Why in an RPA stack

Most of the branch points in a bot are closed questions:

• routing work items (invoice / credit note / reminder / junk)

• classifying failures (business exception vs. application exception)

• guardrails before a side effect (is it safe to auto-reply to this email?)

• deciding when to hand off to a human, with the confidence as the gate

RPA.Decision turns each of these into one keyword, records every call in an audit log, and keeps the backend swappable: hosted Jev today, local Laya when data must not leave the machine, a replayed log in CI.

Install

pip install rpaframework-decision          # hosted backends (Jev)
pip install rpaframework-decision[laya]    # + local open-weight Laya

Robot Framework

*** Settings ***
Library    RPA.Robocorp.Vault
Library    RPA.Decision    threshold=0.85    audit_log=output/decisions.jsonl

*** Tasks ***
Route a document
    ${secrets}=    Get Secret    TypeSafe
    Authorize To Jev    api_key=${secrets}[api_key]

    ${result}=    Decide    ${document_text}
    ...    question=What kind of document is this?
    ...    options=invoice=A request for payment,credit_note=A credit against an invoice,other
    Log    ${result.choice} with confidence ${result.confidence}

    ${safe}=    Check    ${email_body}
    ...    question=Is this a routine request that a template answers fully?
    ...    threshold=0.9
    IF    ${safe}    Send Template Reply

    TRY
        ${team}=    Decide Or Escalate    ${ticket}
        ...    question=Which team owns this ticket?
        ...    options=billing,technical,sales
    EXCEPT    LowConfidenceError*    type=glob
        Escalate To Human    ${ticket}
    END

Python

from RPA.Decision import Decision, LowConfidenceError

decide = Decision(backend="jev", api_key="...", threshold=0.85)
result = decide.decide(text, "What kind of document is this?", ["invoice", "other"])
print(result.choice, result.confidence, result.probabilities)

local = Decision(backend="laya", device="cpu")   # nothing leaves the machine
if local.check(email, "Does the sender ask for a refund?"):
    ...

Keywords

What a decision model is not good at

Decision models are strong on recognition and weak on arithmetic and multi-step reasoning. Compute totals in Python; ask the model whether a document is an invoice, not whether its total is correct. Measure accuracy on your own data from the audit log before trusting a threshold.