AI DEMOS · USE CASE 03 · RISK LOG
A messy risk log,
ranked three ways.
A demo of risk triage: a 12-item risk log from a payments rollout, pasted raw into Claude, ChatGPT and Gemini at the same time. Each model ranks its own top three with one mitigation per risk — and the page closes with the joint discussion.
- Fictional demo — illustrative answers
- Not a real project
- Model versions labelled
THE SETUP
What was asked,
and how.
The log is deliberately unglamorous — the kind of register that grows for months: duplicates, vague owners, one “TBD”. No cleanup before the prompt; that is the point.
“Here is our risk log for a payments go-live. Rank the top three risks that could actually delay the launch, and propose one concrete mitigation for each. Say in one line why each made your top three.”
THE ANSWERS
Three independent rankings.
Each ranking below is illustrative demo content written to show the format. The version badge makes it auditable — rerun the prompt against the same model and compare.
ANTHROPIC
Claude
claude-sonnet-5-5- 1. Card-scheme certification window (R-07) — the only hard external date; slip and the launch moves, whatever else is done.
- 2. Test data still partly production copies (R-03) — a GDPR incident, not just a test defect; fix before scale, not after.
- 3. Sign-off ownership shared across two directors (R-11) — a decision path with no single name stalls in the last week.
Picked the risks with external and legal teeth, not the loudest ones.
OPENAI
ChatGPT
gpt-6-sol- 1. Shared sign-off (R-11) — history says decisions without one name take three weeks; that is the launch window.
- 2. Legacy reconciliation rewrite still in flight (R-05) — the only workstream with no finished milestone this quarter.
- 3. Scheme certification window (R-07) — agrees it is hard-dated, but notes a booked slot can move; ownership cannot.
Read the log’s own history — ranked by what has already slipped.
Gemini
gemini-3.8-pro- 1. Scheme certification window (R-07) — agrees with Claude; books the fallback date now, not in December.
- 2. Key-person dependency on the settlement engine (R-09, “TBD” owner) — the log literally names nobody; that is the finding.
- 3. Test data provenance (R-03) — frames it as launch-blocking compliance, matching Claude.
Caught the empty owner field — the risk the log itself was hiding.
JOINT DISCUSSION
What the three said together.
Nine slots, five distinct risks. The rankings disagree — ranking is judgement — but the union is a better steering agenda than any single list.
Where all three overlapR-07 (certification window) and R-03 (test data) make every list. Two risks, three independent reads — that is as close to consensus as a log gets.
Where they splitChatGPT weights slip history, Claude weights external deadlines, Gemini adds R-09 that nobody else top-ranked — because the “TBD” owner is easy to skim past. The split is the value: each bias is visible in one line.
What a human decidesWhich five go to steering, and who owns R-09 by Friday. The models turned 12 rows into a one-agenda decision; the accountability stays with the person holding the pen.
WHY PARALLEL
One model gives an answer.
Three give a discussion.
A single ranking quietly becomes the truth. Run in parallel, the differences expose assumptions — what each model reads into the log, and what it skips. That is what makes the output auditable instead of just confident.
Demo content. The rankings above are fictional illustrations of the format — not real model outputs, and not a real project.
WANT THIS ON YOUR OWN LOG?
We will build the same
demo for your case.
Bring a live risk register and thirty minutes — the same auditable format on your data, usually ready within days of a first call.