How to export AI governance evidence to Vanta, Drata, or your GRC tool

Back each AI governance control in your GRC tool with the record from the control itself.

12 min read
In short: Arcjet's evidence page maps the AI inventory, owners, provider decisions, reviews, and runtime records to NIST AI RMF 1.0, ISO/IEC 42001, the EU AI Act, and frameworks you define. Every requirement is listed as supported or not covered. Export it as JSON or CSV, or push it to Vanta or Drata, then map each requirement to a control by its reference.

How do you export AI governance evidence to a GRC tool?

Compliance automation tools track controls, and they need evidence from the systems those controls govern. Export Arcjet's framework evidence as JSON or CSV, or push it to Vanta or Drata, so each AI governance control in your governance, risk, and compliance (GRC) tool is backed by the current inventory, owners, decisions, and runtime records, not by a screenshot.

AI governance evidence is the record that shows a governance requirement is met in practice. For an AI inventory requirement, that's the list of AI systems and how complete it is. For an ownership requirement, it's the named owner on each system. An AI system is an application or agent together with the model providers and tools it uses, and Arcjet's evidence is built from the AI systems, actors, and providers it observes.

This is the last of five how-to guides. The others build what the export contains: the AI inventory, owners, provider decisions, and reviews. For the full governance picture, see what is AI governance.

What does an AI governance evidence export contain?

An export contains one framework's requirements, each marked as supported or not covered, with the Arcjet evidence behind each supported one. Arcjet maps evidence to three frameworks and to frameworks you define:

  • NIST AI RMF 1.0, the National Institute of Standards and Technology AI Risk Management Framework.
  • ISO/IEC 42001, the international standard for AI management systems.
  • The EU AI Act, Regulation (EU) 2024/1689.
  • Customer-defined frameworks, such as your own AI policy, mapped to the same evidence.

The export lists every requirement in the framework, not only the ones Arcjet supports. That's deliberate. A file that lists only supported rows reads like full coverage. A file that lists every row, with "not covered" on most of them, shows your auditor exactly where other evidence has to come from.

The following table describes what each requirement in the export carries.

PartWhat it holds
Reference

The framework's own identifier, such as GOVERN 1.6 or Article 26(6)

Title

A short title in Arcjet's own words. The export doesn't reproduce framework text

CoverageSupported or not covered
Signals

The named Arcjet measurements behind the requirement, with their values

Note

What the evidence doesn't show, such as "only what Arcjet observes"

The export rolls up every standard project in your team. Temporary projects are left out. The framework itself is identified by name, version, and a link to its source, so a reader knows which edition the references belong to.

Which Arcjet signals back which governance requirements?

A signal is a fixed computation that Arcjet names and versions, such as the share of records with an owner. Frameworks map to signals, and each signal value drills down to the records behind it. The following table lists the main signals and the requirements they supply evidence for. The EU AI Act column is Arcjet's mapping, not text from the Act. The evidence page also maps these signals to ISO/IEC 42001, which the table leaves out because the standard's text is sold rather than published openly.

SignalWhat it measuresNIST AI RMFEU AI Act
inventory.systemsLive actors and observed providers, reviewed or notGOVERN 1.6None
inventory.coverageHealthy enrolled devices as a share of enrolled devicesGOVERN 1.6None
accountability.owner_assignedShare with a named ownerGOVERN 2.1Article 26(2)
accountability.purpose_documentedShare with a documented purposeMAP 1.1Article 26(1)
review.currentShare reviewed within 180 daysGOVERN 1.5None
third_party.decidedShare of observed providers with a decisionGOVERN 6.1None
third_party.violationsObservations of a prohibited provider after the decisionGOVERN 6.1, MANAGE 2.4None
lifecycle.retiredRetired actors, with the date and who retired themGOVERN 1.7, MANAGE 2.4None
runtime.policy_enforcedShare of live actors with at least one live Guard policyMEASURE 2.4None

Purpose drift, risk tier and EU AI Act role, Hold decisions, and long-term decision retention also appear as evidence where a framework asks for them. For example, long-term retention supplies evidence for EU AI Act Article 26(6). For the requirement-by-requirement view of each framework, see the NIST AI RMF checklist, the ISO/IEC 42001 Annex A checklist, and the EU AI Act checklist.

Should you export JSON or CSV, or push to Vanta or Drata?

Push to Vanta or Drata if you use one of them, and export JSON or CSV for any other GRC tool or for your own archive. The content is the same in each format. The following table compares them.

FormatFitsWhat you do with it
Push to Vanta or DrataTeams that track controls in Vanta or DrataMap each pushed requirement to the control it backs
JSON (JavaScript Object Notation)

Other GRC tools with an import, scripts, and an archive that keeps the full structure

Import it, transform it, or store it as the dated record
CSV (comma-separated values)Spreadsheets and auditors who want a tableAttach it to an audit request, or filter it by coverage

Keep a dated JSON copy even if you push. It's the format that keeps every field, and it's the one you can reread later without the GRC tool.

How do you map Arcjet evidence to controls in a GRC tool?

Map Arcjet evidence by framework reference: each requirement in the export carries the framework's identifier, so it attaches to the control in your GRC tool that covers the same requirement. If your GRC tool includes NIST AI RMF or ISO/IEC 42001 as a framework, match on the reference. If it doesn't, create one control per requirement that you want to track, named with the reference, and attach the evidence there.

