Five creative-software failures worth rehearsing before one of them happens

Rehearse failures before wider use. Five events cover the most consequential boundaries: an unsupported claim, personal data in a prompt, uncertain asset rights, compromised access and an unusable exit.

These drills describe hypothetical breakdowns in an English advertising workflow, not events observed in any organisation. Each scenario becomes useful only when it names a trigger, decision owner and recovery record.

What to take away

Rehearse five failure scenarios quarterly to prepare for unsupported claims, data leaks, rights issues, compromised access and unusable exits.
Each scenario needs a named trigger, decision owner and recovery record to be useful.
Missing asset provenance is an unresolved gate, not proof of infringement or permission to proceed.
A scenario passes only when the buyer can demonstrate the agreed response.

1. A generated qualification changes the claim

An approved offer is resized automatically. The small qualification disappears while the headline price remains. The trigger is any mismatch between source copy and an assembled advert.

Stop release, preserve the output and identify every affected variant. Require advertising review of the actual communication and destination before republishing. CAP's misleading-advertising section addresses substantiation, qualifications, prices and availability within its remit. It does not decide the facts of this imagined event.

The recovery evidence is a version list showing withdrawal, correction and renewed approval. A template lock that nobody tested is not recovery evidence.

2. Personal information enters an external model

A collaborator pastes a customer message or recognisable image into a creative assistant. The trigger may be a log entry, disclosure from the user or supplier incident notice.

Disable the affected route, preserve proportionate evidence and involve the privacy and security owners. Establish what information moved, why, where it went and whether deletion is possible. The ICO's AI and data-protection guidance describes risk assessment and proportionate controls when AI processes personal data. A qualified reviewer must apply current law to the real circumstances.

3. Nobody can prove the asset permission

A synthetic or retrieved image reaches final review, but the team cannot link it to permitted source material or likeness consent. Treat missing provenance as an unresolved gate, not proof of infringement and not permission to proceed.

Quarantine the asset, substitute cleared material and retain the prompt, source register, applicable terms and human edits where lawful. The government's March 2026 copyright and AI report records consultation responses and policy options. Its proposals do not grant a licence for the asset in this scenario.

4. An account or integration is compromised

Unexpected publishing activity, changed roles or unexplained exports should trigger containment. Remove affected credentials, suspend automation and switch to the documented manual route. Check downstream destinations rather than assuming disabling the creative service retracts released adverts.

The voluntary UK AI Cyber Security Code of Practice sets baseline principles for AI developers, operators and data custodians. It is neither certification nor evidence that a buyer's configuration is secure. The incident owner should preserve logs, record decisions and test a clean restoration.

5. The supplier changes or closes the service

A material term, model, export function or service boundary changes. The earliest trigger could be a contract notice, failed export or missing dependency. Freeze avoidable changes and compare the notice with the inventory agreed at purchase.

The NCSC's guidance for software customers encourages buyers to request evidence for supplier security claims and maintain assurance. Apply that discipline to exit: export approved assets, source files, metadata, permission records, users and logs, then open them outside the service.

Run the exercise

Choose one scenario quarterly and give the team a harmless specimen, starting the clock at the trigger. Record who noticed, who could stop release, what evidence survived and whether ordinary work resumed.

A scenario passes only when the buyer demonstrates the agreed response. Change the exercise after a new data flow, advertising category, contract, law or supplier architecture alters risk.

Before you act

Name a trigger and decision owner for each scenario.
Preserve outputs and evidence when a failure occurs.
Involve privacy and security owners for personal data incidents.
Quarantine assets with missing provenance and substitute cleared material.
Remove affected credentials and switch to manual routes.
Export assets, metadata and logs before a supplier closes.

Common questions

What should you do if a generated qualification changes an approved claim?

Stop release, preserve the output and identify every affected variant. Require advertising review of the actual communication and destination before republishing. The recovery evidence is a version list showing withdrawal, correction and renewed approval.

How should you respond when personal information enters an external model?

What steps should you take if an account or integration is compromised?

Remove affected credentials, suspend automation and switch to the documented manual route. Check downstream destinations rather than assuming disabling the creative service retracts released adverts. The incident owner should preserve logs, record decisions and test a clean restoration.