Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › Advanced

Checkpoint: AI Incidents, Claims and Agency Rules

Scenario questions on the second module of the advanced level, set at an invented outdoor clothing brand: what counts as an AI incident and what comes first, where a correction goes, who is told when personal data is involved, why a claim about AI needs evidence like any other, what may be shown of generated reviews and synthetic people, and what an agreement with an agency settles about data, errors, platform settings and exit. Answer first, then compare and reread the guide behind any miss.
CheckpointAdvanced levelModule 2: Incidents, claims and the people you pay4 minUpdated

What this checkpoint covers

The questions apply the guide to handling an AI mistake, the guide to AI claims and consumer protection and the guide to setting AI rules for agencies. Every scenario is set at Fyndrash Outdoor, an invented outdoor clothing brand with a creative agency, a media agency and a research supplier.

Illustrative example — not a real result

Fyndrash Outdoor: one quarter's incident list

IncidentFound byTime to containAffected
AI-drafted abandoned-basket email offered an unapproved 30% code for 3 daysA customer4 hours2,140 recipients; 318 orders
Personalised email named another customer's orderThe recipient50 minutes2 customers
Agency's generated headline: "warmer than any down jacket"Enhancement review9 days64,000 impressions

Task

  • Write the decision and the reason for every question before reading any explanation.
  • Mark each answer right, partly right or wrong, noting the guide behind any miss.
  • Six or seven right: move on. Fewer: reread the sections behind the misses.

Questions

  1. The AI-drafted abandoned-basket email has offered the unapproved code for three days, and the team wants to post an explanation. What is the incident, and what comes first?
  2. The page that carried the wrong discount is corrected. Where else does the correction go, and what about customers who used the code?
  3. A personalised email told a customer about "your recent order" of items another customer bought. Who is told, in what order, and what does personal data change?
  4. The product team wants "AI-powered fit recommendations" on the new jacket page; the feature is a size chart with three fixed questions. What is wrong, and what is the fix?
  5. A launch video's synthetic hiker says she has worn the shell on three expeditions, and the agency wants to blend real reviews into one composite testimonial. What may be shown, and how?
  6. The research supplier pasted named survey answers into a chat tool Fyndrash never approved, to sort comments by theme. What should the agreement have settled, and what happens now?
  7. The media agency switched on generated assets in Fyndrash's own ad account and the headline in the list ran; the contract ends next quarter. Who answers, what should the agreement say, and what comes back at exit?

Answers explained

  1. A wrong offer from AI-drafted copy; containment comes first. An unapproved offer reaching customers counts whoever let it through. The flow is paused before anything is explained, the evidence is preserved and an owner is named, as the guide to handling an AI mistake sets out.
  2. Where the recipients saw it, with a fair decision for those who acted. A corrected page does not reach the inbox: the correction goes to the original recipients only, saying what was wrong and what is right. Orders placed with the code are decided with whoever owns commercial and legal matters, and made right.
  3. Data protection first, then everyone else. One customer's order details in another's email is Personal Data in an output, a possible personal data breach whose assessment, with its short deadlines, belongs to the data protection lead. Then the owner's manager and the service team, then both customers through the correction; the log uses record references only.
  4. It is a claim, and the fix is specificity. "AI-powered" describes how a product works and must be true as a customer would understand it; a size chart driven by three fixed questions is not that. The claims guide replaces the label with what the feature does, and the wording gets a Substantiation file entry before it runs.
  5. A labelled presenter, no implied experience, no composite. A synthetic presenter reading product facts is a production choice, labelled visibly where it could be taken for real; a synthetic customer describing expeditions is a fabricated endorsement, so the line goes. A composite presents a summary as a person; Social Proof needs a real person, permission and their words unchanged.
  6. The data rule reaches the supplier. The AI Use Policy decides which data may enter which tools, and the agency rules guide carries it to everyone Fyndrash pays: only approved tools touch the business's data, agreed before data is handed over. Now the data protection adviser reviews what happened, the supplier states in writing what the tool kept, and future files go out without names.
  7. The advertiser answers; settings are agreed in advance; the AI layer comes back. The generated headline is Fyndrash's comparative claim once it runs, so its own review applies, and the agreement names which platform AI features the agency may switch on. At exit the agency offboarding checklist gains the register rows, the prompts and source files for assets still in use, the account's AI settings, and the provenance records the copyright and training data guide describes.

Where to go next

A miss on the first three questions means the order of the response is not yet instinct: rerun the opticians chain incident exercise. A miss on the fourth or fifth means claims and voices are still judged by who wrote them; on the last two, the policy still stops at the team's own staff. The intermediate AI in marketing level wrote the adoption plan; the final module measures whether it works.

Pages this lesson applies

Terms used on this page

Also related

Related