How do AI governance platforms differ?
AI governance platforms differ in where their record comes from. GRC-led platforms manage policies, assessments, and registers. Model-led platforms govern models through their development life cycle. Compliance automation tools map AI controls to audits. Device and data tools see AI use on endpoints and in data. Runtime-led platforms build the inventory and evidence from what they observe and enforce.
Pick by where your AI risk lives. If you train and deploy your own models, a model-led platform tracks the thing you build. If most of your AI is third-party providers, coding agents, and agents built on someone else's model, the risk is in what those systems reach and do, and the record has to come from somewhere that sees it. Every approach needs a register that people maintain, and every approach needs evidence that the register matches reality. They differ in which of the two they start from.
Arcjet publishes this comparison and sells a runtime-led platform, which is one of the five approaches on this page. Vendor details come from each vendor's public documentation, reviewed on November 16, 2026. Vendors appear as examples of an approach, not in a ranking, and several span more than one approach. For the governance model this page assumes, see what is AI governance.
How do the five approaches compare?
Each approach starts from a different source of truth, and that source decides what its inventory includes and what its evidence proves. The following table compares them.
| Approach | Where the inventory comes from | What enforces decisions | What its evidence proves | Example vendors |
|---|---|---|---|---|
| GRC-led | Intake forms, questionnaires, and connectors to cloud and machine learning platforms | Approval workflows; some add runtime guardrails through integrations | That assessments, approvals, and policies exist and were signed off | Credo AI, OneTrust AI Governance, Holistic AI |
| Model-led | Models registered in the development and deployment pipeline | Deployment gates and model monitoring thresholds | How each model was built, evaluated, and how it performs in production | IBM watsonx.governance, ModelOp |
| Compliance automation | Controls, policies, and use cases you enter, plus integrations with your infrastructure | Automated control tests that flag failures | That controls mapped to a framework pass their tests | Vanta, Drata |
| Device and data | Endpoint telemetry, browser extensions, and data classification | Device configuration, data loss prevention policies | Which AI apps people use, and what sensitive data reaches them | Microsoft Purview, Jamf AI Governance |
| Runtime-led | Agents that report decisions, coding-agent hooks, and device observation of AI providers | Policies that decide each tool call and provider connection before it happens | That each decision held, with the record of each enforcement | Arcjet |
The lines between approaches blur. Several GRC-led platforms add runtime guardrails, and device tools add inventories. The column that changes least is the first one: where the inventory comes from decides what the platform can find without someone telling it.
What does each approach do well, and where does it stop?
The following sections take each approach in turn, with the vendors that document it.
GRC-led AI governance platforms
A governance, risk, and compliance (GRC) led platform holds the register of AI systems, the policies that apply to them, and the assessments and approvals that each one passed. Its strength is the process around a system: intake, risk tiering, review routing, and framework templates.
Credo AI's product page describes a centralized inventory of AI systems that includes agents, models, apps, and shadow AI, with records of each agent's purpose, tools, data sources, and guardrails. It lists policy packs for the EU AI Act, NIST AI RMF, ISO 42001, and SOC 2, approval gates, and a third-party AI registry with discovery across cloud environments. It also describes trace ingestion and continuous evaluation of agent traces, and lists enforcement integration with CI/CD, CASBs, and API gateways as planned. OneTrust AI Governance inventories AI systems, models, agents, datasets, vendors, projects, and use cases, routes intake and approval workflows to owners, automates risk tiering, and includes templates for the EU AI Act, NIST AI RMF, and ISO 42001. It connects to Amazon Bedrock, Microsoft Foundry, Google Vertex, and Databricks Unity Catalog, and describes runtime guardrail enforcement that can block, allow, redact, escalate, or route actions, with MCP policies. Holistic AI describes discovery of models, agents, and AI applications across AWS, Azure, GitHub, and Databricks, pre-launch and production testing for bias, hallucination, privacy, and robustness, control mapping to the EU AI Act, NIST AI RMF, and ISO/IEC 42001, and agents that monitor and can block, pause, or redirect non-compliant agent actions.
The register is only as complete as its sources. Connectors to cloud and machine learning platforms find what runs there. A coding agent on a developer's laptop or a ChatGPT tab doesn't appear in a cloud account, so it reaches the register only if someone reports it or another source feeds it in. For a direct comparison, see Credo AI vs Arcjet and OneTrust AI Governance vs Arcjet.
Model-led AI governance platforms
A model-led platform governs models through the life cycle that produces them: development, evaluation, approval, deployment, and monitoring. Its strength is the model itself, including the metrics a data science team and a model risk function need.
