Conversational AI

A website conversation should help a visitor answer a real question or reach the right person. FyerX designs conversational AI for service discovery, qualification and handoff, with approved sources, explicit limits and a reliable path to human help.

Discuss your marketing brief

When to bring in FyerX

Consider conversational ai when you are dealing with visitors asking the same service questions before they can enquire. Our starting point is a bounded visitor intent. We work with customer or marketing operations to establish what evidence is available, what decision depends on it and which improvement the team can realistically put into practice. The first scope defines a reviewed conversation design and the way your team will judge its usefulness.

Intent design

We start with site search, chat history and repeated sales questions. The immediate task is to group enquiries by goal and decide which need human handling. Review what visitors actually ask immediately before making an enquiry. A question about support coverage, a request for a quote and a complaint after purchase each need a distinct conversational route.

Knowledge preparation

For knowledge preparation, the starting material is approved service information and policy content. We define source ownership, freshness and answers for unsupported questions. Use examples of what the assistant must decline or hand over. Unpublished pricing, private account information and speculative delivery commitments all need boundaries and an appropriate route to a person.

Qualification and handoff

Test a conversation that changes subject halfway through. The original intent should not force a user into a form after they ask for human support or clarify that they are an existing customer. We use sales intake fields and team availability to collect only relevant context and transfer a useful summary.

Conversation evaluation

Have a person review a sample of completed conversations. A short chat may signal efficient resolution or an abandoned dead end. Read the transcript alongside handoff outcomes before claiming success. The work draws on representative questions, ambiguous requests and edge cases. With that in hand, we test accuracy, recovery and accessibility before expansion.

Illustrative engagement

Route a pricing question to the right team

  1. Starting point

    A visitor asks for a quote that depends on contract size, region and implementation scope.

  2. Decision

    Use approved information for the initial answer, gather only what sales needs and transfer unanswered questions with their context.

  3. Deliverable

    A documented conversation flow with escalation, consent and transcript review.

A typical brief

A visitor asks whether a provider supports a platform version. A generic assistant improvises an answer. A grounded conversation checks approved service material, explains what is known and offers to send the remaining question to the appropriate technical team.

Give a human the context needed to continue

A visitor seeking implementation support may already have named the platform, explained the business problem and asked for a callback in a particular time window. The handoff should preserve that information in a concise summary with the original question available for review. The receiving rep needs to know which parts were explicitly supplied and which were inferred. If the assistant guessed the wrong intent, the rep should be able to correct it without asking the visitor to begin again. Discuss how the team handles missed callbacks, out-of-hours requests and information that the visitor wants removed or corrected. Test one unresolved question with the actual receiving queue. Confirm the person responsible can see it and understands the expected response. Then test a routine question that should finish within the conversation. Check whether the answer contains approved facts and gives a clear next step without unnecessarily collecting contact details. Both cases matter. A chatbot that hands every visitor to sales creates work, while one that avoids handoff at all costs leaves difficult questions unanswered.

Writing the boundary between assistance and commitment

Choose one common commercial enquiry and agree exactly what the assistant can answer from approved material. It may explain a service, collect the buyer’s requirements and offer a handoff. A personalised quote, appointment guarantee or contractual commitment may need a person with authority. Define that boundary in the conversation design and test questions that approach it indirectly. For example, a buyer asking whether installation can finish before an event is asking about a real commitment, even if they never use the word guarantee. The assistant should capture the date, explain how the team will confirm feasibility and send the conversation to the correct owner. Review the transcript to confirm that the actual wording preserves that distinction.

What the receiving salesperson needs

A chatbot handoff should provide the question, relevant answers, consented contact route and any deadline the buyer has volunteered. Avoid transmitting irrelevant personal details or filling the CRM with speculative conclusions about buying intent. Give the salesperson a concise summary and a route to the original conversation when more context is needed. Test a handoff from a returning customer, an existing opportunity and a new prospect using the same business email domain. Check record matching and ownership before enabling automatic follow-up. Define what the visitor sees if the CRM is unavailable and who reviews unresolved conversations. The quality test is whether a human can continue the exchange naturally without promising anything the assistant was not authorised to offer.

A conversation test pack for the questions people actually ask

Select an enquiry journey with a clear owner. A visitor may need to identify the right service, explain a delivery constraint and arrange a conversation with sales. We document what the assistant may answer from approved material, what it needs to ask, which details may be recorded and when a human should take over. The test pack includes an ordinary enquiry, a vague opening, a correction to an earlier answer, a request outside scope and an explicit request to speak to a person. Reviewers score factual accuracy, the usefulness of the next question and whether the handoff reaches the right team with sufficient context. A technically successful transfer can still be a poor experience if the human asks the visitor to start again. At handover, the conversation owner gets approved responses, escalation routes, unresolved topics and a review queue for failed conversations. Keep a named business owner for changing product facts and a measurable way to discover where people abandon the exchange.

Test a conversation that starts with uncertainty

A visitor might say they need help with an application and only later clarify whether they want technical support, implementation services or a job. Offer a short clarifying question with meaningful choices. If the visitor changes direction, carry the useful context across and discard irrelevant qualification questions. During review, ask whether a human receiving the handoff can identify the original request and the point at which the assistant reached its limits. This tests a more realistic conversation than a script containing only happy-path questions the designer already knows how to answer.

Set the boundaries of a good resolution

An assistant handling a routine question can give a clear sourced answer and finish. A sensitive query or unusual commercial request should reach a person. Measure both routes against the intended outcome. A shorter average conversation is useful only if the visitor still received an accurate answer or a successful handoff.

