Sales and qualification agent

A visitor asks whether a service can help their business. A useful sales agent helps them understand the fit and, if they want, reach the right person. It is not a machine for extracting contact details at any cost. This guide follows a fictional company, Fieldwork Scheduling, which sells appointment-management software to small organisations.

A lead is a person or organisation that has expressed possible interest. Qualification means learning enough about their needs to suggest a sensible next step. A CRM, or customer relationship management system, stores sales records and follow-up activity. Think of the agent as a receptionist who can prepare a meeting brief, not a salesperson authorised to make every commercial promise.

Define a useful outcome #

Fieldwork wants interested visitors to receive an accurate explanation, an appropriate demonstration offer, or an honest statement that the service may not fit. The agent must not invent discounts, promise features, or claim a deadline that does not exist. “No, thank you” is a valid outcome, not a failure to overcome.

Separate three responsibilities before writing the first instruction:

A conversion is a defined next step, such as an agreed demonstration. Define it explicitly: sending a calendar link is not the same as booking a meeting, and a booking is not a sale. Otherwise the dashboard can reward activity that creates no benefit for either party.

Build the conversation around the visitor #

Ask one question at a time and reuse answers already given. If someone says “We need this next month”, do not ask for timing again merely because it is next on a form. A conversational assistant should reduce effort compared with a form, not hide a long mandatory form inside chat bubbles.

The instruction draft can be simple: “Help visitors assess whether our documented service fits. Ask relevant questions without pressure. Do not infer personal characteristics, invent terms, or send follow-ups without the required permission. Record unknown answers as unknown. Offer a person when asked and respect a refusal immediately.”

Score evidence, not people #

A lead score is a prioritisation aid based on stated business facts. It is not a measure of a person’s worth or a reliable prediction of buying behaviour. A hot lead simply means one that meets the company’s documented readiness criteria and wants timely contact.

For this case, the rubric considers product fit, a concrete problem, and stated timing. Budget can improve the completeness of the brief but an unknown budget does not mean “unqualified”. A human should be able to read the same evidence and understand why the agent suggested its route.

Avoid scoring from names, accents, disability, age, or other sensitive characteristics and unjustified proxies. A proxy is an indirect substitute, such as using a postcode to guess income. Review the rubric for unfair effects, particularly if it influences access to services or commercial terms.

Related: AIVAX decision models can evaluate defined questions, while structured responses provide a machine-readable output format. A valid structure only shows that the fields fit a required shape; it does not prove that the facts or recommendation are correct.

Put approval before the write #

A CRM write is an external action: it changes a business record. Keep the proposed summary separate from the saved record. The proposed data should include the visitor’s stated need, unknown fields, approved contact method, and evidence for the routing recommendation. Avoid storing the entire chat when a short summary is sufficient.

Visitor agreement and staff approval serve different purposes. The visitor confirms what will be shared; the staff member checks whether the business should create or update that record. Neither automatically provides every legal basis needed for processing personal data. Establish the appropriate basis, privacy notice, retention period, and channel rules with the responsible team.

Read Human in the loop for approval design. The approval screen should show exactly what will change. If important fields change after approval, ask again rather than treating the old approval as unlimited permission.

If the CRM request times out, the application may not know whether the write happened. Check for the existing result before retrying so the visitor does not become several duplicate leads. A webhook is an event notification sent from one system to another; it can trigger follow-up after a confirmed change. See Webhooks, events, and automations for the connection pattern.

Evaluate the funnel without rewarding pressure #

A funnel shows how many people move through successive stages. Use clearly defined stages and the same observation period. The chart below is illustrative, not a conversion forecast or a claim about any product.

Review whether sales staff received accurate briefs, visitors understood the handover, and declined contact was respected. Track incorrect product claims, duplicate records, unwanted messages, and complaints alongside meetings. A rise in bookings accompanied by more pressure complaints is not an unqualified improvement.

Test missing budgets, contradictory answers, requests for unsupported features, withdrawal of contact permission, CRM failure, and an attempt to bypass staff approval. Pilot with reviewers available, define who owns each follow-up, and stop automated writes if approval or permission checks fail.

What’s next: build an internal knowledge assistant, where the central question is who may see which information.