Managing AI mistakes and escalation.

Sooner or later, someone in your business will send AI output that is wrong. For the owner, operations manager or privacy officer who gets that phone call, this lesson builds a one-page escalation ladder your team has rehearsed, so the first real mistake is handled calmly and correctly. Allow 30 minutes, ideally with two or three colleagues for the tabletop exercise.

Grade the mistake first

The right response depends on how far the mistake travelled and what it affected. Grading first stops two failures: panicking over a caught typo, and treating a privacy breach as a wording problem.

Every New Zealand organisation must appoint at least one privacy officer under section 201 of the Privacy Act 2020. If you have not named yours, do it before you finish this lesson.

The six-step escalation ladder

The same six steps apply at every level; only the speed and the people change.

Stop

Pause the workflow or tool for the affected case. Do not keep sending.

Contain

Correct the error with the customer, recall or delete where possible, revoke access if a tool or account is involved. Keep a copy of what happened.

Tell

The people your table names for that level, plus the AI vendor if the tool itself misbehaved. The NCSC’s joint guidance advises contacting the vendor first for tool issues and reporting cyber incidents to NCSC (NCSC).

Assess

If personal information is involved, decide whether it is a notifiable privacy breach: one reasonably believed to have caused, or be likely to cause, serious harm (s 112). If so, notify the Privacy Commissioner as soon as practicable (s 114); OPC asks for this ideally within 72 hours (OPC).

Fix

Find the cause: prompt, missing context, skipped review, wrong tool. Change it and re-test on the failed case.

Record

Log the date, level, what happened, who was told, the fix and the decision on notification, including a decision not to notify.

Wrong claims need the same care. If an AI-drafted advert or webpage made a claim you cannot back up, withdraw it: the Commerce Commission says businesses need reasonable grounds for claims at the time they make them (Commerce Commission).

Exercise: build and rehearse your ladder

1. Fill in the table. Copy the four levels and write real names, with a backup, in the “Who to tell” column. Adjust the time limits to suit your business.
2. Write four scenarios. Base them on your own workflows: one near miss, one minor, one involving personal information, one involving money or safety.
3. Grade each scenario. Do this alone first, then compare with a colleague. Disagreements show where the levels need clearer wording.
4. Run a 15-minute tabletop. Use the facilitator prompt below. Talk through each step of the ladder out loud for one scenario.
5. Note the gaps. Every “we would not know who to call” or “where is that recorded?” is a fix for your ladder.
6. Publish and log. Put the one-page ladder where your AI policy lives and start an incident log, even if it is empty.

Worked example: a tabletop with AI as facilitator

Illustrative example: the owner of a Whanganui café and catering business runs a tabletop with the operations manager. AI drafts function confirmations from enquiry emails. The scenario: a confirmation for a 40-person work lunch left out a guest’s nut allergy mentioned in the enquiry, and the lunch is in two days.

Act as a facilitator for a 15-minute incident tabletop exercise for a small
New Zealand café and catering business.

Scenario: an AI-drafted function booking confirmation for a 40-person lunch on
Friday left out a guest's nut allergy that was in the original enquiry email.
The confirmation went to the client yesterday. The kitchen works from confirmations.

Our ladder: Stop, Contain, Tell, Assess, Fix, Record. Our levels: near miss, minor,
serious, critical.

Run it as a conversation. Ask me one question at a time, starting with how I would
grade this. Wait for my answer before the next question. After the Record step,
list three gaps in our process that my answers revealed. Do not give legal advice.

Example output (illustrative and abridged), after several exchanges:

You graded this as Serious, which fits: it is a safety issue that has not yet
caused harm. Your Contain step (calling the client today and amending the kitchen
sheet) is time-critical, so make sure it happens before anything else.

Three gaps your answers revealed:
1. Nobody checks AI-drafted confirmations against the original enquiry for
   dietary requirements. Add "dietary and allergy notes" to the review checklist.
2. The kitchen works from the confirmation, not the enquiry. Consider attaching both.
3. You were unsure where to record this. Start the incident log today.

The owner adds an allergy check to the review step for confirmations, the same day.

Where escalation goes wrong

• Blame. If people are punished for reporting AI errors, near misses stop being reported. Thank the reporter.
• Letting AI decide on notification. An assistant can explain the test; the decision on whether a breach is notifiable is yours and your privacy officer’s.
• Deleting the evidence. Containment does not mean erasing the record. Keep a copy for the log.
• Fixing the prompt, not the process. Many errors are review failures. See planning human review for AI output.
• Missing manipulation. Some errors are caused by instructions hidden in pasted content. Learn the signs in recognising prompt injection.

The AI incident log template and AI risk register template give you the records. Engineering teams running AI in production can go further with our guide to AI incident response, and our AI governance advisory helps larger organisations set this up formally.

Check your ladder

• I have a four-level severity table with real names and backups.
• I know who our privacy officer is.
• Everyone involved knows the six steps: stop, contain, tell, assess, fix, record.
• I know the Privacy Act test for a notifiable breach and where to find OPC’s self-assessment tool.
• I ran a tabletop on at least one scenario and wrote down the gaps.
• An incident log exists, and near misses go in it.

Next lesson: Using AI for meeting preparation moves from risk back to everyday work.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.