Planning an AI assistant for company information

Choose the first task

Start with a recurring question or drafting job. Examples include finding the current product specification, explaining an approved procedure or preparing a meeting brief from permitted records. Write down what a useful response should contain and what the user needs to verify.

For a policy question, that might be a short explanation with a link to the current source. For a meeting brief, it might be a draft with missing information clearly identified. Choose a task your team can evaluate against examples of the work it already does.

An AI assistant supports a person with information and tasks. A chatbot provides a conversation interface. An AI agent can carry out a sequence of tool-based steps. Decide which behaviour the first version needs before choosing a platform.

Prepare the information

Select a small set of current material and name the person responsible for each source. Remove superseded versions from the pilot, resolve conflicting instructions and check that essential information is present as readable content.

Record where each item came from and how often it changes. A catalogue updated every week needs a different maintenance process from a stable operations manual. If the assistant uses a copy of a source, agree how that copy will be replaced when the original changes.

• Which sources are approved for this task?
• Who owns them and resolves conflicting information?
• Which version should be treated as current?
• What should the assistant do when the source does not answer the question?

Define access and actions

The source connection must respect the user’s access. A person who cannot open a restricted record should not receive its contents through the assistant. Access checks need to be part of the application and retrieval design, rather than relying only on an instruction written to the model.

Keep the first pilot read-only where that meets the task. If a later phase needs to send email, update a CRM or create a booking, define the allowed action, required fields, approval step and failure handling separately.

Documents and external pages can also contain instructions that try to change a model’s behaviour. OWASP identifies this as a prompt injection risk and recommends limiting privileges, separating untrusted material and requiring human approval for consequential actions. These controls reduce risk; they do not justify a promise that every answer or action will be correct. Read OWASP’s prompt injection guidance.

Test a limited pilot

Prepare representative requests before the build is evaluated. Include ordinary questions, ambiguous wording, missing answers, outdated information and requests for restricted material. Keep the expected source or outcome beside each test so reviewers have a consistent basis for checking the result.

Evaluate whether the assistant uses the right source, makes unsupported statements, shows useful references and recognises when it needs help. Record the type of failure instead of treating a fluent response as a successful result.

For example, an illustrative product assistant might be asked for a specification that is present in the catalogue, a discontinued option and a confidential commercial note. The evaluation should check the correct response to each, including declining or escalating requests outside the user’s permitted scope.

Plan updates and handover

Before wider use, agree who maintains the sources, reviews unresolved questions and checks changes to integrations or models. The handover should describe how to report an error, disable an unavailable connection and update the evaluation set when business requirements change.

The operating estimate should state the assumed users, request volume and connected tools. Platform subscriptions, model usage and support should be visible separately from the initial build. Compare the pilot’s usefulness with the effort needed to maintain it before expanding the scope.

If the first task or source material is still unclear, a consultancy review can define the project before development. You can also discuss an AI assistant brief directly.