AI lead qualification: usable fields instead of another paragraph

Retrospective · week of 5 October 2026

A polished summary can leave a salesperson with the same problem as before: reading the entire conversation to work out what to do next. For lead qualification, a few clearly defined fields are often more useful than a fluent paragraph. But every value needs a clear meaning, a known origin and a sensible way to represent missing information.

A correct format does not guarantee correct meaning

On 6 October, OpenAI launched the Decisions API beta for typed answers from text and images. Source: official changelog.

For the business, the real task is still defining useful information. An answer may follow the expected format while containing a mistaken interpretation. Qualification design should therefore start with the sales team: which details genuinely change the next action, and which simply add noise to the record?

Start with the next action

If the aim is to prepare an initial conversation, a short list may be enough: stated need, relevant tool, stated deadline, contact method and the question still to ask. Each field should have an accessible definition, permitted values where appropriate and a rule for missing information.

We recommend separating three states: stated by the prospect, not provided, and an estimate requiring verification. An absent deadline does not become “urgent” because the message is short. An undisclosed budget does not become zero. These distinctions prevent assumptions from being presented as established sales information.

Illustrative example: annotate rather than guess

Consider this fictional message: “We use two applications for incoming enquiries and would like to simplify things before our next campaign.” A demonstration record might contain:

• Stated need: simplify the handling of incoming enquiries.
• Tools involved: two applications; names not provided.
• Stated deadline: before the next campaign; date not provided.
• Budget: not provided.
• Suggested next question: which applications do you use, and when does the campaign begin?

The system can retain a short supporting excerpt for useful fields, within the company's agreed retention boundaries. It should not invent company size, purchasing authority or likelihood to buy. This example illustrates a data structure, not a real customer record. Its purpose is to show how incomplete information can remain useful without being artificially completed.

Test each field separately

Build a representative set of messages with expected answers approved by a salesperson. Include corrections made during a conversation, relative dates, several needs in one message and contradictory information. Evaluate extraction, missing-data handling and the usefulness of the next question separately.

An overall score can conceal a serious mistake. If the need is summarized correctly but the date is invented, the record should not pass without review. Define blocking errors and fields that may remain empty. Keep a “requires verification” state rather than forcing the system to choose an uncertain value simply to fill every field.

Connect qualification to sales follow-through

A useful record must reach the right place with clear ownership. Who takes the enquiry? How does a salesperson correct a field? How does that correction feed into future tests? Without this feedback loop, structured qualification is merely another formatting step, leaving operational problems unresolved.

Measure usable requests, clarification needed and accepted handovers rather than just the number of records created. Avoid rankings based on sensitive attributes or information unrelated to the requested service. Qualification should clarify the next action, not assign a supposed value to a person.

The connection between marketing and acquisition and conversational agents depends on this actionable handover. Agreeing on a modest, well-tested schema is a practical place to start.

Let's define the few pieces of information your team needs to handle each enquiry.