IBM's model governance page for watsonx.governance describes factsheets that log model facts and can be exported for audits, monitoring of performance metrics for drift, quality, and safety, and alerts on model inputs and outputs. It says watsonx.governance governs generative AI models built on watsonx.ai and on third-party platforms, including Amazon Bedrock, Microsoft Azure, and OpenAI. The product page shows logos for the EU AI Act, NIST AI, and ISO 42001. ModelOp describes a system of record for machine learning, generative AI, agentic, and vendor AI, with lifecycle automation, policy applied at delivery time, and operational visibility into cost, token use, and risk across more than 50 connected technologies.
A model-led platform fits teams that build and ship models. If you mostly consume models through agents and third-party tools, the model's accuracy is the provider's concern, and your risk sits in what the agent does with the model's output. For a direct comparison, see IBM watsonx.governance vs Arcjet.
Compliance automation tools
A compliance automation tool maps controls to frameworks, collects evidence from your infrastructure, and runs tests that show whether each control passes. Its strength is the audit: one place where a framework's requirements, your controls, and the evidence for each meet.
Vanta's ISO 42001 page describes mapped controls, policy templates with version control, automated tests that monitor controls hourly, and work with ISO 42001-accredited auditors. Its AI governance page says Vanta helps teams work toward ISO 42001, EU AI Act, and NIST AI RMF requirements. The same page's FAQ describes its AI inventory and guardrails as a separate paid product that wasn't available on the review date, so this page doesn't count them. Drata's ISO 42001 page describes a control library with clear ownership, centralized evidence, continuous control monitoring, and policies aligned to controls with tracked reviews.
These tools collect evidence from the systems you connect, and they don't observe AI use themselves. That's why Arcjet integrates with them rather than replacing them: Arcjet's evidence page pushes to Vanta and Drata, so the AI-specific evidence lands in the audit workspace you already use. For the export, see AI governance evidence export.
Device and data tools
A device or data tool sees AI use where it happens on the endpoint, or where sensitive data moves. Its strength is coverage of people using AI apps, which neither a register nor a model pipeline sees.
Microsoft Purview Data Security Posture Management (DSPM) includes an AI observability page that inventories AI apps and agents with activity in the last 30 days, and an activity explorer that shows AI interactions, including prompts and responses and whether they contained sensitive information. Microsoft's classic DSPM for AI documentation says third-party AI sites need the Purview browser extension and onboarded devices, and that its recommendations include regulatory templates from Compliance Manager for AI regulations. Jamf AI Governance describes an inventory of AI tools, agents, and MCP servers running on Macs, vendor AI configuration deployed and scoped by user or group, and a record of every policy decision, configuration, and administrative action. The page mentions the EU AI Act's transparency rules in a note, and names no other compliance framework.
These tools see use, not purpose or ownership. An inventory of apps on a device doesn't say who answers for each one or what it's for, and a data loss prevention match doesn't say which AI system the data belonged to. For a direct comparison, see Jamf AI Governance vs Arcjet.
Runtime-led AI governance
A runtime-led platform builds the governance record from the controls that enforce it. The inventory comes from what reports to the platform and what it observes, the decisions are enforced by the same platform, and the evidence is the record of each enforcement.
Arcjet works this way. Discovery finds the AI providers people reach from managed macOS and Windows devices. The SDK and coding-agent hooks for Claude Code, GitHub Copilot, Cursor, OpenAI Codex, and Muse Code report the agents you build and run. Arcjet groups each agent with the providers and MCP servers it uses into an AI system that a person confirms. Owner, purpose, risk tier, and review sit on that system. Provider decisions are enforced on devices, and Guard decides each tool call before it runs. The evidence page maps the result to NIST AI RMF 1.0, ISO/IEC 42001, and the EU AI Act, plus frameworks you define, and lists every requirement as supported or not covered.
The limit is the mirror of the GRC-led approach. A runtime-led platform sees only what reports to it or runs on enrolled devices, and it holds no policy documents, training records, or impact assessments. For more about how the two layers relate, see AI governance vs AI runtime security.
What should you ask an AI governance platform vendor?
Ask each vendor where its inventory comes from and what its evidence proves, because the answers separate the approaches more than feature lists do. The following questions work in a demo or a proof of concept:
- Where does each inventory entry come from? A survey answer, a cloud connector, a device, and a runtime record each prove something different. For the trade-off, see observed vs survey-based AI inventory.
- Does the platform state its own coverage? An inventory that reports how many enrolled devices are healthy, or how many agents report to it, tells you what it might be missing.
