openauditor Get the code

Where it earns its place

Any step where software asks "which one", "how much", or "yes or no" and needs the answer in milliseconds with a number it can act on.

Every use case below assumes you fine-tune on your own labeled rows. The public checkpoint proves the pipeline and the interface; your data makes it accurate on your domain. Unseen policy tasks are its weakest measured area out of the box.

1. Review every agent action

An agent in an ERP, CRM or ticketing system is about to write. Before the write, ask the auditor whether to allow it. The state is the agent's context and intended action; the question is a Noul. Run it on every step, not a sample, because it costs milliseconds and no output tokens. Gate on the probability, log the rest.

auditor.review(
    {"agent": "ap-clerk", "intent": "post_journal_entry",
     "entry": {"account": "6100", "amount": 48250.00, "memo": "Q3 consulting accrual"},
     "context": {"approver_present": false, "vendor_verified": true, "similar_entries_30d": 0}},
    {"allow": Noul("Should this action proceed without a human?",
                   {"true": "Post the entry", "false": "Hold for review"}),
     "reason": Choice("If held, why?", {"amount": "Unusual amount", "approval": "Missing approval",
                                        "vendor": "Vendor risk", "none": "No issue"})},
)

Train it on your own action logs labeled with outcomes: succeeded, reversed, flagged by a human. The repository includes a converter for agent transcripts and a synthetic agent-trace task family.

2. Ticket and case routing

Which team, which queue, which priority. This is the model's strongest measured area: 8-way intent routing at 97% on the pilot. Replace a prompt-and-parse chat call with one review that returns a distribution over your teams and a Score on your urgency rubric.

{"state": {"subject": "Invoice INV-2291 shows tax twice", "channel": "email", "plan": "enterprise"},
 "questions": {
   "team":     {"type": "choice", "instructions": "Which team owns this?", "criteria": {"billing": "...", "tax": "...", "technical": "..."}},
   "priority": {"type": "score",  "instructions": "How urgent?", "criteria": ["low", "normal", "high", "critical"]},
   "sla_risk": {"type": "noul",   "instructions": "Is there a risk of breaching the SLA?"}}}

3. Exception triage in finance operations

Accounts payable, expense reports, reconciliations. Most items are routine; the job is finding the few that are not. Ask a Noul per policy rule and a Score for severity, then route by threshold: auto-approve above 0.9, queue for review in the middle, escalate below 0.2. Because the probabilities are calibrated on your validation set, the thresholds mean what they say.

{"state": {"expense": {"amount": 412.80, "category": "meals", "attendees": 6, "receipt": true, "city": "Austin"},
           "policy": {"meal_per_person_max": 75, "receipt_required_over": 25}},
 "questions": {
   "compliant": {"type": "noul",  "instructions": "Does this expense comply with the policy?"},
   "severity":  {"type": "score", "instructions": "If not compliant, how serious?", "criteria": ["minor", "material", "escalate"]}}}

4. Approval gating in workflows

Purchase orders, discounts, refunds, credit limits. The workflow engine asks one question at each gate and branches on the answer. The native environments the model trains in are exactly this shape: an enumerated action set, a state that changes after each step.

5. Document and record classification

Contract type, invoice versus receipt, GL account suggestion, product category. Feed the extracted fields as the state and the taxonomy as Choice criteria. Descriptions in the criteria are read by the model, so a good taxonomy description is most of the work.

6. Scoring on your own rubric

Lead quality, churn risk, review sentiment, vendor risk, ticket satisfaction. A Score question with your levels returns a distribution and an expected index, which you can average, trend and threshold like any other metric.

7. Picking the next step for a generator

Pair it with a chat model. The chat model drafts candidates (three possible replies, three possible tool calls); the auditor scores them and picks one. The generator supplies creativity, the auditor supplies a calibrated judgment about which candidate to act on.

What to avoid