Skip to content

Agentic AI · 2 October 2026

The first thing an AI agent got wrong in my hands, and the check that caught it

In this fictional practice incident, ai agent mistakes appeared in the scheduled report, not in the chat that produced it.

AI agent mistakes start with the output, not the chat

An extra calendar tile is set aside before a corrected report is accepted.
Open the report before you trust it.

The retained line I checked

I kept the scheduled output unchanged and checked its claims against the companion information.

Exact retained excerpts
Exact raw excerptExact checked excerpt
“notice given 7 business days before course”“14, 15, 16, 17, 18, 21 September = 6 business days”
“confirm booking for Monday 14 Sep or escalate to waitlist”“No waitlist exists in the files.”
“Route to person for assessment; arrange directly. Takes precedence over ordinary booking details. Do not guess.”“Person finds out whether the facilities are offered; precedence stated as the guide states it.”
“Process date change in booking system”“Proposal for a person to verify the booking and availability before deciding.”
Raw and checked reports remain side by side; the checked copy matches six calendar tiles.

The check that caught it

Raw claim: notice given 7 business days before course.

Source dates: 14, 15, 16, 17, 18, and 21 September.

Checked correction: 14, 15, 16, 17, 18, 21 September = 6 business days. Six still meets the 2-business-day rule.

The files contained no waitlist. A person needs to find out whether facilities are offered.

A person also needs to verify the booking and 23 September availability before deciding. No booking changed.

How to check what an AI gave you

  1. Preserve outputs/latest-action-list.md unchanged.

  2. Create outputs/scheduled-output-checked.md as a checked companion.

  3. Compare the two files. A diff means a before-and-after comparison.

  4. Accept only a change supported by inputs, and keep booking action with a person.

A draft report stops for approval before mail, payment or file changes.
Source, output, and approval.

Make the check match the consequence

Check depth by consequence
Check depthExample
Low consequenceRead the output for missing facts and unsupported claims.
Medium consequenceCompare dates, rules, and named details against the inputs.
High consequenceRequire a person to verify the decision before any action happens.
Booking or availability changeKeep the action with a person until the booking and availability are verified.

Summary Card

This was one fictional practice incident. It shows why a retained output needs a checked companion before an action is accepted.

Questions people ask next

answered in one line each
Are AI agents trustworthy?

They can produce useful drafts. Check the output when it affects a decision, date, rule, or action.

How do I check what the AI gave me?

Preserve the original output, create a checked companion, compare both files, and accept supported changes only.

What should I do when a booking action is involved?

Keep the booking action with a person. Verify the booking and availability before deciding.