Many organizations already have their own control set, such as an internal AI policy with a clause for approved tools. In that case, map from your control to the framework references it covers, then attach the Arcjet evidence for those references. A control for "approved AI tools only" maps to NIST AI RMF GOVERN 6.1, so third_party.decided and third_party.violations back it. For a clause-by-clause example, see the control map in the shadow AI policy template. If you'd rather keep the mapping in Arcjet, define the policy as a customer-defined framework and export it like any other.

Give the "not covered" rows owners in your GRC tool too. They're requirements Arcjet doesn't supply evidence for, such as AI literacy training or impact assessments, and each needs evidence from somewhere else. An auditor reads a requirement with no evidence and no owner as a gap.

Why is the export the point-in-time record?

The export is the point-in-time record because Arcjet doesn't recompute the evidence page for a past date. The page always shows the current state. Provider observations are kept for a limited retention period, so a value from six months ago can't be rebuilt honestly from what remains, and Arcjet doesn't try to.

That makes the export schedule part of your control. Export or push at a fixed interval, such as monthly and at the start and end of each audit period, and keep each file. An auditor who asks what your inventory looked like in March reads the March export. For the decision records behind the runtime signals, opt in to long-term decision retention, or send decisions to your security information and event management (SIEM) tool. For how long to keep them, see AI system log retention.

Changes between exports are recorded elsewhere. Every change to an owner, purpose, review, or provider decision is written to the WorkOS audit log with the fields that changed and the new owner, and the purpose text itself is left out.

How do you export AI governance evidence, step by step?

The following procedure produces evidence for one framework and keeps it current.

  1. Choose a framework. On the evidence page, choose NIST AI RMF 1.0, ISO/IEC 42001, the EU AI Act, or a framework you've defined.
  2. Review its coverage. Read which requirements are supported and which aren't. For each supported one, drill into the signal values and fix gaps first, such as systems with no owner or overdue reviews.
  3. Export or push. Download JSON or CSV, or push to Vanta or Drata. Keep a dated JSON copy either way.
  4. Map the evidence to controls in the GRC tool. Attach each supported requirement to the control with the same reference, or to your own control that covers it. Assign an owner to each requirement Arcjet doesn't cover.
  5. Repeat on a schedule. Export again at a fixed interval and at each audit milestone, so that you have a record for every period an auditor may ask about.

What Arcjet does not cover

Arcjet supplies evidence for some requirements in each framework. It doesn't satisfy any framework on its own, and the export says so by listing every requirement it doesn't support. The following areas stay with you:

  • Management system work. ISO/IEC 42001 specifies requirements for an organization's AI management system. Running that system is organizational work, and no product performs it for you.
  • Policies, training, and culture. Most of NIST GOVERN, such as AI risk management training, risk tolerance, and executive accountability, is organizational.
  • Impact and rights assessments. Arcjet records a risk tier and an EU AI Act role. The assessment behind them, including a fundamental rights impact assessment under Article 27, is yours.
  • Model quality. Validity, fairness, explainability, and accuracy measurements belong to the teams that build and evaluate models.
  • Systems Arcjet doesn't observe. Devices without the endpoint agent, applications without the SDK, and coding agents without hooks aren't in the evidence.
  • Past states. The evidence page shows the current state only, so keep your exports.
  • OSCAL. Arcjet exports JSON and CSV and pushes to Vanta and Drata. It doesn't export OSCAL (Open Security Controls Assessment Language).

Where Arcjet fits

Governance frameworks ask for the same four things: a list of the AI systems you run, a named owner and purpose for each, decisions about which providers are allowed, and evidence that those decisions hold. Arcjet builds that record from what it observes. Discovery finds the AI providers people use on managed devices, the SDK and coding-agent hooks report the agents you build and run, and Guard records every decision it enforces.

The evidence page maps those records to NIST AI RMF 1.0, ISO/IEC 42001, the EU AI Act, and frameworks you define, lists every requirement as supported or not covered, and exports JSON and CSV or pushes to Vanta and Drata. The evidence comes from the control itself, not from a survey about the control.

For the full governance picture, see what is AI governance. For what auditors ask of the runtime records, see compliance evidence for AI agent activity.

Frequently asked questions

Does Arcjet export OSCAL?

No. Arcjet exports AI governance evidence as JSON and CSV and pushes it to Vanta and Drata. If your GRC tool needs OSCAL (Open Security Controls Assessment Language), it isn't an Arcjet export format.

Does exporting Arcjet evidence satisfy NIST AI RMF, ISO/IEC 42001, or the EU AI Act on its own?

No. Arcjet supplies evidence for some requirements in each framework. The export lists every requirement and marks most as not covered, because policies, training, impact assessments, and management system work stay with your organization.

Can you get the evidence as it was on a past date?

No. The evidence page always shows the current state, and Arcjet doesn't recompute it for a past date. Export or push on a fixed schedule and keep each file. The export is the point-in-time record.

Which projects does the evidence cover?

The evidence rolls up every standard project in your Arcjet team. Temporary projects are left out. It covers only what Arcjet observes: enrolled devices, actors that report to Arcjet, and coding agents with hooks installed.

Can you map Arcjet evidence to your own AI policy?

Yes. Define the policy as a customer-defined framework, mapping each clause to Arcjet signals, and export it like the built-in frameworks. You can also map your own controls to NIST, ISO/IEC 42001, or EU AI Act references inside your GRC tool.

AI runtime security in your code

Protect your AI agent workflows with Arcjet

Get allow, deny, and redact on agent actions before the side effect.