What a useful AI assistant needs beyond a chat window

A chat box is the easy part. Useful assistants depend on approved knowledge, clear limits, a route to a person and a way to tell whether answers are right.

Close view of hands typing on a laptop keyboard, with a blurred messaging window on the screen.

Adding a chat window to a website or an internal tool takes an afternoon. Making it useful takes deliberate decisions about what it knows, what it may do and what happens when it cannot help.

This article sets out the parts we look for before an assistant is shown to customers or staff. None of them is specific to one vendor or model.

Start with a narrow job

“Answer questions about our business” is not a job. “Help a prospective customer understand which service fits their problem, and pass a summary to the right person” is. A narrow job tells you which knowledge to provide, which questions to decline and how to judge success.

Give it approved knowledge, not everything

An assistant can only be as accurate as the material it draws on. Decide which sources are approved, who owns each one and how changes reach the assistant. Published service pages and reviewed answers are good candidates. Draft documents, internal notes and personal data usually are not.

  • Keep a list of approved sources and an owner for each.
  • Exclude drafts, demonstration content and anything not yet reviewed.
  • Ask answers to link back to the page they came from, so people can check them.
  • Refresh the knowledge when the source changes, not on an occasional manual export.

Decide what it must not do

Limits are part of the design. Write down the topics the assistant should decline, the actions it may never take on its own, and the information it must not request. A clear, polite refusal that points to a person is better than a confident guess.

Risk frameworks can help structure this conversation. The NIST AI Risk Management Framework, for example, organises the work around governing, mapping, measuring and managing AI risks. [1]

Design the route to a person

Every assistant reaches the edge of what it should answer. Design that moment as carefully as the conversation itself: who receives the handoff, what context travels with it, and what the person on the website is told about when to expect a reply.

  • Offer a person at any point, not only after the assistant fails.
  • Pass a short summary and the relevant page, so nobody has to repeat themselves.
  • Be honest about response times and working hours.

Measure whether the answers are right

Before launch, write a set of realistic questions with the answers you expect, including questions the assistant should decline. Run them whenever the knowledge, prompt or model changes. After launch, review a sample of real conversations with the owners of the source content, and record the corrections you make.

Privacy belongs in the same review. Tell people what is recorded, keep only what you need, and agree how long conversation records are kept.

A short checklist

  1. A narrow job, written in one or two sentences.
  2. Approved sources, with owners and a refresh path.
  3. Written limits and a polite way to decline.
  4. A route to a person that carries context.
  5. A test set of questions, run before every change.
  6. Clear privacy information and retention rules.

Sources

Product-specific statements refer to the official documentation below. Other guidance is Sixsparks’ proposed approach.

  1. AI Risk Management Framework — National Institute of Standards and Technology (NIST). Voluntary framework for managing risks of AI systems, organised around the Govern, Map, Measure and Manage functions.

A next step

Working on something similar?

If you are considering an assistant for customers or staff, we can help define its job, its knowledge and its limits before anything is built.

Explore automation and applied ai Start a project