Define What Your Website Agent Should Answer—and When to Hand Off
Define What Your Website Agent Should Answer—and When to Hand Off
A website agent is most useful when its job is clear. It can answer questions grounded in your published information, help visitors find the right details, and acknowledge when the answer depends on something it cannot know. The work starts before you add a chat widget: decide which questions belong in scope, what source supports each answer, and what the agent should do when the source material is missing.
This is website agent scope in practice. A clear scope makes answers easier to check and gives you a concrete way to set AI agent answer boundaries.
1. Group visitor questions by the task they support
Start with the decisions visitors are trying to make. List the questions they ask, then group them by task instead of treating every phrasing as a separate topic. This visitor question planning helps you see whether your website has enough information to support an answer.
Common groups include:
- Understand the offer: What does the service include? Who is it for? What problem does it address?
- Compare options: What is the difference between plans, products, or service tiers?
- Check practical details: What are the requirements, schedule, availability, or terms?
- Prepare to take action: What information should I have ready? Where do I start?
- Resolve an issue: What should I do if I cannot log in, find a document, or complete a step?
Use real questions from sales conversations, support messages, and website feedback where available. Keep the visitor’s wording. “Can I change my booking?” may point to a specific policy, while “What if my plans change?” could involve exceptions that require a person.
For each group, write down the visitor’s task and the outcome a useful answer should provide. For example, a question about plan differences should help someone understand the published distinctions—not decide which plan is right for their particular situation.
2. Match each in-scope question to a current source
An answer should have a clear home in your website content. Match each question you want the agent to handle to a specific, current page or document. Record the source next to the question so you can spot gaps before they become confusing answers.
| Visitor question | Intended task | Source to check |
|---|---|---|
| What does the service include? | Understand the offer | Current service or product page |
| How do the plans differ? | Compare options | Current pricing page |
| What should I prepare before applying? | Get ready to act | Application guide or checklist |
| How do I request a change? | Resolve an issue | Current policy or support instructions |
A source match is not just a title that sounds relevant. Check that the page directly answers the question, uses current terms, and does not conflict with another page. If two documents give different instructions, correct the content or narrow the agent’s scope until the conflict is resolved.
Avoid treating an old email, an unpublished internal note, or a vague paragraph as support for a broad answer. If visitors frequently ask something your pages do not answer, add or update the source first. The agent can only give a reliably grounded response when the relevant information is available to it.
For setup and product details, see the Simple Agent documentation. Keep the source list simple: question, decision or task, approved page, and any relevant boundary.
3. Mark questions that need a person
Not every useful question is suitable for an automated answer. Mark questions that require personal judgment, details the visitor has not provided, or an action that your website content does not explain.
For example:
- Personal judgment: “Which option is best for my situation?” may depend on preferences, constraints, or professional advice.
- Missing information: “Am I eligible?” may require details the visitor has not shared or information your site does not publish.
- An exception: “Can you make an exception to this policy?” cannot be resolved by repeating the standard policy.
- An external action: “Can you change my account?” requires an action beyond explaining the instructions on a page.
- A high-impact decision: Questions about health, legal, financial, or similarly consequential choices should not receive a confident answer based only on general website text.
For each boundary, write a practical response rule. The agent might point to the relevant policy, ask the visitor to contact your team, or explain that it cannot determine the answer from the available information. Make the next step specific: name the page, form, or contact route that exists on your site.
These are your AI agent handoff rules. A useful rule says both what the agent should avoid and where the visitor can go next. Do not promise a live transfer, ticket, booking, or other action unless your configured setup actually supports it.
4. Try representative questions in your configured agent
Before embedding anything, prepare a small set of representative questions from each group. Include a straightforward question with a clear source, a question that could be phrased ambiguously, and one that should fall outside the intended scope.
Before you publish, save these questions as tests on the agent configured with your own sources and compare each answer with the cited page.
When retrieved source material is available, inspect the answer and its citations. Check whether the cited source actually supports the specific claim. A citation is useful only when it leads to relevant source material; it does not make an unsupported answer reliable. If the answer states more than the source says, or skips an important qualification, treat that as a problem to address in the source or scope.
Also try questions the agent should not answer directly. Does the response make clear what it cannot determine? Does it point to a practical next step? Testing your own sources and boundaries requires your configured agent.
5. Embed the widget and refine scope from real questions
Once your questions, sources, and boundaries are ready, add the script-tag widget to your website and review how it behaves with your own configured agent.
After visitors begin using it, review the questions they actually ask and compare them with your plan. Look for patterns:
- Questions that belong in scope but have no clear source
- Sources that are outdated, incomplete, or in conflict
- Repeat questions that suggest a missing page or unclear wording
- Requests that need personal judgment or an action outside the published content
- Answers that cite a page but draw a conclusion the page does not support
When you find a gap, decide whether to update the source, narrow the allowed scope, or make the handoff instruction clearer. Then try the representative question again with your configured agent. This keeps refinement tied to source material from your website and its visitors.
A short, current source can be more useful than a large collection of loosely related pages. Keep the mapping between questions and sources readable, and revisit it when your offer, policies, or instructions change. The widget is currently measured at 18.56 KiB gzip; your main planning task is still deciding what it should answer and what it should leave to your team.
Start by listing the questions visitors ask most often, then match each one to a current source and a clear boundary. When that map is ready, pick a Simple Agent plan and review answers from your configured agent against your own content.
Ready to build your AI agent?
Guided setup in 90 seconds. One flat price, published monthly volume.
Create my agent