AI Agent

An AI agent that starts with your business.

Visitors ask questions in their own words. Engage finds the passages of your published content that bear on the question, answers from those, and tells you which page it used.

One question, end to end

What happens between the question and the answer.

Six stages. None of them is a model being asked to remember your business.

  1. In their own words, without knowing how your site is organised or which page holds the answer.

    “Do you support emergency bookings?”

  2. The text is turned into a vector so it can be compared against your content by meaning rather than by keyword. This is retrieval, not a judgement about the visitor.

    1,536-dimension query vector

  3. The closest passages from your indexed pages are selected. Only your content is searched.

    3 passages from 2 pages

  4. Those passages, and the conversation so far, are given to the model as the material it may use.

    Services → Emergency support

  5. The model writes from the supplied passages. If they do not cover the question, it returns an explicit refusal instead.

    supported · 2 citations

  6. The conversation and every model call are stored. On your own website the conversation is kept for your workspace, though the console cannot show it yet; in a public demo it goes when the demo expires.

    conversation + usage row

    Partly available

The agent never reaches for general knowledge to fill a gap. If the retrieved passages do not answer the question, that is the answer.

What that buys a visitor

Answers that behave the way a person would expect.

Natural questions

No navigation, no site map, no guessing which page holds the answer. A visitor asks the thing they came to ask.

Follow-ups keep their thread

The exchange so far is part of the context, so “and on weekends?” means what it should within a conversation.

Grounded in your pages

Answers are written from passages retrieved out of your own published content, not from what a model assumes about your industry.

It will not invent your policy

Pricing, guarantees, hours and eligibility are the things a confident model gets wrong. Without a source, Engage declines.

Continuity

A visitor who comes back to your site in the same browser, while their visitor token lasts, picks up the conversation where they left it. In a public demo, the conversation ends when the demo expires.

On your own website

An embeddable widget puts the agent on your site: one snippet, shown only on the addresses you allow.

Reliability

Built to answer — and to refuse.

Useful AI doesn't need to pretend it knows everything.

Supported answer

Same-day service is available in Miami for selected areas.

Cited to Services → Emergency repairs.

Insufficient information

I don't have enough information to confirm that.

No evidence found. The agent does not guess to fill the gap.

Human escalation

I can have someone from the team help you with that.

The question and its context stay attached to the conversation.

Reliable beats confident-sounding.

A distinction worth making

Understanding a question is not the same as reading a customer.

Both get called “intent”. One of them is retrieval, and it is built. The other is commercial classification, and only its narrowest form exists here: noticing that a visitor has asked to buy, book or be contacted, so the widget can offer them a form.

Commercial intent scoring — not built

  1. Scoring how likely a visitor is to buy
  2. Treating a visitor as a lead before they ask to be contacted
  3. Ranking one lead above another
  4. Routing an opportunity to a person
  5. Recording a lead the visitor did not send

Query understanding — built

  1. Turns the question into a vector
  2. Finds the passages closest in meaning
  3. Handles wording your pages never used
  4. Keeps the thread across follow-ups
  5. Decides what the model may write from
Specifics

How it actually behaves.

How does a refusal look to the visitor?

As a plain sentence saying it could not verify that from this business's own information, with no citations attached. It reads as an answer rather than as an error, because refusing is the correct outcome and not a failure of the system.

Can it tell an uncertain answer from one needing a person?

Internally the domain declares four outcomes — supported, clarification, unsupported and handoff. The public API returns two: supported, or unsupported. So a visitor sees a grounded answer or a refusal, and the finer distinction is not something we can claim they experience today.

Does it hand off to a human automatically?

Not on its own. When a visitor asks to buy, book or be contacted, or asks something your content does not cover, the widget offers them a form, and the details they leave are emailed to your owners and admins with a verified address. Nothing routes a live conversation to a person, and nothing reaches your team unless the visitor sends it.

Does it remember visitors between sessions?

Only in the same browser. A visitor who returns to your site picks up their earlier conversation, for as long as their browser keeps the token that names it. There is no visitor profile, and nothing is carried across browsers or from one conversation into another.

Next step

Ask it something only your website could answer.

That is the test that separates a grounded agent from a general one, and it takes about a minute.

Build your agent