Your campaign brief says 'subject to availability'. The AI headline says 'guaranteed'. That is not a tone change. It is a different offer.
Your prompt is not the advert the customer sees
An approved campaign brief can become an unapproved promise in seconds. Ask an AI writing tool for a shorter headline and it may remove the condition that made the original claim defensible. Ask for a stronger opening and a measured benefit can become a guarantee. Neither change needs an obviously invented statistic to mislead a customer. The dangerous version often sounds clearer, more confident and more commercially attractive than the approved original.
Consider an illustrative maintenance business offering a next-working-day visit within a defined service area, subject to availability. A generated advert that says 'We will fix it tomorrow' has changed three things: a visit has become a repair, a conditional service has become a promise, and the geographic restriction has disappeared. These are invented examples to explain the risk, not findings about a real company. The point is that grammatical quality and commercial accuracy are different acceptance tests.
The CAP Code section 3 makes that distinction useful rather than theoretical. Rule 3.7 requires documentary evidence before distribution for objective claims capable of substantiation. Rule 3.9 says significant limitations and qualifications must not be omitted, and qualifications must not contradict the claims they qualify. A sensible prompt helps, but the published wording still needs to satisfy those requirements.
What this means in practice is simple: approve the customer-facing promise, not merely the prompt that requested it. Compare each final headline, image caption and landing-page statement with the authorised claim. Ask what a reasonable customer would think they were being offered. If the answer has changed, route it back for review. Keeping the same brand voice does not make a materially different promise an approved one. This applies whether the copy comes from ChatGPT, an agency, a platform's built-in generator or a person working from a template.
An approved prompt is not approval for every promise it produces.
Build a small claim register before generating variations
Most small businesses do not need another compliance platform to begin. They need one controlled list of what they can actually say. A shared spreadsheet or a structured document can record the approved wording, the evidence behind it, the product or service it covers, the conditions that must travel with it and the person who can authorise a change. Start with claims already used in your highest-spend campaign rather than trying to catalogue the entire business.
Keep evidence and wording together. 'Our team answers calls during office hours' is not the same as 'instant support'. A result from one customer's pilot is not evidence that every buyer will achieve the same saving. Record whether a number describes a particular test, a range, an average or a guaranteed outcome. Note when the evidence was collected and what would make it stale, such as a price change, a different service package or a new measurement method.
For each claim, identify the parts that must not drift. These might include the VAT treatment, eligibility, minimum commitment, geographic coverage, renewal terms or the difference between response and resolution. Under CAP rule 3.18, quoted prices generally need to include non-optional taxes and charges applying to all or most buyers; VAT-exclusive advertising has specific conditions. Do not let a generator decide which price presentation applies just because the campaign is described as business-facing.
Use the register as input to generation, but do not treat retrieval as proof of compliance. The tool might select an outdated row, combine incompatible offers or lose a qualification when shortening text. Assign every generated variant to a claim record and retain its exact wording. If the business cannot locate supporting evidence, remove or narrow the claim before spending money to distribute it. The useful output is a manageable set of defensible promises, not a large archive nobody can navigate.
If a claim has no evidence owner, it has no reliable approval route.
Review the actual combinations, not just individual assets
An advert is more than a collection of separately acceptable sentences. A headline, image, price and call to action can create a misleading combined impression even when each component looks harmless in isolation. A synthetic demonstration might imply a result the real product cannot deliver. A testimonial-style voice can suggest genuine customer experience. A short price statement can appear to describe a one-off purchase while the landing page reveals a recurring commitment.
The ASA's 16 September 2026 Animcha ruling investigated two issues and upheld both. It concerned harmful and offensive presentation in an advert for an AI companion app, not an ordinary pricing or substantiation dispute. Its relevance here is narrower and important: the regulator assessed the finished communication and its overall presentation. The fact that the subject was an AI product did not remove established advertising responsibilities.
The 21 October 2026 Bingchuan ruling offers a separate, named example of outsourced creative oversight. The advertiser described third-party production and promotion arrangements, said it had withdrawn the advert and terminated the arrangement, yet the ASA still upheld the complaint. The ruling does not establish that AI made that advert. It does show why outsourcing production is not a substitute for control over what represents the business.
