Build a Lead Router That Replies in Minutes

Inbound leads decay fast. The form fill that sits in a shared inbox until Tuesday is the one a competitor answers on Monday. AI lead routing closes that gap: every lead gets judged the moment it arrives and lands in front of the right person while the prospect still has the tab open. Jev, TypeSafe AI's decision model, does the judging part.

What to score and why it beats the points spreadsheet is the territory of AI lead scoring. This page wires the router.

The receipt

Romàn gave Jev 700 high-intent leads with their personalised outreach messages. In 40 seconds it predicted how each message would perform, attached a confidence score, and flagged leads that didn't match their message, for $0.09 total, as reported (build). That's outbound scoring, not inbound routing, but it's the same verdict machinery and it tells you per-lead cost rounds toward zero.

A Clay workflow does something similar for prospect lists, giving every row the same typed question (build). Numbers here are builder-reported; no official Jev benchmarks exist.

The recipe

1. Catch the lead at the source. Form submission, demo request, chatbot handoff, "contact sales" email. Send each to one webhook. If you use a no-code tool, the pattern in Jev automations covers this step without writing a service.

2. Enrich before judging. Add what you can look up cheaply: company size from the email domain, existing CRM record, open deals. Jev judges what you give it, so give it the rep's view of the lead.

3. Ask routing questions, not a score. Four questions tend to cover inbound:

   • "Is this a sales inquiry, a support request, a partnership pitch, a job application, or spam?"
   • "Does this message describe a specific need or project with a timeline? YES / NO / UNCLEAR"
   • "Which segment fits this company: self-serve, mid-market, enterprise, unclear?"
   • "Does the requester appear to be an existing customer? YES / NO / UNCLEAR"

   The first question does a lot of quiet work. Support requests and job applications in the sales queue are a surprising share of "leads".

4. Call Jev and route.

# pseudocode, not real API syntax
on lead(l):
    ctx     = pack(l.message, l.form_fields, enrich(l.email))
    kind    = jev.choose(KIND_Q,    ctx, ["sales", "support", "partnership", "job", "spam"])
    intent  = jev.choose(INTENT_Q,  ctx, ["YES", "NO", "UNCLEAR"])
    segment = jev.choose(SEGMENT_Q, ctx, ["self_serve", "mid_market", "enterprise", "unclear"])
    assign(l, route_table(kind, intent, segment))   # low confidence -> triage queue

Per ecosystem documentation, the model is typesafe-ai/jev via the Vercel AI Gateway, returning per-choice probabilities. Syntax at docs.typesafe.ai.

5. Alert the owner, instantly. High-confidence sales plus clear intent pings the right rep with the lead and the verdict table. Enterprise goes to the account team. Support goes to support. Anything low-confidence goes to a shared triage queue someone checks every hour.

"Replies in minutes" without a robot replying

The obvious objection: "Why not have AI send the reply too?" You can draft one, with a chat model, since Jev doesn't generate text. But an email to a prospect can't be unsent, and a confidently wrong reply to an enterprise buyer is expensive. So the rule here: no outbound message fires on a lone verdict.

What does fire automatically is a generic acknowledgment that every submitter gets regardless of routing ("Thanks, a person will reply shortly"). The fast, specific reply comes from the rep who was pinged 30 seconds after the form fill, optionally starting from a draft they review. That's how you reply in minutes without letting a model freelance with your pipeline.

Inbound lead automation hygiene

Measure speed-to-lead before and after. It's the metric this build exists to move, and it's easy to track from CRM timestamps.

Check misroutes weekly. Ask reps to tag wrong assignments. Those tags are your tuning set, and they usually point at one ambiguous question.

Keep a human in the loop on disqualification. Marking a lead as spam is fine to automate for obvious junk, but "not a fit" verdicts on real people should be reviewable, since they're where good leads quietly die.

Log verdicts on the CRM record. When a rep asks "why did this come to me?", the answer should be one click away.

If your inbound looks more like a support inbox than a sales funnel, support triage with Jev is the sibling recipe.

Frequently asked questions

How fast can an AI lead router assign a lead?

The verdict itself is fast; one builder reports 700 leads scored in 40 seconds, as reported. In practice, speed-to-lead is limited by how quickly the pinged rep responds.

Should the router send replies to leads automatically?

Only a generic acknowledgment that everyone receives. Specific replies should come from a person, optionally from a drafted starting point, because outbound emails can't be taken back.

What's the difference between lead routing and lead scoring?

Scoring judges how good a lead is; routing decides where it goes right now. The strategy side of scoring lives in AI lead scoring.

Can I build this without code?

Yes, with a webhook step in your automation tool and branch logic on the verdicts. See Jev automations for the pattern.