Buyers use “AI governance” for five different jobs. Risk platforms (IBM watsonx.governance, Credo AI, Holistic AI) inventory and assess AI. OneTrust adds AI to a privacy and GRC suite. Collibra and Atlan extend a data catalog to AI assets. Microsoft and UiPath govern agents inside their own platforms. Trail, the memory and context layer for enterprise AI agents, holds the approved business rules those agents read.

What is an AI governance tool?

An AI governance tool keeps a record of the AI an organization runs, assesses each system’s risk, ties it to policies and regulations, and checks that controls hold once the system is live. The newer products also act at run time, blocking or escalating an agent action that breaks a policy.

NIST’s AI Risk Management Framework is the usual reference point. It “describes four specific functions to help organizations address the risks of AI systems in practice”: Govern, Map, Measure and Manage, with Govern described as “a cross-cutting function” that informs the other three nist.gov. NIST says the framework is “intended for voluntary use” nist.gov. Most vendors below ship templates for it, for the EU AI Act, or for ISO/IEC 42001.

Agents add to the job: a model scores a loan, while an agent reads an invoice, calls tools and posts an entry. Governing an agent means knowing that it exists, what it may touch, and which company rule it applied on each case.

How do the nine AI governance tools compare?

They differ most on where they sit: beside the AI as a record, inside the runtime as a guardrail, or under the agent as its rules.

Nine AI governance tools from each vendor's own pages, checked 19 October 2026
ToolKindWho it is forWhere it is strongWhat to check
Trail Rules layer for agentsBusiness owners whose agents must follow written policyRules cited to document, page and line; named approver per rule; versions kept; every read loggedNot an AI inventory, risk assessment or model-testing tool; developer docs page is a placeholdertrailhq.com
IBM watsonx.governance Lifecycle and risk platformRisk and compliance teams governing models and AI use casesGovernance graph, shadow AI detection, monitoring, GRC; published prices and a 14-day trialWe found no mention of agents on its product or pricing pageibm.com
Credo AI Lifecycle and risk platformGovernance teams that want regulatory policy packsAgent registry and agent cards; agentic risk library; packs for EU AI Act, NIST AI RMF, ISO 42001Agent Governor is a research preview; no published pricescredo.ai
Holistic AI Lifecycle and risk platformTeams that need testing and red teaming as well as inventoryDiscovery across AWS, Azure, GitHub and Databricks; Guardian Agents that block, redirect or escalateNo published prices; check which agent runtimes its Sentinels can watchholisticai.com
OneTrust Privacy and GRC suite with an AI modulePrivacy and security teams already on OneTrustInventory including agents; approval workflows; guardrails that block, redact or route; MCP policies with audit logsNamed runtime monitoring covers Amazon Bedrock and Microsoft Foundry; pricing on requestonetrust.com
Collibra Data catalog ledData governance teams extending to AIRegistry for use cases, models and agents; lineage from dataset to agent; agent contracts; guardian agentsNo published prices; announced the trail ML acquisition on Oct 5, 2026, so check the roadmapcollibra.com
Atlan Data catalog ledData teams that want AI assets next to data assetsAI asset registry, Jira and ServiceNow approval workflows, policy monitoring, certified context over MCPAutomatic intake, versioning and compliance readiness are labelled betaatlan.com
Microsoft Agent 365 and Purview Platform-native controlsIT and security teams on Microsoft 365Agent registry, least-privilege access to tools and MCP servers, lifecycle rules, audit loggingPurview lists no compliance management for ChatGPT Enterprise or Claude Enterprise agentslearn.microsoft.com
UiPath Platform-native controlsTeams building agents in UiPathAI Trust Layer policy for access, guardrails and models; publish limits for agentsPolicies apply to agents in a UiPath tenantdocs.uipath.com

Facts in this table were checked against the sources listed at the end of this page, between and .

Which AI governance tools cover AI agents?

Six of the eight products other than Trail name agents on their governance pages: Credo AI, Holistic AI, OneTrust, Collibra, Microsoft and UiPath. We found no mention of agents on IBM watsonx.governance’s product and pricing pages or in the feature list of Atlan’s AI governance module, though Atlan’s homepage serves certified context to agents.

We split “covers agents” into four questions and answered each from the vendor’s own pages; no vendor publishes this grid.