In practice, request previews of the combinations that can actually appear, including short mobile layouts. Check whether important wording is truncated, hidden after a click or separated from the price it qualifies. If a platform automatically assembles assets or generates new wording after upload, establish what controls and review visibility it provides before enabling those features. Do not claim to have approved combinations you have never seen. Where exhaustive review is not feasible, constrain the variation or remove the higher-risk claims.
Review what the audience can see, not just what the asset folder contains.
Keep fast testing, but restrict what AI may change
The strongest counterargument is reasonable: checking every variation could erase the speed advantage that made AI attractive. A small team cannot send every adjective through a formal approval committee. The answer is not to pretend all wording changes carry the same risk. Separate creative freedom from permission to change facts. Let the tool explore tone, sentence order and non-factual framing inside clear boundaries, while treating prices, guarantees, comparisons and customer results as controlled material.
That boundary needs a practical consequence. A variant that introduces a new number, a superlative such as 'best', a testimonial, a guarantee or an unsupported comparison should leave the fast lane. So should a version that removes a qualifying condition. It may be perfectly usable after assessment, but it should not inherit approval from an unrelated original. Automated comparison can flag these changes; a person still needs to decide whether the resulting impression is accurate.
Recent industry commentary supports this operational focus. Intercepta AI writing on 29 October 2026 discusses how variations can alter approved claims. It is commercial commentary, not an independent measurement of how frequently errors occur. Similarly, Vane AI's 12 October 2026 UK advertising guide argues for starting with product truth rather than production tools. Neither source establishes a universal error rate, so do not turn their examples into one.
A useful small-business rule is to automate formatting more freely than factual commitment. For an ordinary low-risk campaign, nominate one reviewer and use a short checklist rather than creating a new committee. For health, financial or other specially regulated claims, use the appropriate specialist route. If the tool changes the commercial meaning, treat the output as new copy. You keep the benefit of rapid drafting without allowing a language model to expand the offer on your behalf.
Automate creative variation, not permission to invent a better offer.
Make withdrawal as deliberate as approval
A campaign can be correct when approved and wrong a week later. Stock runs out, a promotional period ends, a service boundary changes or a supporting result is superseded. If the business retains only the prompt and an approval email, it may struggle to find every live variation that carries the old claim. The claim register therefore needs a second purpose: helping the owner locate, pause and replace affected communications quickly.
Record where each approved version runs, which offer it uses and who can stop it. Include the advert, the destination page and any generated follow-up material that repeats the promise. Retain the exact approved asset rather than assuming a platform will always preserve a convenient preview. When an agency or platform supplies the creative, agree access to version history and withdrawal arrangements before the campaign starts. These are operational checks, not a requirement to purchase a particular monitoring product.
Test the process with a harmless scenario: an approved offer expires tomorrow. Can the owner identify every associated campaign, stop new generation from using the old wording and confirm replacement wording is live? If a platform keeps producing fresh variants, pausing one visible asset might not be enough. The Bingchuan ruling also records a publisher explaining that blocking one advert did not necessarily prevent a similar one appearing in a new campaign. That is a useful reason to check campaign-level controls, not evidence that every platform behaves identically.
Measure review burden alongside output volume. Count variants rejected for changed claims, time spent correcting them and how quickly a withdrawn promise disappears. These are suggested internal measures, not published benchmarks. A small campaign with fewer trustworthy variations may outperform a larger one once wasted spend and correction work are included. Before the next AI-assisted launch, choose one important promise, locate its evidence, inspect its actual mobile presentation and prove someone can stop it. That is a concrete governance improvement a business owner can verify without commissioning a new system.
A promise is not controlled unless somebody can find it and stop it.
Key Takeaways
- Create a claim register with exact approved wording, supporting evidence, required conditions and a named owner.
- Require fresh approval when a variant changes a price, guarantee, comparison, testimonial or significant qualification.
- Inspect actual mobile advert combinations and landing pages, not only source assets and prompts.
- Limit automatic variation when you cannot see or control the combinations a platform may publish.
- Keep campaign locations and withdrawal access linked to each claim, then test an expired-offer scenario.
