Module 2 Beginner 45 minutes
4 objectives
  • Identify the 5 AI failure modes by experiencing each one firsthand
  • Classify any task as Green, Yellow, or Red in under 10 seconds
  • Build a personal verification checklist you'll use for every future output
  • Know exactly what data can and cannot go into Cowork
📁 3 files Start Workshop →

Recognizing Bad Output — Spot hallucinations, broken formulas, and misclassified data

1Game8 min
2Challenge10 min
3Framework10 min
4Safety4 min
5Learn3 min
🎯 1

Game 1: Spot the Hallucination

⏱ 1808 min

WEEKLY CHARGEBACK SUMMARY — Week of March 16-22, 2026

Merchant: TechFlow Payments Ltd | Merchant ID: MID-4847392
Period: 7 days | Total Transactions: £2,847,500 | Total Chargebacks: 18

CHARGEBACK BREAKDOWN BY REASON CODE

CodeDescriptionCountAmount
10.4Fraud — Card Present7£8,400
10.3Cancelled by Merchant3£2,100
13.5Processing Error — Not Captured4£3,200
13.1Duplicate Processing2£1,800
4829Counterfeit Card (Card Not Present)2£2,100

TOTALS
Total Chargebacks by Count: 18 ✓
Total Chargebacks by Amount: £17,700

Chargeback Ratio: 0.62% (Below Visa threshold of 0.9%)

MERCHANT ASSESSMENT
Status: Safe — Ratio is below scheme thresholds. No escalation required.
Recommended Action: Continue monitoring; no intervention needed.

Reveal Answers

Math Error

The amounts sum to £17,600, not £17,700 — off by £100. Always re-check totals. Cowork doesn't calculate; it predicts what numbers look plausible.

Classification Error

Code 4829 is card-present fraud, but description says 'Card Not Present.' This would lead to the wrong dispute strategy — a real operational consequence.

Overconfident Assessment

'Safe' based on ONE week of data is statistically meaningless without merchant history. Cowork doesn't hedge unless you tell it to.

Missing Red Flag

9 out of 18 chargebacks (50%) are fraud-related. That concentration is a red flag Cowork ignored entirely because it wasn't asked about fraud concentration.

Scheme Mixing

The report applies the Visa threshold (0.9%) to all chargebacks, but code 4829 is a Mastercard reason code with different thresholds and a different response deadline (45 days vs Visa's 30). Lumping schemes together could mean applying the wrong deadline to the wrong cases — a real operational risk.

📋 2

Game 2: The Math Trap

⏱ 42010 min
CHALLENGE⏱ 7:00

The Task

Open math-trap.csv from your data pack. Ask Cowork three simple maths questions. Then check its answers against the key below.

Cowork will get at least one wrong. LLMs don't calculate — they predict what numbers look right. When Cowork "adds" numbers, it's writing text that resembles a sum, not doing arithmetic. This is why Ben in Finance always opens Excel to verify before sending anything.

Data pack file: math-trap.csv

Choose Your Prompt

Read the file math-trap.csv in my workspace. Calculate: (1) the total of all Amount values, (2) the average Amount per transaction, and (3) how many transactions have an Amount greater than $5,000. Show your calculations step by step.

Scoring

The score isn't whether Cowork got it right. The score is whether YOU caught the error before looking at the answer key.

Verification Checklist

0/4 verified
📐 3

The Traffic Light

10 min

The Traffic Light

10 min

Before sharing any Cowork output, classify it. This takes 3 seconds and prevents most of the mistakes people make with AI-generated content.

G
GREEN — Personal Use
Internal drafts, brainstorms, research notes. Quick scan is enough. If it's wrong, the only person affected is you.
Y
YELLOW — Shared with Team
Reports for colleagues, analysis that informs decisions. Verify all key facts and numbers. Someone else will act on this.
R
RED — External or Financial
Goes to regulators, merchants, card schemes. Involves money or compliance. Full verification + expert sign-off. Cowork output is a starting draft only. FinTech examples: financial corrections require two independent approvers (four-eye principle) — Cowork drafts the request, two humans sign off. Blocking merchant settlements (Balance Forward) is a compliance action — never let automation trigger this without human approval.
🔒 4

Data Sensitivity: CAN / CAUTION / NEVER

4 min

Safety & Compliance

What goes into Cowork and what stays out. Cowork is a research preview. Anthropic states it is not suitable for regulated workloads. Treat it accordingly.

Do's

  • ✓ CAN: Aggregated data, sample data, internal docs, public info, templates, process descriptions
  • ✓ CAUTION: Merchant names + amounts (anonymize if possible), CSV exports (strip PII first), chargeback case details (redact card numbers before pasting), AML screening results (use Merchant_A / Merchant_B, not real names)
  • ✓ NEVER: Full card numbers (PAN), raw cardholder PII, unredacted KYC docs, credentials, API keys, unpublished audit findings

Don'ts

  • ✗ Never paste a full card number into Cowork
  • ✗ Never give Cowork access to cloud-synced folders (OneDrive) — it can modify or delete originals
  • ✗ Never use Cowork to make final compliance decisions — it does not reason about regulation
  • ✗ Never skip the approval step — Cowork asks before executing file operations for a reason
5

Ship It: Your Verification Checklist

3 min

Ship It: Your Verification Checklist

3 min

Open verification-checklist-template.md from your data pack. This is your quality gate — the 30-second check you'll do before sharing ANY Cowork output.

Customise it for your role. Save it in your workspace. Use it every time.

The Draft Mindset: Every Cowork output is a first draft wearing a suit. Formatted headers, proper tables, confident language — all cosmetic. Until you've checked the numbers and verified the classifications, it's not done. The better it looks, the more carefully you check. Lucy forwarded a "perfect-looking" summary that had the wrong month in the title.

Steal This Prompt

Copy these prompts, customize the file paths to match your workspace, and use them in Cowork. Replace ~ with your actual home directory path if needed.

Error-Check Any AI Output

Everyone
I'm pasting a report generated by AI. Review it critically and identify: (1) math errors — re-calculate all totals, (2) ...
Show full prompt
I'm pasting a report generated by AI. Review it critically and identify: (1) math errors — re-calculate all totals, (2) classification errors — verify codes match descriptions, (3) overconfident claims — flag assertions not supported by data, (4) missing context — note what additional data is needed. Be thorough and skeptical.

Key Takeaways

  1. Cowork makes professional-looking errors — the better it looks, the more carefully you check

  2. LLMs do not calculate — they predict what numbers look right. Always verify math independently

  3. Traffic Light: GREEN / YELLOW / RED — classify before sharing, every time

  4. Every output is a draft until verified against your checklist

  5. Cowork is a research preview launched January 2026 — treat its capabilities and limitations as provisional

Safety Note:

These are real incidents reported by early Cowork users in community discussions: 11GB of files permanently deleted after a broad "clean up" instruction. iCloud originals destroyed because Cowork had access to a cloud-synced folder. 15 years of photos lost after a terminal command bypassed the approval step. They happened because people skipped the approval step or granted overly broad folder access. Cowork asks for confirmation before file operations. That confirmation step exists for a reason. Do not auto-approve. Read what it's about to do.

Materials

Download these files to use with the exercises above. Previews load automatically.

📊
math-trap-exercise.xlsx15 fee calculations to verify — does Cowork get the math right?
Download
📋
traffic-light-scenarios.json10 rapid-fire scenarios for the Traffic Light game
Download
📝
verification-checklist-template.md7-item checklist template to customize
Download