Before support staff paste a customer conversation into an AI tool

Define the purpose and the minimum content first

Start with the task, not the tool. Summarizing a complaint theme, drafting a reply, and translating a message each need different content, and none of them obviously needs a customer's full name, address, order identifiers, or payment references. Under the UK ICO's articulation of the data minimisation principle, personal information should be adequate, relevant, and limited to what the purpose requires. That principle is UK guidance under review, but the underlying discipline travels well: if the purpose works with a redacted extract, the full conversation is not justified. The extract is still customer-derived information, so moving it to an outside tool needs the same authorization as any other transfer; redaction shrinks what is exposed without deciding whether the transfer is permitted.

Write down what a minimized working extract would look like for each task under evaluation: the issue type, the relevant order facts without identifiers, the customer's question in substance. While authorization is pending, staff can develop and test their drafting approach with invented, non-customer material only. What they cannot do is place customer-derived content, however heavily redacted, into a tool the owner has not approved.

Verify the actual tool, plan, and role before trusting it

ICO guidance on controller and processor roles makes a structural point that matters here: roles depend on who determines the purposes and means of each processing activity, not on what the vendor calls itself. When your staff paste conversations into a tool, you need to know what the provider does with that input under the specific product and plan you hold: whether inputs are retained, for how long, whether they train models, who can access them, and what the contract says. Those facts vary between products and tiers and change over time, so verify them from the provider's current official documentation for your exact plan.

Do not accept indirect evidence as a substitute. A colleague's assurance, a marketing page headline, or a remembered settings screen are not the current terms. If the retention and training behavior of your plan cannot be established from official documents, that is an unresolved finding, and unresolved means the transfer is not authorized yet.

Check authority on both sides of the transfer

Your own published privacy statements and any notices given at collection describe what you told customers about how their information is handled. A support-drafting use that sits outside those statements is a representation problem as well as a data-handling one. The policy owner should compare the intended tool use against the promises the store has actually made, with qualified privacy advice where the lawful basis or jurisdiction is uncertain.

Also confirm the tool's terms permit the use you intend and that your account is a business arrangement rather than a personal account an employee signed up for. An individual employee's personal AI account holding customer conversations is an access and retention problem regardless of that tool's terms.

Record the authorized pattern and the fallback

The owner's decision should be written as an operational rule: which tasks, which tool and plan, what content may be included, what must be redacted, and who approved it. Staff need a concrete permitted pattern, not a principle. Include the fallback: when a task does not fit the pattern, staff work inside existing support systems or escalate for a decision rather than improvising, and any experimentation outside those systems uses invented, non-customer material only.

Revisit the decision when the tool's terms change, the plan changes, or the tasks expand, and record the review date. If you want help aligning what your public privacy and policy pages say with a documented internal practice, that is a scoped website-review question you can raise with Prism; confirm the pages, scope, fees and terms before work begins.

Support-data handoff sheet

Complete before any customer material is transferred, and review on tool or plan change. Record evidence references and settings states, never conversation content. A tool is usable for a task only when every row is resolved for that task, and redaction alone does not resolve one.

Worksheet entries are not submitted by Prism’s worksheet and are not saved by the site. Use record types, availability, anonymized observations, or match/mismatch results. Do not enter government identifiers, customer names or addresses, customer messages, receipt-access links, card or bank details, passwords, or keys. Send sensitive documents only through the provider’s verified secure channel.

These are temporary notes. Leaving or reloading this page may clear them. Worksheet entries are not sent automatically. If you copy notes into the consultation message and submit the form, Prism receives them as part of your request.