Put recurring governance work on a schedule
Scout and the platform agents quantify risks, assess controls and run your checks on the cadence you choose. Every result keeps the sources it read, and the decisions that belong on the record stay with your team.
One sample week. On Monday at 09:00 a weekly job runs risk quantification, control assessments and agent tests. One of the agent tests fails. Metric tests check production every morning at 06:00; on Tuesday a data drift test fails and the control owner is notified, and on Thursday a data test fails. Over the week the Risk Agent makes 36,000 estimates, the Control Assessment Agent assesses 438 controls, 3 checks fail, and people take 3 decisions.
Governance work repeats every quarter and goes stale in between
Every quarter someone re-reads the evidence, re-checks each control and re-estimates each risk by hand.
The record is accurate on the day it is finished, but models, vendors and usage change every week, so the next review starts from an old picture.
Each agent works from a clear brief
Every worker on the platform reads a defined input, produces a defined output and runs on a defined trigger. That keeps each result predictable, and every output names the records it came from.
- Risk AgentUsage data, controls and evidenceAn annual loss estimate in euros, with its reasoningOn a schedule or on demand
- Control Assessment AgentThe evidence attached to each controlA readiness score with cited gaps and recommendationsOn a schedule, per framework
- Agent testsThe connected sources you nameA verdict of pass, fail or error, with a run reportOn a schedule or Run now
- Evidence processingEach document you uploadA proposed name, summary and control mappingsWhen a document arrives
- ScoutYour platform records and connected sourcesAnswers that link the records they usedWhen someone asks
Describe a check in plain language and get a verdict
Agent tests check what a threshold cannot, such as whether a process was followed across two systems. Write the objective and the pass criteria the way you would brief an auditor and point the agent at the sources. It inspects them, applies your criteria and returns pass, fail or error, with a report that shows the evidence behind the verdict.
Run a test once or put it on the weekly job. The inputs are frozen for each run, so every verdict matches the criteria it was judged against.
Edit Agent Test
Name*
Claim denials are reviewed by a handler before the customer is told
Test Objective*
Check that the Claims triage assistant never tells a customer a claim is denied or reduced before a claims handler has reviewed it, as required by MCF-179. Look at every conversation in the last 30 days where the assistant recommended denying or reducing a claim, and match it to the handler review in the claims ledger.
Pass Criteria*
PASS if every denial or reduction recommendation has a review by a named claims handler, recorded before the customer received the decision. FAIL if a customer heard first or a review is missing. ERROR if a source cannot be read.
Owner*
Head of claims operations
Sources
Langfuse · Claims triage assistantSnowflake · Claims ledger
Mapped Controls
MCF-179
Agent Test Report
Test
Claim denials are reviewed by a handler before the customer is told (AT-7), aligned to MCF-179.
Evidence inspected
- Langfuse traces from 21 August to 21 September: 4,812 conversations, 316 of them with a denial or reduction recommendation.
- Claims ledger review records for the same 316 claims, with reviewer and timestamp.
Findings
Every one of the 316 recommendations has a handler review in the ledger. In three conversations on the new web chat channel, the decision reached the customer 2 to 11 minutes before the review was recorded.
Adjudication
- A named handler reviewed every recommendation: Met, 316 of 316.
- The review came before the customer heard: Not met, 3 of 316 sent first.
Verdict
Failed. Three customers learned of a denial before a claims handler had reviewed it.
Group related checks into a job and pick the cadence
A job holds the work that belongs together, from metric tests and agent tests to control assessments and risk quantifications. A framework or an AI system’s risks expand into a step for each control or risk. Pick daily, weekly, monthly or quarterly, or press Run now, and every run keeps its date and each step its result.
In this job
- EU AI Act · Claims triage assistantControl assessments · 56 controls
- Claims triage assistantRisk quantification · 30 threats
- Damage photo classifierRisk quantification · 30 threats
- Agent tests4 tests, 1 disabled
Cadence
Monday 09:00
Run nowNext run · Mon 28 Sep, 09:00
Last run
Run 21 · Mon 21 Sep, 09:00
- 119 steps completed
- 1 skipped, its test is disabled
- 1 agent test failed its check · AT-7
People decide what goes on the record
Evidence processing proposes a name, a summary and control mappings for every document you upload, and a person creates the evidence. The Control Assessment Agent scores readiness and cites the gaps it found, while the control’s status stays with its owner. Scheduled risk estimates update on their own, and each keeps its method and reasoning for review.
When a metric test fails, the control it proves reopens and its owner is told.
Evidence processing · proposals for the uploaded documents
- bias-audit-q3.pdfBias audit, third quarterBias and fairness testing
- model-card-v4.mdModel card, version 4Technical documentation
- vendor-soc2-2026.pdfVendor SOC 2 report 2026Third-party assurance
- hitl-override-log.csvHuman override log, SeptemberHuman oversight
Control Assessment Agent · Human oversight
Readiness, with cited gaps
- No override log before July [1]
- Escalation owner not named for the new channel [2]
5 recommendations, each citing the evidence it read
Ask Scout about any control, risk or source
Scout sits in a drawer on every screen. It reads your controls, evidence, frameworks, risks, policies and tests, and it researches your connected sources in read-only mode. Each answer links the records it used, so you can check it in one click.
You • 21.09.2026 10:12
Which controls cover human oversight for the Claims triage assistant, and what is missing?
Scout • 21.09.2026 10:12
Two controls in this AI system cover human oversight. MCF-178 has current evidence. MCF-179 was reopened at 09:14 today: agent test AT-7 found three claim denials that reached the customer before a handler had reviewed them.
MCF-178 Human Oversight Features and ToolsMCF-179 Human Oversight Deployer RequirementsAT-7 run report
Confluence · Risk policy v3GitHub · model-card.md
Session ID: 7c21f0b4...e91a6d03
Every agent run is metered by feature
AI usage is attributed to the feature that spent it, so you can see what Scout, the Risk Agent, control assessments, agent tests and evidence processing used this month. The platform warns you as the monthly quota approaches its limit.
64% of this month’s AI quota used
As of Mon 21 Sep, 09:00Warning at 80%
- Control assessments23%
- Risk Agent17%
- Scout12%
- Agent tests8%
- Evidence processing4%
Bring a check your team repeats every quarter
Tell us about one review your team repeats by hand. We show how it would run as an agent test or a scheduled job, and what its run report would look like.
Discovery, governance, risk and compliance
- Discovery Find and register every AI system, then send each one through a review.
- Governance Keep every AI system owned and governed from inception to retirement.
- Risk Put a price on every AI risk and see how it changes from week to week.
- Compliance Map one control to many frameworks and test it against production.