Decisions

Decision engines

Decisions.
Answers with odds attached.

A decision engine reads messy input and returns a typed answer with a probability. See where that works, where it does not, and when to hand a case to a person.

Sign in to explore the demosHow it works

01 — What it is

One question in, one typed answer out.

You write the question in plain language and hand the engine some text. It answers in one of three shapes, and every answer carries the engine's own probabilities, so you can see how sure it is.

  • Yes or no

    A probability

    How likely is it that a statement is true of this text?

    Illustrative

    Did the caller confirm the policyholder's identity?

    Yes, 0.93

  • Choose one

    A choice among options

    Which of a fixed set of options fits, and how the odds split between them.

    Illustrative

    Which team should handle this claim?

    Glass 0.71 · Theft 0.19 · Other 0.10

  • Score a level

    A score on a rubric

    Which level of a written rubric the text reaches, from lowest to highest.

    Illustrative

    How well did the agent handle the complaint?

    Level 4 of 5 · 0.62

02 — Where they are used

Six decisions, each with a demo you can open.

Every demo is a set of questions run over a set of cases with known answers, so you can see how often the engine was right and how sure it said it was.

  • Live calls

    Call quality and compliance checklists

    The decision
    Has the agent covered everything the script requires, and is the call being handled well, while it is still going on?
    The demo shows
    Coverage and conduct scores re-evaluated as a call transcript grows, so a supervisor can step in before the call ends.

  • Claims

    Claims triage and numeric checks

    The decision
    Does this claim fall inside the policy: the right period, the notification deadline, the cover limits?
    The demo shows
    Threshold and date checks asked several ways, and per-rule verdicts built from the engine's answers.

  • Languages

    Multilingual claims

    The decision
    Can the same English rules be applied to claim files written in Portuguese?
    The demo shows
    The hosted and open-source engines side by side on Portuguese files, answered against English rubrics.

  • Matching

    Fund-fit matching

    The decision
    Does this organisation meet a funding programme's eligibility, geography, award size and purpose?
    The demo shows
    Eligibility answers per pair, with the labels flagged as not yet reviewed by their author.

  • Routing

    Routing and guardrails

    The decision
    Where should this go, and is the engine sure enough to send it there without a person looking?
    The demo shows
    A small worked example with a review band: move it and watch which cases are decided and which are held for a person.

  • Engines

    Choosing an engine

    The decision
    Which engine should make this decision for us?
    The demo shows
    What each engine is, what its confidence number means, and which ones can run today.

03 — See the uncertainty

Decide the sure ones. Send the rest to a person.

A probability tells you more than a verdict. Pick a band around the middle: answers outside it are decided automatically, answers inside it go to a person.

Widen the band and more cases reach a person; narrow it and fewer do. The demos let you move the band on real cases and see what it would have got wrong.

Answers placed by probability of yesDecide: noSend to a personDecide: yes00.51P(yes)
  • Filled dot: decided automatically
  • Open dot: sent to a person
Answers placed by probability of yes
P(yes)Outcome
0.03Decided no
0.07Decided no
0.12Decided no
0.31Sent to a person
0.46Sent to a person
0.55Sent to a person
0.68Sent to a person
0.87Decided yes
0.93Decided yes
0.97Decided yes

04 — Three engines

Same questions, three engines.

Decisions runs the same question sets through different engines so the difference is visible.

  • Hosted by TypeSafe

    Jev

    Available

    The model the demos were built on. The demos show its precomputed results on each demo's cases.

  • Open source

    Laya

    Precomputed results

    Shown on the same cases with precomputed results. On these cases it does worse than Jev, often close to chance on questions with several options. The question sets were written for Jev, and Laya's own documentation describes its checkpoints as a base to fine-tune.

  • Preview only

    OpenAI Decisions

    Preview, not yet available

    Announced on 29 September 2026 in limited preview, with no public contract. It is listed so it is not missed, but it cannot be run here.

For Patchworks staff

Sign in with your @patchworks.ai account to open the demos. Accounts are for Patchworks staff.

Sign inCreate an account