Decide what business data an AI tool actually needs

Make three working lists: approved public facts, internal operating information and sensitive records. Public facts might include services and business hours. Internal information might include assignment rules. Sensitive records might include private customer details, financial information or access credentials.

The lists are not universal classifications. A detail that is public for one business may be private for another. Assign someone who understands the information to decide what may be used, for which purpose and by whom. Avoid treating everything in a shared folder as automatically approved.

An assistant answering general service questions may need an approved FAQ, service area and escalation instructions. It probably does not need every customer conversation or the company’s full financial history. Start with a clean, current reference document rather than a bulk export.

For example, a repair shop can provide inspection steps, opening hours and instructions for requesting an estimate. Customer names and payment details add little value to that general task. If a later use case needs account-specific information, design that access separately with a defined identity and permission check.

Check the provider’s current terms, data controls and account settings for the exact product you intend to use. A consumer chat plan, business workspace and API can have different settings and agreements. Do not assume a familiar brand means every product handles information identically.

Write down where the data will live, who may access it, what connected tools can do with it and how to remove it when the task ends. If those answers are unclear, keep the pilot on invented or approved public information while the relevant owner reviews the setup.

• What task requires this information?
• Which fields can be removed?
• Who approved this use?
• Which destination and account will receive it?
• How will outdated information be corrected or removed?

Give each reference set an owner and a review date. A service area, policy or offer can change after the assistant is configured. Without an update process, an initially correct answer can become wrong while still sounding confident.

Test questions that should not be answered from the available material. The expected response may be to ask a clarifying question or hand the conversation to a person. Keep uncertainty visible. A useful assistant should know the limits of its approved information rather than fill every gap with a plausible answer.