What the model is given
The input is the conversation transcript, with speaker labels — Customer and Agent — rather than chat roles. Labels matter: a model asked to judge 'the customer's budget' needs to know which lines are the customer's.
It is the whole thread, not the last message. Authority is often established in message four and never mentioned again; a scoring pass that only reads recent messages misses most of what it is looking for.
What comes back
Four booleans and a sentence. Budget, Authority, Need and Timeline, each true or false, plus a one-line explanation of the judgement. Nothing more elaborate, because a longer output is a longer thing to be wrong about.
Asking for a bare number instead would be worse. A model asked to 'score this lead out of 100' will produce a confident 73 with no way for anyone to check it. Four discrete judgements can each be inspected and disagreed with.
Where the number comes from
The score is arithmetic, not a second model call. Each satisfied criterion is worth 25 points, so the possible scores are 0, 25, 50, 75 and 100 — and the label follows from the score: COLD up to 30, WARM to 60, HOT to 80, QUALIFIED above that.
This is deliberately boring. The judgement is the hard part and belongs to the model; turning four booleans into a number is not, and making that step deterministic means the same conversation always scores the same way.
What gets written back
The score and its label land on the lead. So does the authority judgement, which maps onto a decision-maker flag the sales team already reads.
The other three are returned to whoever asked but not forced into fields that were not built for them. Budget and timeline on a lead record are free-text columns meant for '₹5,00,000' and 'end of Q3'; writing 'true' into them would corrupt data a human relies on. Where a system does not have somewhere honest to put a value, the right move is to not put it anywhere.
What it is not
It is not a prediction of whether the deal will close, and it should not be sold as one. It is a structured reading of what the customer already said. A lead can score 100 and still go quiet, and no scoring model will tell you otherwise.
Treated as a way to order today's follow-ups, it earns its place. Treated as a forecast, it will disappoint you.