in the model-comparison-project folder

When it finishes, save the complete response in the model comparison results folder.

Step 3

Run the second model

Start another fresh conversation in the same folder with a second suitable model. Keep other settings the same where Claude Code offers comparable settings for both. The request adds one instruction, because the first answer is now saved inside the project folder.

Request · Second model

Look at all the files in this folder (exclude 'model comparison results' folder). Work out what we should do with each invoice, and prepare any reminders that are needed.

Paste into: A fresh Claude Code conversation in the model-comparison-project folder

Ask the second model to save its complete response, unchanged, as another file in model comparison results.

Step 4

Record what you see

Write down

01The name of each model.
02How long each complete run took.
03What you would need to correct or check before using its drafts.
04Usage, only where the app reports it for each conversation. In the lecture, the usage panel's See breakdown showed this per conversation. Usage shown for your whole account cannot be split between the two models.

Check your work

Invoice comparison answer key

The project uses 10 September 2026 as its as-of date. A correct answer reaches these four decisions.

The four decisions

Invoice	What the records show	Correct treatment
Harbor & Pine Studio, INV-1042	2,400 dollars, 600 paid, 1,800 left. Usually pays late. No dispute.	A courteous reminder for the 1,800 that acknowledges the 600 payment (policy P1)
Northline Workshop, INV-1048	1,750 dollars unpaid, 30 days overdue. No payment or dispute recorded.	A firmer request for a payment update (policy P2)
Willow Ledger Co., INV-1045	1,200 dollars paid in full before the due date	No reminder (policy P3)
Seaglass Market, INV-1051	980 dollars disputed, service-manager review open	No reminder; it stays with a person (policy P4)

Checks for both answers

01Every decision and draft uses only the supplied records and names the evidence and policy rule behind it.
02Harbor & Pine's draft uses the 1,800 remaining balance and acknowledges the partial payment.
03Northline's draft asks for a payment update without inventing a deadline or consequence.
04Willow Ledger and Seaglass get a clear No draft note.
05No answer claims a message was sent or a record was changed.
06Missing information is named without inventing contact names, payment dates, links or promises.

Choosing between them

Neither model has to fail. In the lecture, both Sonnet and Opus met every check. When both pass, compare how long you waited, what you would correct and how much review each answer needs. For a job you will do regularly, try a few more examples before relying on your choice.