Write the handoff before the greeting

Decide where an unresolved conversation will go and what the recipient needs to know. A technical question might need the product, version and use case. A booking request needs availability and contact preferences. If the human queue cannot accept that information, fix the handoff before expanding the range of questions the assistant attempts.

Test the gaps in the knowledge base

Ask about discontinued services, ambiguous terms, confidential pricing and an urgent support request. Evaluate whether the assistant recognises its limits and sets an accurate expectation of follow-up. A pleasant answer that invents an unsupported service is a serious product defect. Record corrections at the source so future answers improve consistently.

What to expect from intent design

For intent design, ask for an intent catalogue and escalation map. Use it with site search, chat history and repeated sales questions to examine a real decision. The issue it needs to prevent is forcing support enquiries through a sales qualification script. A reviewer should be able to trace a recommendation back to the source, explain where an exception belongs and say who takes the next action. Agree with customer or marketing operations which decisions they can make directly and which ones need another team.

What to expect from knowledge preparation

The output is an approved knowledge set and review schedule. To review it, pick an example drawn from approved service information and policy content and ask the person responsible to define source ownership, freshness and answers for unsupported questions. If the example ends up serving an outdated policy as current, examine the point where the decision became ambiguous. Capture the reason for any exception and incorporate it into the next version. This exercise tests whether the work can be operated by the people who inherit it.

What to expect from qualification and handoff

Before accepting a handoff contract and tested notification path, look at a borderline case. Use sales intake fields and team availability to identify when a routine recommendation could lead to asking the visitor to repeat the whole conversation after transfer. Ask the reviewer to describe the appropriate next step in their own words. A workable design lets another colleague reach a consistent decision using the same evidence, or makes the remaining judgement explicit enough to discuss.

Planning your engagement

What to expect from conversation evaluation

Under conversation evaluation, the deliverable is an evaluation pack with improvement priorities. Review an ordinary case and one that could result in judging success solely by the number of chats handled. Compare both with representative questions, ambiguous requests and edge cases and confirm the owner can test accuracy, recovery and accessibility before expansion. Record any change to the operating assumptions with the deliverable so later results can be interpreted against the approach actually used.

How to evaluate the work

Track successful resolution and qualified handoff by intent. Keep the starting point and reporting period beside the observed result. If customer or marketing operations changes how the team qualifies a bounded visitor intent, report the change alongside the figures. The useful comparison is between groups that entered under comparable conditions. A shift in who qualifies can change the reported percentage even if the underlying commercial performance stays the same. Review representative examples as well as the aggregate.

What to bring to a scope discussion

Bring an example that illustrates visitors asking the same service questions before they can enquire. Share the current version of site search, chat history and repeated sales questions and explain how the team handles qualification and handoff today. Describe who needs to use a reviewed conversation design and what is already scheduled that could affect delivery. With that context, FyerX can distinguish the immediate work from later enhancements and identify the access, evidence and approvals needed for a credible estimate.

Frequently asked questions

Conversational AI

What should the assistant do when it cannot answer?

Acknowledge the gap, preserve the visitor’s question and offer the agreed human route. If contact details are needed for follow-up, explain why and collect only the relevant fields.

Can the assistant qualify leads without frustrating visitors?

Yes, when qualification follows the visitor's stated goal. Ask the minimum useful questions, explain the next step and offer a human route for questions the assistant cannot resolve.

Which knowledge sources should it use?

Start with approved product, service and policy content. Give each source an owner and a review schedule. Mark information that must be obtained from a person or an authenticated system.

Should the assistant answer support and sales questions together?

It can recognise both intents, but the workflows often differ. Support may require account checks and a case owner. Sales may require product fit, territory and a callback preference.

How do we test the launch version?

Use real visitor wording, ambiguous requests, unsupported claims and handoff failures. Evaluate whether the response is accurate, whether the next step is achievable and whether the right human receives the context.

Can it operate outside business hours?

Design the after-hours experience explicitly. An assistant may answer approved questions and queue requests while giving the visitor an accurate expectation of when a human will respond.

Which enquiries should a chatbot hand to a person?

Define handoffs for pricing exceptions, complaints, account-specific commitments and questions outside the approved knowledge base. Give the receiving team the conversation context and a reference so the buyer does not repeat everything. Test the experience outside business hours as well, including what response time the assistant is authorised to promise.

How do you test a conversation before putting it on the website?

Build a test set from genuine enquiry patterns, including incomplete questions, follow-up questions and changes of topic. Ask reviewers to assess factual accuracy, useful next action and escalation quality. Test whether the assistant handles an unsupported request honestly and whether the contact details submitted during a handoff reach the right owner.

Can the assistant qualify a lead without making the conversation feel like a form?

Ask for the minimum information needed to take the next useful action. A buyer asking about delivery coverage should receive that answer before being asked for budget or a telephone number. Use the conversation context to avoid repeating questions and offer a direct route to a person when the buyer prefers it.

How should we maintain answers when an offer changes?

Name the owner of the approved answer and the sources the assistant may use. When pricing, service scope or availability changes, update the authoritative material and rerun affected conversations. Include a question using the old terminology to verify that the assistant can recognise it and give the current answer.

What does FyerX hand over for qualification and handoff?

The scope can include a handoff contract and tested notification path. It should describe the inputs, owner, decisions and agreed exceptions. Include a short walkthrough using sales intake fields and team availability so your team can continue the work after delivery.

What determines the scope of conversational ai?

The main considerations are the condition of site search, chat history and repeated sales questions, how much of knowledge preparation already exists and the level of support needed for conversation evaluation. We define scope around a reviewed conversation design, with dependencies and acceptance decisions agreed upfront.

Talk to FyerX about Conversational AI

Contact Marketing