What lead scoring is, and where points-based models stop
Lead scoring ranks prospects so sales spends time on the ones most likely to buy. A classic model adds points for fit, such as industry, company size or role, and for engagement, such as pricing page visits, demo requests or email clicks, and subtracts points for signals like a student email. CRMs such as HubSpot and Salesforce offer both manual rule-based scoring and predictive scoring trained on your historical deals.
Both approaches work on structured fields. Neither reads the sentence “we need to route 4,000 tickets a month before Q3”, which states a use case, a volume and a deadline, three things a good salesperson looks for first. That text sits in a form field or the first email, and it is usually the strongest signal you have for a new lead with little engagement history.
Turn free text into a readiness band
Instead of asking a model for a score from 0 to 100, which invites invented precision, ask it to place the text in one of a few bands defined in writing. Four bands are enough for most teams: strong, when the lead names a use case, a volume and a timeline; developing, when the use case is clear but volume or timing is missing; early, for curiosity without a concrete need; and not a fit, for needs you do not serve.
The band definitions are the part to get right. Write them the way you would brief a new sales development rep, including what must not count: never infer budget, seniority or anything about the person from a name, email domain or writing style. The request below is ready to paste into the playground.
{
"text": "We handle about 4,000 support emails a month in Zendesk and want to auto-route them by team before Q3. Two engineers available.",
"instructions": "Classify buying readiness from what the lead wrote. strong: named use case, volume and timeline. developing: use case but no volume or timeline. early: curiosity only. not a fit: needs we do not serve. Never infer budget, age, gender or location.",
"labels": ["strong", "developing", "early", "not a fit"],
"threshold": 0.75
}Score a lead in the playground · Request and response contract
Combine the band with the score you already have
Keep your existing fit and engagement points and add the band as one more input. A simple scheme maps strong to 30 points, developing to 15, early to 0 and not a fit to minus 30, then lets the CRM add them to the total. Start with numbers like these, then adjust them with the calibration step below rather than by intuition.
Store the band, the confidence value when the model supplies one, and the request identifier on the contact record. When confidence is under your threshold, do not add points at all; mark the lead for a human look instead. A score that silently absorbs a guess is worse than a score that says it does not know.
Three leads, three bands
“We run a 40-person support team on Freshdesk, about 6,000 tickets a month, and want suggested queues live by the end of next month.” This is strong: a use case, a volume and a date. Add the points and let a rep reply the same day.
“Looking for a way to categorise customer feedback automatically.” This is developing: the need is real but there is no volume, tool or timeline. A useful reply asks one question, such as how many responses arrive each month, rather than booking a demo.
“Just exploring what AI can do for our business.” This is early. It deserves a helpful resource, not a sales call. A message asking for image generation or a chatbot builder would be not a fit for a classification API, and the kind response is to say so and point elsewhere.
Check the bands against revenue before you rely on them
Pull 200 to 500 historical leads with their original form text and outcome: won, lost or never engaged. Classify the text with your rule and compare conversion rates by band. If strong leads do not convert noticeably better than developing ones, the definitions are not separating what matters, and the fix is in the wording, not in more points.
Look at the errors in both directions. A strong band on leads that never replied usually means the rule rewards detail rather than intent. A not-a-fit band on deals you won is the expensive mistake, so review those cases one by one. Re-run the check whenever you change the product, the target market or the band wording.
Keep people in the loop and the data clean
The band changes priority, not access: every lead still gets a response, and nobody should be rejected automatically on the strength of one classification. Let reps overrule the band with one click and record the reason, because those corrections are the best material for improving the rule.
Send only the text needed for the decision. Remove phone numbers, addresses and anything you would not paste into a support ticket. If you score existing leads in bulk, classify a file in batch rather than looping over single requests, and write results back to the CRM through your own integration so the scoring logic stays auditable.
Human review for uncertain results · Classify a lead list in batch · Route inbound leads to a queue