A chain you can try to break
Verification doesn't return a green check. It walks every row from an empty genesis hash and reports how far it got — and if a row was edited, which one.
The record proves its own integrity.
Use cases / GRC
Prove control to an assessor, not to a dashboard.
AI-native governance, risk and compliance for agents. Manage policies, assess risk, govern approvals and enforce controls in one platform — and connect the decision to deploy with the actions taken in production and the record your reviewer will ask for.
Independent runtime governance. Your policies. Your environment. Your evidence.
A control, in action — and the record it writesSample data · hashes computed here
Customer support agent
Issue a refund
support.issue_refund · original payment method
$12,400
Held before dispatch
Finance approval required.
The refund exceeds the $10,000 limit. The action is held for a named approver, and nothing is dispatched.
Policy appliedrefunds · rule R-3
Human reviewFinance approver · pending
Agent version assessedrefund-agent 3.2 · same version acting
DispatchNot dispatched
Record written to the ledgerVerify the chain
#1,285policy.decision · hold4bad…3e
Move AI reviews forward
Bring a controlled workflow to the table, not a policy document.
Keep accountability visible
Know which rule applied, who reviewed it, and what happened next.
Start reviews with the record
Give assessors evidence without reconstructing the story.
See the real product
Evidence, control mapping and export in the console. Watch the walkthrough
In the product
See the record an assessor reviews, in the console.
Evidence, control mapping and export in the product: evidence per framework, each item mapped to a control, an audit entry that carries its own hash and the one before it, and the export your reviewer takes away.
The record, step by step
Real product screens from a demo tenant with seeded activity. Nothing was collected, exported or changed while recording.
Recorded against a demo tenant with seeded activity. Real product screens, not a mock-up.
The record
Open the record. See the decision.
The refund above is now a ledger entry, a control mapping, and an export. Switch between them — then edit a row and watch verification recompute every SHA-256 and stop at the one that no longer reconciles.
- #1,281policy.decision · allowinvoice-agent → netsuite.post · ap-controls · args sha256:7c1e…prev genesis
- #1,282policy.change · approvedauthor j.okafor · approver m.reyes · refunds v13 → v14prev 9d32…e6
- #1,283assurance.verdict · passrefund-agent 3.2 · snapshot sha256:b02d… · bars preregisteredprev ee99…0b
- #1,284kill_switch.activatedops-oncall · scope: invoice-agent · reason: anomaly 0.91prev bc15…4a
- #1,285policy.decision · holdrefund-agent → stripe.refund · $12,400 · refunds R-3 · explain: OVER_LIMIT_NEEDS_APPROVAL · agent 3.2 b02d…prev 9633…0c
5 entries linked from genesis. Verify to walk the displayed sample.
Refund authorization — what governed it
Actionsupport.issue_refund · $12,400 DecisionHeld before dispatch Policy basisrefunds · rule R-3 Policy approved bym.reyes (author j.okafor) Human responsePending · finance approver Assessed versionrefund-agent 3.2 · asr-0417 · pass Version that actedrefund-agent 3.2 — match Ledger entries#1,285
Which control it supports
Internal controlFIN-REFUND-01 · refund authorization ISO/IEC 42001Human oversight of AI decisions partial NIST AI RMFMANAGE · approval workflow signal partial SOC 2Change management (author ≠ approver) live signal Remaining to attestApprover training record; threshold review cadence Control ownerFinance controls · attestation task open, due 31 Oct
bundle evidence-2026-Q4-refund-authorization.tar
range 2026-10-01 → 2026-10-05 entries 1,285
merkle_root 5e7c…b914 signer ed25519:gw-prod-02
chain genesis → #1,285 reconciled
controls FIN-REFUND-01 · ISO/IEC 42001 · NIST AI RMF · SOC 2
formats PDF · JSON · CSV · HTML
verify POST /api/v1/orgs/{org_id}/evidence/verify valid
An assessor checks the signatures against the deployment’s public keyring through GovernorAI’s verify endpoint. An offline verifier for decision evidence records, governorai-verify, is built but not yet released to assessors.
