← BlogPerspective

Jev — Thousands of GTM decisions a day, automated for pennies

Jev is a new model built for yes/no, pick-one and rating decisions — not conversation or research. It makes a whole class of GTM and post-sales tasks affordable for the first time.

Nicholas Sainsbury

Co-Founder, Atlas

7 min read
Contents

Every week the X-sphere loses its collective mind over a new AI release. Jev is the latest to join the hype train. This time, though, it may matter more for GTM and post-sales work than it first appears.

Sales, customer success and operations workflows are full of small decisions, many of them made without noticing. Does the sentiment in these emails suggest this account will renew? How urgent is the request the client just made? Which products and features need configuring before go-live?

We make thousands of these calls every day, and with more time and headspace we would make thousands more. Today they get made in one of three flawed ways:

  • People — slow and inconsistent.
  • Rules — brittle keyword matches.
  • Large language models — seconds per call, a real cost, and prone to inconsistency.

If you asked Claude to make every one of those micro-decisions, you would go out of business very quickly.

Jev is a model released by a new company called TypeSafe AI. They describe it as a ‘System 1’ model because, like Daniel Kahneman popularised, it works very like our System 1 thinking. Fast, automatic, and unconscious processing mode that handles rapid judgments, instincts, and daily routines without mental effort.

Why should I care?

Jev makes previously unviable use cases viable. The cost of a single decision, or classification, has dropped massively.

Take a real example. Each month your implementation and CS teams generate 10,000 hours of customer call transcripts, and you want to QA every call against five questions:

  1. 01Were the success criteria agreed? (yes/no)
  2. 02Did the customer raise any implementation challenges? (yes/no)
  3. 03Was a go-live date committed to? (yes/no)
  4. 04Which of the following blockers was the primary challenge? (pick one)
  5. 05How clearly were next steps summarised on the call? (rate it)
ModelCostTime
Jev~$6Minutes
Sonnet 5$170+Hours
Five QA questions asked of 10,000 hours of implementation and CS call transcripts in a month.

At that price you stop sampling a handful of calls and start checking all of them.

How does it work?

You give Jev one item and one narrow question with a fixed set of answers. It returns a probability for each answer rather than free text, and you decide what happens at each confidence level.

Every small decision fits one of three question shapes. Jev answers in under a second and returns how sure it is, so you choose what happens at each confidence level.

Ten places Jev fits in post-sales and GTM

Customer success

01

Decomposed health signals

Don’t ask “is this account at risk?” Ask five narrow questions of every call, email and ticket: Did the champion say they are leaving? Was a competitor named? Is there a billing dispute? Did usage come up as a problem? Is an escalation open? Then learn the weights from your own renewal outcomes.

02

Handoff completeness

After the sales-to-CS handoff call, run a checklist as yes/no questions against the transcript: success criteria agreed, technical owner named, go-live date discussed, commercial promises recorded. Flag the gaps before kickoff, not at the QBR.

03

Every call tagged, not a sample

Tag every Gong call for competitor mentions, pricing objections, feature requests and expansion interest. At this price you stop sampling and start counting.

Support

04

Triage with an honest middle

Classify product area, severity and churn-risk language on every inbound ticket. Use three bands: high confidence routes automatically, low confidence is rejected, and the middle goes to a person.

05

Duplicate and linked-issue detection

“Is ticket A the same problem as Jira issue B?” is a choice question. Run it across the open queue every night and link what matches.

Sales and RevOps

06

CRM hygiene at full coverage

For every open deal, every night: does the next step match the stage? Does the last email support the close date? Is the listed champion still on the thread? Flag the contradictions to the owner. This is the review a RevOps lead wants to do and never has time for.

07

Inbound routing

Segment, intent and product line for every form fill and reply, in under a second, before the lead goes cold.

Ops and governance

08

Guardrails on AI drafts

Before an AI-written customer email goes out, ask Jev: does this commit to a date? Offer a discount? Promise a feature? Any “yes” goes to a person. The model that writes should not be the only model that checks.

09

Policy gates

Is this refund request inside policy? Does this discount need approval? Let Jev read the request, and let code do the arithmetic and dates — Jev is weak at both.

10

Context gating for your other AI

Most AI quality problems in GTM are input problems. A cheap relevance gate in front of the expensive model keeps noise out of summaries, briefs and account plans. It is also a privacy control: what never enters context cannot leak into an answer.

Getting started with Jev

  • Ask atomic questions. Narrow questions beat broad ones by a wide margin.
  • One item per call. Don’t let items influence each other.
  • Use three bands, not two. Confirm, reject, or send to a person.
  • Shadow first. Run Jev beside your current process for a while, compare, then automate.
  • Calibrate on your labels. TypeSafe says the probabilities are calibrated, but no public calibration curve exists yet. Check on your own data.
  • Keep receipts. Store the input hash, model, score and policy version for each decision. When someone asks “why was this excluded?”, you have an answer.
  • Keep code for maths, dates and counting, and a System 2 model for writing and multi-step reasoning.

Written by

Nicholas SainsburyCo-Founder, Atlas

Keep learning