jevusecases.work

AI decisions your board can sign off.

Jev makes short, typed judgments on routine operational work in Microsoft workflows: route this, score that, flag what doesn’t fit. Your policy stays in code. Uncertain or high-stakes cases go to a person.

Explore the use cases See it work

Independent guide. Not affiliated with TypeSafe or Microsoft.

Incoming work passes through Jev and your policy rules. Each item leaves as a raised ticket, a draft for review, or a case for a person to decide. POLICY IN CODE Jev +RULES TICKET RAISED DRAFT FOR REVIEW A PERSON DECIDES
ChoicePicks one option from a closed set.
ScoreRates how good, or how severe, on a scale.
NoulSays when a condition is present, or that none fit.
RulesYour policy, in code, decides what happens next.

Read every case, not a sample.

Many operational judgments are too slow or too expensive to make on every item, so teams sample. Reading every support request for routing, every completed agent run for failures, or every invoice for missing explanations becomes affordable when the judgment is small and bounded.

  • Operational decisions with observable outcomes
  • Stable categories your teams already use
  • First actions that can be reversed
Illustration: today a small sample of items is checked; with Jev every item is read and doubtful ones go to a person. TODAY WITH JEV A SAMPLE GETS CHECKED EVERY ITEM IS READ DOUBTFUL ONES GO TO A PERSON
Illustration, not measured data.

Confidence is not permission.

Jev returns an answer with a probability. Your code decides what that answer is allowed to trigger. Cases below your threshold go to review, high-stakes actions stay with a person whatever the score, and every final action records the policy that produced it.

What to measure first
Illustration: cases plotted by confidence. Cases above your threshold may be acted on by policy; cases below it go to a person. YOUR THRESHOLD 0 CONFIDENCE 100% BELOW: A PERSON REVIEWS ABOVE: POLICY MAY ACT
Illustration of the policy gate, not model output.

Spend follows difficulty.

A small classification step filters the workload, so expensive reasoning and human attention go to the cases that need them. Budget for the whole path: orchestration, retrieval, licences, review, mistakes and support, not only model tokens.

Read the board brief
Illustration: every item passes through rules, Jev judges the rest, a frontier model drafts for fewer, and a person handles the smallest share. RULES · EVERY ITEM JEV · BOUNDED JUDGMENTS FRONTIER MODEL · DRAFTS A PERSON · DECIDES
Illustrative shape, not measured volumes.

Watch it sort a service-desk inbox.

An imaginary company mailbox. Pick a message and follow it through rules, Jev, a frontier model and a person. Some paths raise a ServiceNow ticket, some draft a reply for approval, some go straight to a human queue.

Shared mailbox · service desk Synthetic demo Proposed design
Inbox · 8 messages
    RulesPolicy in code
    JevChoice · Score · Noul
    Frontier LLMDraft only
    HumanAccountable action
    Inputs

    Select a message.

    Jev + rules

    Judgments appear here.

    Outputs

    Tickets, drafts, queues.

    Synthetic data, client-side only. Full rule table with editable toggles: service-desk email demo.

    One pilot. Twelve weeks. A stop rule.

    Fund one bounded pilot with a named process owner, baseline measures and a stop rule. Require evidence before expanding authority.

    Weeks 1–2

    Confirm the case

    Buyer, data permissions, baseline and labels.

    Weeks 3–4

    Build

    The adapter and a Power Automate demonstration.

    Weeks 5–6

    Shadow

    Evaluate alongside the current process; tune on development data.

    Weeks 7–8

    Validate

    Held-out results and a failure-recovery demonstration.

    Weeks 9–12

    Decide

    Limited routing if approved, and an evidence report your team agrees.

    Before the pilot

    Decision discovery

    A shortlist of decisions, owners and allowed actions, with a pilot recommendation or a clear reason not to proceed.

    The pilot

    Bounded pilot

    One process, one data source, one decision policy and an agreed evaluation set, with acceptance criteria set in advance.

    After acceptance

    Managed service

    Monitoring, review quality, controlled releases and monthly outcome reporting.

    Timings depend on customer access and procurement. Pricing is quoted per engagement. Services in full.

    Where it works first.

    Repetitive decisions, stable categories, reversible first actions, lawful access to representative data, and a sponsor who owns the queue. Each page runs an interactive demo on imaginary data.

    All twenty use cases Where it sits in Microsoft

    Not a first pilot.

    Autonomous payment approval, employment selection, diagnosis, or irreversible account sanctions. These stay out of scope until a bounded pilot has earned the evidence.