The AI GRC program
Run your AI GRC program in GovernorAI. Connect it to your enterprise.
Bring AI assessments, policies, approvals, control monitoring and evidence together in one platform — the AI GRC platform, not a replacement for the enterprise GRC suite that runs the rest of your control universe. Connect the two without separating policy decisions from the actions they govern.
- 1Assess risk
Classify the system, gate the version, set the bar before the run.
- 2Approve policy
Author, test, simulate, approve — with a named approver on the record.
- 3Enforce controls
Allow, hold or deny the action at the boundary where the agent acts.
- 4Monitor outcomes
Anomaly, risk score, kill switch, attestation campaigns with owners and due dates.
- 5Demonstrate compliance
Status per control, program workspaces, exports your reviewer verifies, signed where enabled for the deployment.
Keep your enterprise GRC
ServiceNow IRM, Archer, OneTrust or Vanta stay the system of record for the rest of your controls. GovernorAI is the system of record for AI agents — programs, framework packs, attestation, evidence — and exports into them, mapped to the controls they already track.
PDF · JSON · CSV · HTML → your enterprise GRC
Keep your policy authority
Native rules or OPA/Rego, including the OPA engine you already operate. Your repository, your review process, your approvers. GovernorAI applies those decisions to agent actions and records the result.
your OPA → AI action → ledger
Keep your platforms
One policy and one ledger across Bedrock, Azure, Google Cloud, Databricks, Kubernetes, ServiceNow Now Assist, Salesforce Agentforce and MCP. The platform running the agent doesn't grade its own homework.
one record · every runtime
Independence is the point. A reviewer trusts a record more when it wasn't produced by the system being reviewed — and when the platform that governs the agent is the one that enforced the control.
Assessor questions
What an assessor asks about an agent that took a consequential action
Five questions. Each is answered with a shipped surface, not a slide.
What did this agent actually do, and when?
Each control-plane audit entry is written with a sequence number and the SHA-256 of the entry before it. Verification walks the chain from genesis and stops at the first row it cannot reconcile — naming that row.
Evidence & Compliancedecision recordWho allowed it, and against which rule?
Gateway decisions carry the matched policies, an explain code, and — where escalated — the approver and decision latency. Gateway decision events record a digest of the call arguments; approval requests keep the arguments.
Policy & Governanceattributable decisionsWas the version that acted the version you assessed?
Assurance evaluates an immutable snapshot — model, prompt and tool manifests, artifact and runtime digests — against bars set before the run. A config change stales the verdict; the record shows whether the assessed version is the one that acted.
Assuranceversion matchWhich control does this evidence support?
Each control in a framework pack carries a status derived from live signals — audit volume, active policies, approval workflow, kill switch, RBAC — with the fields a person still attests to named and tracked to an owner and a due date.
Evidence & Compliancestatus per controlCan I take it away and verify it myself?
Framework exports across a date range, an AI-BOM with a content hash, and signed evidence bundles — Ed25519 with Merkle manifests, where a deployment enables signing — that verify against the deployment’s public keyring through GovernorAI’s verify endpoint.
Evidence & Compliancesigned bundles
Where it differs
Where GovernorAI differs
Four mechanisms you can see operating in the product — above, on this page.
The version that mattered
A runtime decision records, where it can tell, whether the acting agent snapshot is the one assurance assessed, so a reviewer can see whether the assessment covered today's action. It says nothing about whether that assessment passed, and the record does not carry the policy revision.
An assessment of another version isn't an assessment of this one.
Evidence minted at deployment, not reconstructed in November
The assurance gate is a pre-deployment decision surface with five distinct exit codes a deploy step can consume, shown today as a demo, not yet the automated stop in your pipeline. It writes the artifact you would otherwise rebuild eight months later from memory and Slack.
The timestamp is the moment of approval.
