Designing AI assistants people can actually trust

Start with a task, not a model

Before choosing a model or a chat interface, identify the user, the question they need answered and the decision or action that should follow. “Add an AI chatbot” is not a requirement. “Help a customer find the correct support policy and route unresolved cases to a person” is much closer to one. Write down the accepted answer, the unacceptable answer and the escalation path.

Give the assistant a controlled source of truth

For company-specific questions, an assistant should draw from approved service descriptions, policies, product documentation and current business information. Keep those sources owned and reviewable. If information is missing, the assistant should say it cannot verify the answer rather than fill the gap with a plausible invention. Retrieval does not automatically make a system correct; source quality, freshness and retrieval behaviour still need testing.

Evaluate beyond a friendly demo

Create a small test set before release. Include common questions, ambiguous requests, outdated information, misspellings, prompt injection attempts, requests for confidential data and questions the assistant should not answer. Score whether it used the right source, stayed within scope, gave a useful fallback and handed off when needed. Repeat those tests whenever the knowledge base or model changes.

Design for privacy and human control

Collect only information needed for the task. Tell visitors what is stored and where it goes. Do not invite passwords, payment credentials or sensitive personal details into a general chat. For actions with consequences—submitting an enquiry, changing an account or issuing a refund—use confirmation and an appropriate human review step.

Measure whether the experience helps

Track task completion, unanswered questions, escalation rate, user correction and harmful or misleading responses. A high conversation count is not a success measure by itself. A small assistant that reliably solves one narrow problem can be more valuable than a broad assistant that improvises across every topic.

Putting the ideas into practice

The dependable path is deliberate: define the task, curate the knowledge, test difficult cases, protect user data and make human help visible. Treat the assistant as a product feature with an owner and an evaluation cycle—not a decorative widget.