An assistant should solve a defined communication or workflow problem. “Add AI” is not a useful requirement. “Answer approved service questions, identify serious enquiries and connect them to sales” is specific enough to design and test.
Start with the smallest capability that can create value. More intelligence, integrations and automation should be added only after the initial assistant proves that customers and staff actually benefit.
Define one accountable job
Choose a primary outcome such as answering repetitive questions, guiding product selection, collecting a structured enquiry, helping staff find an approved policy or triaging support requests.
Write down what the assistant must never do. High-risk decisions, commitments, payments, confidential disclosures and exceptional cases usually need a person.
- Primary audience and business outcome.
- Questions and tasks included in scope.
- Actions that require human approval.
- Success and failure conditions.
Control the knowledge boundary
Approved knowledge should have an owner, version and review process. Website pages, price lists, policies and service descriptions frequently contradict one another; the assistant should not be expected to resolve governance problems by guessing.
When an answer is not supported, a professional assistant says so and offers the correct next step. Confident invention damages trust faster than a polite limitation.
- Use named, approved sources.
- Remove expired prices and policies.
- Record when knowledge was last reviewed.
- Provide an “I do not know” and escalation path.
Design natural handoff and privacy safeguards
The conversation should preserve context when it moves to WhatsApp, email, a form or a staff queue. The user should understand that a person is taking over and what information will be shared.
Public assistants should warn users not to submit passwords, secret keys, financial credentials or sensitive personal records. If regulated or confidential information is required, establish a protected channel outside the public chat.
- Clear identity and capability statement.
- Consent-aware collection of contact details.
- Direct human escalation for uncertainty or commitment.
- Minimal retention and role-controlled access.
Measure whether it earns its place
Useful measurements include questions resolved, qualified enquiries, handoff completion, unanswered topics, response time and customer satisfaction. Message volume alone does not prove value.
Review failures as product intelligence. Repeated unanswered questions may reveal missing website content, unclear pricing, a new service opportunity or a broken business process.
- Resolution and escalation rate.
- Qualified lead and appointment rate.
- Top unanswered or misunderstood topics.
- Cost per useful outcome.
- Human review of a sample of conversations.
Add paid intelligence only when justified
A controlled knowledge assistant can begin with rules, curated answers and search. Generative models become valuable when language variation, larger knowledge sets or more complex interpretation produces a measurable improvement.
Model cost is only one part of ownership. Include knowledge maintenance, privacy review, monitoring, testing, escalation and incident response in the commercial plan.
- Prove demand with a focused first version.
- Test quality against real questions.
- Compare paid intelligence with the simpler baseline.
- Maintain a safe fallback when an external service is unavailable.