Approvals on the record
Policy changes follow a lifecycle — author, test, simulate, approve, roll out — with retained versions and a named approver. A rollback cannot be approved by its own author. The governance timeline replays every approve, defer, escalate and revoke.
Change management is evidenced, not asserted.
Before deployment
The cheapest evidence is the evidence recorded before deployment
“It went to production in March. We are asking in November.” If nothing was recorded at the moment the agent was approved, no effort in November reconstructs it. Assurance sits in the pipeline and records it then.
The bar is set first
Every measured domain reports its raw numerator and denominator, sample size, confidence interval, and a bar chosen before the result was seen.
Absence of evidence is not a pass
A domain whose prerequisite is missing returns not_assessed with the missing item named. A critical unit reporting it blocks the gate.
Unreachable never exits zero
Pass, blocked, malformed input, control plane unreachable and auth failure are five distinct exit codes. A plane that read no evidence cannot produce a green gate.
The result leaves as an artifact
A self-describing bundle, stamped with its staleness as of generation. A separate operator client retains and re-verifies captured bundles.
Frameworks and programs
Connect obligations to operating evidence
Each of these is a surface in the product. Framework detail lives on Evidence & Compliance; this is what the GRC conversation needs.
AI management system
ISO/IEC 42001 and NIST AI RMF
Every clause and subcategory with a coverage status and the named fields an owner attests to. Assessment runs, findings and attestation campaigns with owners, due dates and history.
Program workspace
EU AI Act
Register a system, get a risk tier from a scored classifier, track obligations against Articles 6, 9–16, 26, 50 and 52. Human-oversight and transparency records attach to the same record.
Program workspace
Colorado SB 26-189
Records for covered automated decision-making technology in consequential decisions — notice, post-decision explanation, correction and human review — ahead of the January 2027 effective date. Each obligation links to a specific audit event as its evidence.
Data-subject operations
GDPR
Right-of-access exports, erasure with explicit confirm and retry, consent grant and withdrawal, retention and residency. Erasure is gated by the legal-hold registry and fails closed.
Live control assessment
SOC 2, HIPAA, PCI DSS
Controls scored from real system state rather than a questionnaire — pass, fail, partial or not assessed.
Cross-framework
Control references
A seeded map of equivalent and overlapping controls between SOC 2, GDPR, HIPAA and PCI DSS, so evidence gathered once can be pointed at its counterpart.
Questions
Before we talk
The questions GRC leaders ask first.
Is GovernorAI a GRC platform?
Yes — an AI-native one. Program workspaces for the EU AI Act, Colorado SB 26-189 and GDPR, framework packs for ISO/IEC 42001 and NIST AI RMF, assurance gates, attestation campaigns, an evidence catalog and exports, signed where enabled for the deployment. It is not an enterprise-wide GRC suite: your system of record for the wider control universe stays, and GovernorAI connects to it. What separates GovernorAI from any GRC tool is that it enforces what it governs, so the evidence comes from the control acting, not from a questionnaire.
Can we keep our existing policies?
Yes. Native rules or OPA/Rego, including a connection to the OPA engine you operate. Your repository and approval process stay yours.
What can our assessors actually review?
Governed decision records with policy references and approval history, the assurance verdict for the version that acted, the control mapping with its status, and signed bundles they verify against the deployment’s keyring.
How does a design-partner evaluation start?
Bring one consequential agent workflow and its accountable owner. We identify the control, confirm integration fit, and agree success criteria for a scoped six-week evaluation. Discovery and shadow mode run before you enable enforcement.
Start with one workflow
Bring one workflow — and your hardest assessor question.
In a 30-minute briefing: the policy that should govern it, the control that enforces it, and the record your reviewer will ask for — answered with the artifact, not a slide.
Continue
Evidence & ComplianceHash-chained audit record, mapped to controls Policy & GovernanceOne policy, from intent to evidence AssuranceEvaluate an agent snapshot before production Decision ObservabilitySee the decision behind the action SecurityPer-call authorization and stop controls Platform & AI EngineeringGovern without touching the agent