- Does it find coding agents and their MCP servers? Ask the vendor to show a Claude Code session on a developer laptop and the MCP servers it called.
- Where is a decision enforced? Look for a prohibited provider blocked on the device, or its use counted as a violation and attributed to a person.
- What happens when an owner leaves? Look for ownership that lapses automatically rather than pointing at a former employee.
- How does the evidence count? Prefer a platform that lists every framework requirement and marks each as supported or not covered, over one that reports a compliance percentage.
- Can you export the evidence? Ask for JSON or CSV, and a push to the compliance automation tool you already run.
Which approach fits your organization?
Find the row that matches your organization. Each row pairs a primary approach with the second one that covers its largest gap.
| Your situation | Start with | Add | Why |
|---|---|---|---|
| You train and deploy your own models | Model-led | GRC-led or runtime-led | Model metrics cover the model; agents built on it still need an owner and enforced decisions |
A large enterprise with a risk function and many AI use cases | GRC-led | Runtime-led or device and data | The register needs a source that finds AI nobody reported |
A software company whose AI is coding agents and third-party providers | Runtime-led | Compliance automation | The runtime builds the record; the audit tool holds it next to your other frameworks |
| An ISO/IEC 42001 audit is on the calendar | Compliance automation | Runtime-led | The audit tool covers the management system; runtime evidence covers the AI systems in scope |
| Your concern is staff pasting data into AI apps | Device and data | Runtime-led | Data loss prevention covers the paste; provider decisions cover which apps are allowed |
Most organizations end up with two approaches, because each one records a different part of the same governance program. The question to settle first is which one holds the inventory, because the others feed it.
Where does Arcjet's approach fit, and where doesn't it?
Arcjet covers the runtime-led approach. It fits teams whose AI risk is in agents and third-party providers: engineering organizations running coding agents, companies building agents on hosted models, and fleets of managed Macs and Windows laptops where people reach AI services. The inventory, owners, provider decisions, violations, and purpose drift sit on the same records Guard and Discovery produce. Every change to an owner, purpose, review, or provider decision is written to the WorkOS audit log. For how the inventory is built, see AI inventory.
Arcjet is the wrong choice in the following situations:
- You need a policy library, training records, and impact assessments. Arcjet records a risk tier and an EU AI Act role per AI system, but the assessment behind it belongs in a GRC-led platform.
- You need model evaluation. Arcjet doesn't measure accuracy, bias, or model drift. Use a model-led platform.
- Your AI runs where Arcjet can't see it, such as unmanaged devices or agents that don't report to Arcjet. Use a source that reaches them.
- You want a management system audit run end to end. Use a compliance automation tool, and push Arcjet's evidence into it.
To see the inventory and evidence for your own AI systems, book a demo. For the governance model behind this page, see what is AI governance.
Frequently asked questions
What is an AI governance platform?
An AI governance platform is software that keeps the record of an organization's AI systems, the decisions about them, and the evidence that those decisions are followed. Platforms differ mainly in where that record comes from: questionnaires and connectors, the model development pipeline, compliance control tests, device and data telemetry, or the runtime controls that enforce decisions.
Do I need a GRC tool for AI governance?
Not always. A GRC-led platform suits organizations with a risk function, many AI use cases, and a need for assessments and policy libraries. A software company whose AI is mostly coding agents and third-party providers can start from a runtime-led inventory and push its evidence into a compliance automation tool such as Vanta or Drata.
Is Arcjet a replacement for Vanta or Drata?
No. Vanta and Drata map controls to frameworks such as ISO/IEC 42001 and support audits. Arcjet observes and enforces AI use and pushes its AI governance evidence to Vanta and Drata, so the evidence sits in the audit workspace you already use.
Which AI governance approach finds coding agents?
Approaches that observe developer machines or the agents themselves. Coding-agent hooks report each tool call and MCP server, and endpoint observation sees connections to model providers. Cloud connectors and questionnaires usually miss coding agents, because they run on laptops and are installed without an intake form.
How did Arcjet check the vendor details in this comparison?
Each vendor detail comes from that vendor's public documentation or product pages, reviewed on November 16, 2026, and each is linked where it appears. Vendors appear as examples of an approach, not in a ranking. Arcjet publishes the comparison and sells the runtime-led approach.
AI runtime security in your code
Protect your AI agent workflows with Arcjet
Arcjet runs inside your application, where it can use runtime context to enforce agent actions and budgets, detect prompt injection, and protect sensitive information before a workflow acts.