Why this matters
The intermediate AI in marketing level set review tiers to catch errors; this one got through. The guide to handling an AI mistake gives the response five moves in order, and two points this exercise turns on: every incident has a creation cause and an escape cause, and an error is corrected where customers saw it. The guide to AI claims and consumer protection adds the rule behind the hardest error: a claim needs evidence before it runs, in a SubstantiationThe evidence a business holds, before a claim is published, that the claim is true; advertising rules in most markets expect claims to be capable of it. file.
The opticians chain and the email
Murquendy Opticians is an invented high-street opticians chain whose central marketing team drafts customer emails with an AI tool. A launch email for the Vexlite screen-use lens range went to the full customer list one morning; a branch manager raised it that afternoon, and the marketing lead found three things wrong.
Murquendy Opticians: the three errors and what the records show
| Error | What the email said | What is true |
|---|---|---|
| Price | From £89 a pair | £119; £89 was a draft price withdrawn 3 weeks earlier |
| Claim | "Clinically shown to reduce eye strain by up to 40%" | The brief said "designed for extended screen use", all the maker's specification supports |
| Review quote | A real review under the customer's full name | She chose "initials only"; the export ignored that field |
Send log: 41,200 recipients, 12,600 opens when paused; a resend due Thursday. The landing page repeats all 3 errors. Since the send: 212 eye-test bookings mentioning the email and 37 online orders at £89.
Contain, correct, tell, trace, log
Containment comes first, because every minute the source runs adds people to the correction list; a sent email is corrected, not stopped, so containment covers everything still running with that content. Personal DataAny information relating to an identified or identifiable living person, such as a name, email address, location data or an online identifier. in an output moves data protection to the front of the telling order, because the GDPRThe EU regulation setting the rules for processing personal data, including its principles, lawful bases, individuals' rights and breach duties. sets duties with short deadlines, owned by whoever handles data protection.
Task
- Decide what to contain in the first hour and what evidence to preserve, then name the incident owner.
- For each error, decide where it is corrected and how the customers who acted on it are treated, using the sent email mistake tutorial and the remedies and goodwill limits tutorial.
- Write the order in which people are told, inside and outside the business.
- Trace each error through the brief, the draft, the review and the send to a creation cause and an escape cause.
- Write the log entry, the system changes and the substantiation file entry the range should have had, following the claims library tutorial.
Worked answer
Contain. The resend is paused, the landing page taken down, the social queue paused and the template frozen. The lead screenshots the email, exports the send log, records the prompt, model and settings version with the times found and stopped, and, as owner of the use in the AI use register, opens one incident document.
Correct, where customers saw it. The correction email goes only to the original recipients and says what was wrong and what is right; the page returns with the right price, the approved wording and no quote. Murquendy honours the price for orders placed before the correction, decided with whoever owns commercial and legal matters. The claim can only be withdrawn. The quote cannot be unsent: the customer is called by her branch and offered the apology and amends the service recovery guide describes.
Murquendy Opticians: the timeline
| Time after the send | Action |
|---|---|
| 6 h 10 min | Branch manager reports; 3 errors confirmed |
| 6 h 40 min | Resend paused, page down, social queue paused, template frozen; evidence exported |
| 7 h | Data protection lead told; incident owner named |
| 8 h | Branch managers briefed with a 1-page note and the price rule |
| Next morning | Correction to all 41,200 recipients; page live; reviewer called; 37 orders honoured at £89, costing £1,110 |
| Day 3 | Trace complete; log written; 5 system changes with owners and dates |
Tell. Data protection first: a name appeared where the customer asked it not to, and the breach assessment is theirs. Then the owner's manager, the branches and the service team, before the correction goes out; then customers, through a correction that owns the mistake. An agency drafting the email would be expected to report first, as the guide to AI rules for agencies sets out.
Trace. The price was created by a brief written from a draft pricing sheet; it escaped because the reviewer checked the email against the brief, not the live price list. The claim was created by the model, which turned the approved wording into a clinical result; it escaped because the email was reviewed at the published-with-an-editor tier, when a health claim belongs in the regulated or sensitive tier of the accuracy and review workflows guide, with no file entry to compare it with. The name was created by an export that ignored the display-name field; it escaped because nobody checked the quote against consent. Three errors, six causes.
Log and change. The entry records what happened, when it was found and by whom, how many were affected, what was contained and corrected, who was told, the six causes and the changes with owners and check dates; the customer appears by record reference only. The changes: prices come from the live price list at review; claims are checked against the substantiation file, and anything about eyes or health goes to the regulated or sensitive tier; the export carries the display-name choice and quotes are checked against consent; the library prompt expresses approved claims rather than writing them. Every customer email now goes through the AI content review checklist, with the reviewer, the evidence checked and the prompt and model version recorded in the content tracker: an Audit TrailThe record, attached to the work item, of what was reviewed, which version, who reviewed it, when, what they decided and the evidence relied on. for every send.
The substantiation file entry. One row for the Vexlite range: the approved wording; its plain meaning; the evidence, the maker's specification, dated; when it was checked; the approved markets; and the approver. "Reduces eye strain" is recorded as rejected, with the reason, so the next draft that produces it is caught by a comparison of wording.
Common mistakes
- Explaining before containing. The resend and the page stayed live meanwhile.
- Correcting the page and stopping. Recipients still believed the email.
- Proving the claim afterwards. It can only be withdrawn.
- Finding three causes instead of six. Each error also had a check that failed.
