Product

Governance that runs on every inference.

Every AI decision is judged against the policy that governs the workflow it belongs to. Setup, live operation, and the human checkpoints in between. What an assessor gets is a verdict and the control behind it. Not a log to interpret.

Posture

Setup & configuration governance

Governance starts at the workflow, before the first inference. The workflow is declared once. The agents, the tasks, the tools those tasks need, the policy, the risk tier, and an owner. Configuration is judged against that.

Which tools an agent may use, which model version it carries, what instructions it was given, and where data may leave. Each check binds to the same control that governs live decisions, so a setup failure and a live failure carry the same control. An assessor is never reconciling two systems.

Noncompliant setup is blocked before it ships. Drift detection compares what was approved against what is actually running and raises a finding when they diverge. They do diverge.

Configuration checkAgentStatus
Permitted tools within the approved list intake-agent Pass
Model version pinned credit-assist Pass
Instructions match the approved version claims-triage Drift
Outbound destinations declared care-navigation Pass
Secrets referenced by name only intake-agent Fail

Evaluation

Continuous evaluation

Policies bind to workflows and are evaluated on every decision. Each verdict is written against the active policy set and does not change once written.

Violations surface as they happen with the decision context attached, so nobody goes back for it later.

app.evidentai · /overview

Overview dashboard: compliance score, policies, controls, agents, alerts.

Enforcement

Runtime enforcement

EvidentAI blocks a risky action or holds it for human sign-off before the action completes.

Every enforcement decision carries the control that fired. An examiner can see what was stopped and why.

app.evidentai · /connections

Runtime connections: native control-plane connectors that feed AI Posture.

Security

Vulnerability monitoring

An AI vulnerability catalog ships with the product. Occurrences are drawn from live findings and ranked by severity. Each one maps to , so security teams read AI risk in a language they already use.

Coverage views show which defenses are registered and where the gaps are. Same controls, same verdicts as everything else.

ATLAS techniqueDefenseStatus
LLM Prompt Injection Input integrity verification Covered
LLM Jailbreak Policy gate before completion Covered
LLM Data Leakage Output PII scan Covered
LLM Meta Prompt Extraction Response filtering Partial
ML Supply Chain Compromise No defense registered Gap

Human oversight

Human-in-the-loop enforcement

When policy requires a person, the action stops. It lands in the sign-off queue with the decision context already attached. The reviewer approves or sends it back, and what they decide becomes part of the verdict.

Oversight is a governed step here. No screenshots in shared drives. Every review is attributable.

Held actionWorkflow · policyStatus
Credit line increase credit-decisioning Pending review
Claim payout above threshold claims-processing Approved
Rate change communication customer-comms Escalated
Prior authorization denial draft care-navigation Rejected

Workflows

Workflow-based governance

Governance attaches to the workflow. Register one and EvidentAI governs everything inside it: which agents acted, what policy applied, where a human signed off.

The workflow is shown as it actually executes, which is usually not the diagram.

Decision · credit-decisioningWhat happenedStatus
Agent actionintake-agent Application summary drafted Recorded
Tool callretrieval Lending policy and applicant file, at version Recorded
Policy gateverdict Obligations evaluated before completion Allow
Human checkpointsign-off Reviewer approved, name and time attached Approved
Evidencerecord Decision sealed, reconstructable on demand Sealed

Evidence

Evidence packages & framework mapping

Verdicts and sign-offs compile into packages built for examiner review. Controls map to the frameworks assessors already use, and SR 11-7 among them.

PII shielding and DSAR handling run on those same controls.

Package · Q2 examinationContentsStatus
Decision records Full context, replayable Sealed
Policy evaluations Every inference, verdict attached Sealed
Human sign-offs Reviewer, time, and context Sealed
Framework mapping NIST AI RMF · SR 11-7 Mapped

Next step

See it against real workflows.

Founding design partners are being recruited in banking, insurance, and healthcare. Open the live demo, or book a call.

Controls in an AI context