Start with a narrow scope, verifiable answers and a visible human handoff. Measure correct resolution, not just conversations handled.
If your website does not explain hours, coverage or how to request a quote, a chatbot may be compensating for missing information. If those answers are already clear and the team still repeats them frequently, a useful use case may exist.
Before choosing a platform, define the question it should resolve and the error you cannot accept. The objective is to help customers move forward, not to display AI in a corner of the screen.
Choose between content, rules and AI
A clear page often answers a stable factual question. A form or guided menu may handle a fixed process such as choosing a location and inquiry type. Generated language is more useful when questions vary and reliable information exists to answer them.
Our recommendation draws on NIST’s approach to managing risk throughout AI use. This is not certification or a provider guarantee. Identify potential harms, evaluate behavior and maintain oversight.
Sources: NIST: Artificial Intelligence Risk Management Framework
Limit what the assistant can claim
Create an approved knowledge base covering services, coverage, hours, policies and common questions. Each answer needs a current source and someone responsible for updates. A commercial policy change must reach the assistant as well as the website.
Define topics requiring a person: exceptions, complex complaints, negotiations and decisions the business has not authorized for automation. Do not turn an estimate into a confirmed price. Acknowledging missing information and handing off is a useful response.
Separate conversation from execution
An assistant explaining a service has a different risk profile from one canceling orders or changing records. For the first pilot, avoid irreversible actions and unnecessary access to private information. Use the minimum data and permissions required.
If you add actions later, enforce permissions and confirmations in the system performing them. A user-written message must not override business rules or expose someone else’s conversations and information.
Test real questions and uncomfortable situations
Build an initial set of common, ambiguous, contradictory and out-of-scope questions. Add typos, both languages if you serve bilingual customers, and requests to speak with a person. Record expected answers and the sources supporting them.
Example test: a customer requests service outside your coverage area. Correct behavior: explain coverage and offer contact where appropriate. Incorrect behavior: invent a branch. Another test asks for a nonexistent discount. The assistant should follow approved conditions instead of creating a promotion.
Measure usefulness and full cost
Track correct answers, necessary handoffs, unresolved questions and material errors. A long conversation does not demonstrate satisfaction. Review a sample using defined criteria and appropriate handling of personal information.
Hypothetical example: software costs 60 monetary units monthly and needs four hours of supervision at 20 per hour. The included monthly cost is 140, before integration and other charges. Compare it with time actually saved while maintaining quality. Do not attribute a sale to the bot merely because it participated in a conversation.
Launch with a working human handoff
People should be able to request contact without completing an interrogation. If the team is unavailable, explain how and when requests are handled without promising instant responses. Retain necessary context so customers do not have to start again.
- Scope is documented and each knowledge source has an owner.
- The assistant acknowledges when it does not know.
- Testing includes errors, misuse and out-of-scope requests.
- Human handoff works on mobile and desktop.
- The assistant can be paused while other contact options remain available.
REFERENCES
Sources and review
Source documentation consulted on October 7, 2026. Platforms and features may change; check the official source before implementation. The English adaptation retains the source review date of the original article.
This guide provides general guidance. It does not guarantee search rankings or replace a review of your business, locations and data.