Understand the Customer's Product Enquiry Before It Reaches the Team

Customer enquiries arrive through email, chat, telephone and marketplaces, but messages asking whether a product fits, where a delivery has gone or how to return a faulty item need different evidence and different decision-makers. A small product business can use intent routing to reduce repeated sorting work, provided that classification is never confused with proof of a product claim or permission to change an order. The first question is what the customer needs; the second is who is authorised to resolve it. This article examines those questions for UK trading businesses without claiming that artificial intelligence removes the need for human judgement.

The Information Commissioner's Office provides guidance on AI and data protection. This guidance concerns processing personal data, not an endorsement of an individual customer-service product. Businesses must assess their own processing context, supplier contracts and obligations. Examples below describe practical operating choices rather than certified compliance controls or measured returns.

Organise enquiries by the action customers need

Start with representative messages from real business processes, handled under suitable privacy rules. A customer asking whether an accessory fits model X needs specification evidence; a delivery question needs access to the right order events; a refund request requires policy and legal review. Useful categories include pre-purchase specifications, compatibility, stock availability, order progress, faults, returns, invoices and general questions. The categories should lead to meaningful next steps, not merely coloured labels.

A message can fit several categories. For example, an adapter that fails to work may be incompatible, damaged or incorrectly described. Allow an uncertain classification, a request for clarification or a human handover. Preserve the original customer wording alongside any machine-generated label, and record later corrections separately. That makes it possible to review errors without rewriting the history of an interaction.

Set a permission boundary for each route

Decide which requests can receive a response from an approved public source and which require access to restricted information or an authorised decision. Explaining published opening hours is different from revealing private order details. Quoting a verified product dimension is different from assuring someone that an untested component is safe to install. The system should be permitted to describe information it can support, not infer a commercial entitlement from a confident label.

Refund approval, delivery-address changes, order cancellation and possible product hazards should have appropriate human controls. An automated acknowledgement may confirm receipt, but it must not state that a consequential action has occurred unless the underlying transaction system confirms it. Customers should have a clear means to challenge or correct the category assigned to them.

Collect only context needed to resolve the enquiry

A compatibility query may need a model and revision number; a dispatch query may need an appropriately verified order identifier. Neither inherently requires every available piece of personal information. Avoid forms that ask all customers for full addresses, contact numbers and unrelated account details merely because the same form handles many different enquiries. Define the purpose of each field and remove unnecessary collection.

Restrict access to sensitive information by role. When a case moves to a human colleague, share enough context to avoid unnecessary repetition without disclosing unrelated customer records. Data minimisation, transparency, appropriate security and accuracy are principles addressed by the ICO's AI guidance. The choice of a particular workflow does not itself prove compliance.

Use reliable product records rather than plausible answers

Make the approved catalogue, compatibility tables, current prices, shipping information and relevant policies the authority for customer answers. If an item has multiple revisions, record which revision each specification describes. Otherwise, a helpful-sounding answer may accidentally use a current model's features to answer a question about an older purchase. Make contradictory or missing entries visible and route them to the person who can investigate.

Keep customer-facing responses proportionate to evidence. A model can summarise a published fact but cannot establish a missing test result. If an answer would require a safety assessment or unavailable compatibility information, explain the uncertainty. Misleading certainty damages trust and may have consequences beyond a wasted support interaction.

Illustrate the human handover with licensed context

A human reviewer remains important when a question involves uncertainty, complaints or exceptions. The following photograph provides a contextual example of a customer-service worker using a headset; it does not depict Epic Creative employees or verify any software's performance.

Make every handover useful to the receiving person

Provide the original question, provisional category, relevant product or order context, previous responses and the unresolved decision. Do not merely forward a note saying that automation could not help. The receiving staff member needs to know what has already been verified, which information still requires checking and who is responsible for the next action.

Use plain statuses such as received, awaiting information, assigned, in review and resolved. Record actual ownership so that a customer does not disappear between queues. A technical expert may assess compatibility, an authorised service colleague may decide a return, and an appropriate responsible person should receive a potential product-safety concern. A classifier's confidence should never be treated as authority to bypass those people.

Test ambiguous and high-consequence cases

Test more than routine greetings. Include misspellings, several issues in one message, old product names, order references that cannot be verified and customers who correct an initial label. Include possible overheating, damaged goods and uncertain compatibility, where an erroneous answer has a different consequence from a late response about colour choice. The correct behaviour may be to ask one targeted question or escalate.

Review privacy boundaries by testing that a vague order number cannot reveal somebody else's purchase details. Review whether messages move to an authorised team without exposing entire histories unnecessarily. A successful test should confirm that the system follows its limits, not merely that it generates a fluent reply. Where an issue appears repeatedly, repair its source data or workflow before considering additional automation.

Measure the outcome rather than the label count

Track how frequently provisional categories match the issue confirmed by a reviewer, whether urgent reports reach the right person, whether customers repeat information and whether problems are actually resolved. Separate an automated acknowledgement from a completed service outcome. Report significant misroutes, especially those involving safety or account privacy, instead of hiding them inside an overall accuracy percentage.

Use reviews to improve the categories and the catalogue. If buyers routinely ask whether a cable is included, a clearer product listing may reduce the need for support. If staff keep correcting the same label, revise its definition or provide better context. It is not necessary to claim an unmeasured efficiency uplift to justify a practical pilot.

Maintain accountable routing as products change

Give a named owner responsibility for the category scheme, approved response material, escalation routes and retention rules. Review these as product ranges, suppliers and marketplace processes change. Keep a traceable record of material decisions and provide staff with a straightforward correction route. Make sure outdated information does not continue to be presented as current merely because it was once an approved answer.

For potentially hazardous products, the UK government's product-safety guidance is relevant background to separate from routine customer service. For personal-data uses of AI, consider the ICO's guidance in the actual processing context. A modest, reviewable triage workflow can help a small team provide more consistent responses while retaining human responsibility for the decisions that require it.