Most businesses that say they have added AI have added a chat window. It answers questions, sometimes accurately, and then does what every widget before it did: it hands the visitor back to a form.
The gap between a chatbot and an agent is action. A chatbot answers. An agent answers, then updates a record, books a slot, drafts a quote, escalates to a human or triggers a sequence — and it leaves a trace of having done so.
Design the actions before the personality
Most agent projects start in the wrong place: tone of voice, name, avatar, greeting. None of that determines whether the agent is useful. Start instead with an explicit list of the actions it may perform, and the conditions under which it may perform them.
- Read-only: answer from published content, show availability, explain scope or pricing logic.
- Write with confirmation: capture an enquiry, book a slot, request documents, update a preference.
- Write autonomously: send a confirmation email, tag a record, assign an owner, log the conversation.
- Escalate: hand to a human with the full transcript and the fields already collected.
Anything not on that list, the agent cannot do. That constraint is not a limitation — it is what makes the agent safe to expose to the public.
Constrain the knowledge source
An agent grounded in your published, maintained content is easier to trust and dramatically easier to correct than one improvising from general knowledge. When the answer is wrong, you fix the page, and the agent is fixed too.
That has an underrated consequence: your content becomes the control surface. The team who owns the website owns the agent's behaviour, without touching prompts or code.
In practice this means three things. Keep the source content structured and current. Give the agent an explicit refusal path for anything outside it. And make it cite or link the page it is answering from, so a visitor can verify and a colleague can audit.
Handoff is a feature, not a failure
The most common design mistake is treating human escalation as an admission of defeat, so the agent loops instead of transferring. Visitors notice immediately, and trust does not recover.
A good handoff carries everything forward: the transcript, the fields collected, the intent classification and the reason for escalation. The human picks up mid-conversation rather than starting over, which is precisely the moment the visitor decides whether you are competent.
Set explicit escalation triggers: two failed attempts on the same question, any pricing negotiation, anything involving complaints, refunds, legal or medical judgement, and any explicit request for a person.
Instrument everything
If you cannot see what the agent was asked and what it did, you cannot improve it — and you cannot defend it if something goes wrong. Logging is not an optional phase two.
- Every question, with a topic classification you can group and count.
- Every action taken, with the record it touched.
- Every refusal and every escalation, with the trigger.
- Resolution outcome: answered, escalated, abandoned.
Within a fortnight that log tells you what your visitors actually want, which is usually different from what your marketing assumes. The top ten unanswered questions become next month's content plan.
A realistic first build
The agents that survive contact with real users are narrow. Pick one high-volume, low-risk workflow and take it end to end.
A service business might start with: answer scope and coverage questions from the services pages, collect job type and location, show real availability, book the visit, confirm by email, and escalate anything about price negotiation to a human. That is a week or two of work, not a quarter.
Once that path is reliable and logged, widening it is incremental. Starting wide, by contrast, produces an agent that is vaguely helpful everywhere and trusted nowhere.
What good looks like after 90 days
A working agent shows up in numbers your team already tracks: fewer repeat questions in the inbox, shorter time-to-first-response, more enquiries arriving with complete context, and a measurable share of bookings made outside working hours.
If none of those move, the problem is almost never the model. It is that the agent was given a voice but no actions.