Where AI belongs in a social commerce operation, starting with catalogue exception triage

Start with the task boundary, not a choice of software. The AI foundation models report is context, not proof that AI will improve a social-commerce operation. Move through these decisions in order: protect information, define a checkable job, then decide whether people retain authority and can stop the work.

If live personal information is unnecessary, leave it out

If synthetic or redacted cases can answer the test question, use them rather than live personal information. If live information is necessary, pause until the operation has recorded its purpose, lawful basis, recipients, retention and deletion, and checked how prompts and outputs are handled.

For the specific task, record the personal information involved, its purpose, lawful basis, recipients, retention, deletion, route for individual rights, data sharing and data minimisation measures. Keep the existing recipient reference available in the record through the relevant legislation, alongside review of applicable UK GDPR requirements.

Check whether prompts or outputs train a supplier's service instead of assuming they do not. Apply data protection by design before information enters the system. A supplier providing the service does not resolve what information is used, where it goes or how long it is retained; review access, updates, incidents and exit arrangements as part of the boundary.

Rule of thumb: Use the least sensitive information that can answer the question, and check how the supplier handles it.

If you cannot name one operational task, define it first

If no one can name one operational job and its decision boundary, do not select a system or begin a trial. General uptake does not establish that an application will improve sales or service, or that a particular task can be performed reliably.

ONS research on UK businesses with at least 10 employees, covering 2023 to 2026, describes AI adoption as relatively shallow. Treat that observation as context rather than evidence of value. Record the purpose, input, source of the correct answer, reviewer, error measure and stop rule so a decision to continue or stop can be assessed against an agreed boundary.

Adoption is not a substitute for a defined task. The relevant test is whether the operation can state what information enters, which source establishes the correct answer, who checks the result, how error is measured and when the work stops. If any part is missing, return to task definition rather than widening the proposed use.

Rule of thumb: Treat adoption as context, not proof that your operation should automate a task.

If catalogue exceptions have a reliable reference, test triage only

If an authoritative record lets a named reviewer check each exception, AI may group cases for investigation. It must not edit a listing, invent an attribute or publish a correction; keep the source record and the human reviewer in control of what happens next.

For an Instagram Shop setup, possible cases include restricted-goods flags, missing GTINs, VAT or price mismatches, variant image policy failures and stock count mismatches. These are prompts for investigation, not permission to amend product information. Keep the test within catalogue work that a reviewer can compare with its source.

A retailer such as Next or ASOS selling through TikTok Shop would need to triage these exceptions. The order operation remains the source of truth: the assistant groups cases for human investigation but does not decide what a product record should say.

Build the test from closed, non-sensitive exceptions. One illustration is 200 closed cases, a written limit of 5% false grouping, and zero tolerance for missed restricted-goods flags; suitable limits depend on volume, category and risk appetite. Record false groups, missed high-risk cases and outputs that add claims absent from the product file.

Send restricted products, safety wording and price changes to an authorised specialist. Withdraw the task if the model changes records or its error mix shifts after an update. The test checks a boundary; it does not give the assistant authority over a catalogue or a consumer-protection decision.

Rule of thumb: Let AI sort exceptions; an authorised person decides what happens to each listing.

If AI produces wording, make human approval a release condition

If a system produces draft wording, keep the release decision with a person. A general assistant from OpenAI, Google or Anthropic, or a feature such as Shopify Magic, may prepare a first draft; the approved product record remains the source for factual claims.

A draft of a creator brief, product caption or customer reply needs review against the product record and commercial instructions. Check advertising disclosure rather than assuming plausible wording is ready to publish. A named reviewer should also compare it with the relevant agreement.

Drafting is not a gain if review time, corrections or complaints rise. For sampled work, preserve the prompt, source material, proposed text, edits, approver and release time. That record lets a reviewer compare the released version with the proposal and shows who authorised it, without relying on recollection of how it was produced.

The CMA's AI agents guidance, published in March 2026, says a business remains responsible when a third party provides the agent. It calls for training, monitoring, correction and human oversight. The guidance applies across the UK; interpreting its application to a proposed service requires qualified consideration.

Rule of thumb: AI may prepare wording; a person checks it and owns its release.

If a customer message could affect a remedy, route it to a person

If a message concerns delivery, cancellation or a return, AI may sort it by subject and send delivery, cancellation or return questions to a suitable queue.

It should not approve or refuse a remedy. A person remains responsible for cases that could affect what a customer receives, including questions outside routine categories.

Tell customers when they are dealing with AI if leaving that out could mislead, and provide a workable route to a person. Test ordinary wording, misspellings, multiple questions, distress, accessibility needs and attempts to override instructions. Measure incorrect routing and delayed escalation rather than counting messages handled as the result.

Stop automated handling if the system gives a false price, denies a valid route to help or exposes another person's information. Keep the test focused on routing, not expanding the assistant's authority. A routing result must not decide whether a customer is entitled to a remedy or change the service route available to them.

Rule of thumb: Route the question; do not let AI decide whether the customer receives a remedy.

If ownership or rollback is unclear, defer approval

If no named owner can act on errors, a reviewer cannot check outputs in practice, or stopping the task is not workable, defer approval. Consider a single task and system version only after reference answers, review arrangements, written error limits and rehearsed rollback are in place.

A decision record should identify the task, owner, source of correct answers, reviewer, measures and stop condition. Include platform or API fees, reviewer time and correction work when assessing cost. If a stop cannot be carried out, do not weaken the boundary to make the trial appear workable.

Reconsider approval after a model, data source, platform route, product category or regulator guidance changes. Check consumer law and related guidance when the operating context changes. The evidence here does not predict which application will succeed in 2027; a defensible outcome is a controlled task with evidence, and a reached stop condition returns decisions to authorised people.

The National Cyber Security Centre's secure AI development guidance addresses secure design, deployment, incident management, logging and maintenance. Its security of machine learning principles provide a related reference. A retailer buying rather than building a service still needs evidence about access, updates, incidents and exit arrangements, including supply-chain security.

Rule of thumb: Approve a named task and version only when its limits, owner and rollback are ready.