Lead qualification with Jev: fit, buying stage and timeline

What this check scores

The example above is an inbound message from a 40-person logistics company. They want to classify inbound shipping emails, handle about 20k a month, need something live before Q1, and ask for team pricing and a call this week. Jev answers three questions about it in one call:

This lead is a useful test because it sits between two stages. It says the company is "evaluating tools", which is what evaluating describes, but it also asks for pricing and a call, which is the definition of ready_to_buy. A single label would hide that. The probability split across buying_stage shows it, and the confidence tells you whether to hand the lead straight to a rep.

The message is short and there are three questions, so this is a standard evaluation: 1 credit on JevStation. The questions run in parallel against the same state, so asking about the timeline on top of fit and stage adds almost no latency.

How to write qualification questions that work

Describe your ideal customer in the fit question. The example asks how well the lead fits "a B2B API product", which is generic on purpose. Replace it with what you actually sell and who buys it, and give each Score level a line of criteria. Jev reads instructions literally, so the rubric is where your definition of a good lead lives.

Keep stages mutually exclusive. Each option in the example has a one-line definition: researching is "just learning about the problem", ready_to_buy is "asking for pricing, a contract or a call". If two stages can both be fully right, confidence drops for reasons that have nothing to do with the lead. Write the boundary into the criteria.

Give the timeline Noul criteria. The preset asks only "Does the lead mention a concrete timeline?". "Before Q1" and "this week" are concrete; "soon" is not. Say so:

{
  "type": "noul",
  "instructions": "Does the lead mention a concrete timeline?",
  "criteria": {
    "true": "The lead names a date, a quarter, a week or a deadline for going live or deciding.",
    "false": "The lead gives no time frame, or only a vague one such as soon or eventually."
  }
}

Leave the numbers to code. "40-person" and "20k emails a month" are exactly the details a rep cares about, but TypeSafe documents counting and numeric comparison as weaknesses. If company size or volume decides the tier, enrich the lead, compare in code, and pass the result into the state.

For more on phrasing, see Designing Questions Jev Can Answer.

From one lead to a pipeline step

1. Backfill with leads you already closed. 50 to 200 past leads marked won, lost or never qualified are enough to see where Jev’s answers separate them. The batch page runs one question set over every row of a CSV, so you do not need code for this step.
2. Pick bands, not one threshold. Send leads with a high fit score and a confident ready_to_buy to a rep now, put confident evaluating leads into a follow-up sequence, and review the middle band by hand.
3. Combine answers in one rule. A lead that is ready_to_buy with a high has_timeline probability is the one to call today. Because all three answers come from the same call, that rule is one line in your CRM integration.
4. Log the answers. Store the probabilities and the model field every response returns. Re-check your thresholds when the model changes, and treat reps’ corrections as your next batch of labelled data.
5. Call it from your form handler. The same evaluation runs through the System One API with an API key and charges the same credits as the playground.

Check it before you trust it

• Measure against outcomes, not the example. A threshold that looks right on one message is not your threshold. Use leads whose result you know.
• Watch the middle band. If most leads land between your cut-offs, the stage definitions overlap. Sharpen the criteria before adding more questions.
• Use the Score as a threshold check. TypeSafe describes Score levels as weakly calibrated as magnitudes, so compare fit against a cut-off rather than reading a value between two levels as a precise grade.
• Test separately if leads are not in English. English is where Jev is strongest; other languages are handled, but not equally well.

Where Jev is the wrong tool

• You need the reply written. Jev returns typed answers only. Pair it with a generative model or a rep for the follow-up.
• Qualification rests on numbers or dates. Headcount bands, contract renewal dates and budget thresholds belong in code.
• Your ideal customer changes every month. Thresholds tuned on one definition drift when the rubric does. Re-run your labelled set after each change.
• You need an explanation for every score. Jev gives probabilities, not reasons. If a sales manager must see why a lead was deprioritised, log the inputs and keep a human in the loop.

If you are weighing Jev against a general LLM or a classifier you train yourself, Jev vs LLM classification lays out when each one fits. For routing the support messages that arrive after the sale, see support ticket triage.