Jev lead scoring: fit, intent and disqualifiers at submission time

The problem

Lead qualification is usually a nightly batch job or a rules engine nobody has updated in two years. Doing it with an LLM per lead is affordable but too slow to act on in-session.

How Jev handles it

Score fit and urgency against rubrics built from your actual sales criteria, plus Noul checks for disqualifiers. Fast enough to act on before the prospect leaves.

response = client.system_one(
    model="jev-latest",
    state=f"{form_submission}\n\nCOMPANY: {enrichment_data}",
    questions={
        "fit": Score(
            instructions="How well does this lead match our ideal customer?",
            criteria=["Poor", "Marginal", "Good", "Ideal"],
        ),
        "intent": Score(
            instructions="How far along is this buyer?",
            criteria=["Browsing", "Researching", "Evaluating", "Ready to buy"],
        ),
        "competitor": Noul(
            instructions="Is this a competitor rather than a prospect?",
            criteria=NoulCriteria(true="Works at a competing vendor",
                                  false="Genuine prospect"),
        ),
    },
)

a = response.answers
# fit and intent are 0..3 weighted means: 2 = Good / Evaluating, 3 = Ideal / Ready to buy
if round(a["fit"].score) >= 2 and round(a["intent"].score) >= 2 and a["competitor"].noul < 0.2:
    page_sales_immediately()

Writing a rubric for Jev lead scoring

Jev lead scoring works best with fit and intent as separate rubrics. Fit is about the account — would you want this company — and intent is about timing — are they buying now. A perfect-fit browser and a poor-fit buyer need different follow-ups, and one blended score cannot tell them apart. Write each level the way your sales team defines it in a deal review, with the markers they look for: a named budget, a deadline, a contract that is ending.

Read the distribution, not only the mean. The measured lead scored 2.44 on fit, which rounds to Good, but its probabilities were 0.51 Ideal and 0.44 Good — a near tie between the top two bands. That is why the sample routes on "Good or better" for both rubrics: a rule that required fit to round to Ideal would have sent this lead to nurture.

What to put in the state

A Jev lead scoring call reads the form and the enrichment together, as the sample does: what the person wrote, then a labelled company block. Turn numbers into bands before they reach the state — "800 employees" becomes "SIZE: 500–1,000" — because TypeSafe is explicit that Jev is not a calculator, and a rubric keyed on company size should not depend on arithmetic.

Dates need the same treatment. "Our contract ends in November" is useful as text, but if the rule is "renewal within 90 days", compute that in code and state the result plainly.

What one decision costs

Lead volumes are small, so the bill is rarely what decides Jev lead scoring. First line: the measured lead, form plus enrichment, three questions. Second: 5,000 inbound leads a month at that size. Third: re-scoring a 100,000-lead database once after a rubric change.

# the measured lead, three questions
472 tokens × $0.042 / 1M = $0.0000198

# 5,000 inbound leads a month at that size
5,000 × 472 = 2,360,000 tokens × $0.042 / 1M = $0.0991

# re-scoring a 100,000-lead database once
100,000 × 472 = 47,200,000 tokens × $0.042 / 1M = $1.98

At these volumes the argument is speed and a rubric you can edit, not price: rewording one level and re-scoring every lead you hold costs about two dollars.

When not to use Jev for this

• You have years of closed-won and closed-lost outcomes. A model trained on your own conversion history learns what actually predicts revenue; a rubric only encodes what the team believes does.
• Qualification rests on numeric thresholds — revenue bands, seat counts, budget minimums. Compute those in code rather than asking Jev to do the arithmetic.
• Lead data may not go to a third party. Calls go to a hosted API; TypeSafe does not train on customer requests, but zero data retention is only offered to enterprise customers.
• The next step is a personalised email. Jev scores and routes; the writing belongs to an LLM or a person.

Where it fits in your stack

A Jev lead scoring call runs on the form-submit webhook, after enrichment and before the CRM write. Scores land in CRM fields; a hot result — good fit, evaluating or ready, not a competitor — pages the owner in chat while the prospect is still on the site, and everything else enters a nurture sequence keyed on intent.

An LLM can then draft the first reply with the scores as context. Every closed deal becomes a labelled example for checking whether the rubric's top bands actually convert, which is the moment to revisit the wording.

Notes from the field

• Write rubric levels in your sales team's own language — "ideal" should mean what they mean by it.
• The competitor Noul is the kind of cheap check nobody bothers building a model for.
• Score at submission time and route instantly; speed of first response dominates most other factors.

Common questions