Agent coverage by tool (our reading of vendor pages, checked 19 October 2026)
ToolRegisters agentsActs on agent behavior at run timeFramework templates namedPublished prices
Trail No; it holds approved rulesServes only approved rules; agents below a confidence threshold must stop and askNone found on trailhq.comYes: Free, then $20k to $150k a year
IBM watsonx.governance Not stated for agents; inventories AI use cases and modelsContinuous monitoring of AI systemsNone named; cites a regulatory and risk content libraryYes, from $3,500 a month (Risk and Compliance Basic)
Credo AI Yes, agent registry and agent cardsAgent Governor, in research previewEU AI Act, NIST AI RMF, ISO 42001Not found
Holistic AI Yes, discovers agentsSentinel agents watch tool calls; Operative agents block or pause actionsEU AI Act, NIST AI RMF, ISO/IEC 42001Not found
OneTrust Yes, with a defined purposeBlock, allow, redact, escalate or route by policy; MCP policiesEU AI Act, NIST AI RMF, ISO 42001Not found
Collibra YesGuardian agents intervene; agent contracts set fleet-wide rulesEU AI Act, NIST AI RMF, AI UC-1Not found
Atlan Not stated for agents; registers AI models and applicationsPolicy monitoring for drift and other risksNone named on the module pageNot found
Microsoft Agent 365 and Purview Yes, Microsoft, partner and self-registered agentsAccess control, lifecycle rules, threat protectionNot checkedYes, $15.00 per user per month
UiPath Agents in its tenantGuardrails on agent LLM calls via the AI Trust LayerNot checkedNot checked

Facts in this table were checked against the sources listed at the end of this page, between and .

Write the requirement from the question someone is asking. An auditor asking which agents run and who owns them needs a registry, and a security lead worried about tool calls needs runtime enforcement. A controller asking why an agent approved an invoice needs the rule it applied and where that rule came from.

Which lifecycle and risk platforms fit a governance team?

IBM watsonx.governance pairs GRC and model monitoring with public prices. Credo AI leads with regulatory policy packs and an agent registry, and Holistic AI adds testing and red teaming.

IBM watsonx.governance

IBM watsonx.governance puts shadow AI first: it detects “unmanaged and unapproved AI usage across your organization” and links each system to its risks, controls and policies in a governance graph ibm.com. IBM publishes a price list. Risk and Compliance Basic starts at $3,500 a month for one instance, one module and one concurrent user. Model Management is pay as you go, and a self-managed software edition is priced by virtual processor cores ibm.com. There is a free 14-day trial. Neither the product page nor the pricing page mentions agents, so ask IBM to show an agent in the inventory.

Credo AI

Credo AI catalogs “every AI system, agent, vendor and model”, with an agent registry and agent cards. Its risk library is built for agents, covering “tool misuse, scope drift, and inter-agent risk”, and its policy engine ships ready-made packs for the EU AI Act, NIST AI RMF and ISO 42001 credo.ai. Named integrations include Snowflake, Databricks, AWS, Azure, ServiceNow, Jira and GitHub. Its run-time product for agents, Agent Governor, is labelled a research preview in the site menu, and the pricing URL returned a 404.

Holistic AI

Holistic AI discovers AI “across AWS, Azure, GitHub and Databricks”, and it puts testing up front: Holistic AI says it runs 40+ tests for bias, hallucination, privacy and robustness, plus red teaming for prompt injection. Its Guardian Agents run in two roles: “Sentinels watch each tool call, data access and hand-off, scoring it against your policies as it happens”, and Operative agents can “Block or pause non-compliant actions in flight” holisticai.com. The platform is described as “read-only by default”. Prices are not published, and we found no list on its homepage of the agent runtimes the Sentinels support.

What do privacy and data catalog suites add?

They add AI governance to a system the company already runs. That saves a new vendor review, but the AI module inherits the suite’s focus: privacy and data protection for OneTrust, data lineage and definitions for Collibra and Atlan.

OneTrust

OneTrust says it turns “AI policy into controls that work across homegrown and third-party AI”. Its inventory covers “AI systems, models, agents, datasets, vendors, projects, and use cases”, workflows handle intake, approval, attestation and signoff, and guardrails can “Block, allow, redact, escalate, or route actions by policy”. For agents it lists “Enforce MCP policies with audit logs” and governance of multi-agent handoffs onetrust.com. Runtime monitoring is named for Amazon Bedrock and Microsoft Foundry, with evaluation metrics also from Databricks and Gemini Enterprise Agent Platform, so check your own runtime against that list. The pricing page offers AI Governance as a package with a “Get Pricing” button and no figures onetrust.com.

Collibra

Collibra sells its AI Command Center as a control plane for “every AI agent and system across the full AI lifecycle”. For agents it offers contracts that, in Collibra’s words, “turn enterprise policies into machine-readable controls” across the fleet, and guardian agents that “automate oversight and enforcement in real time”. Lineage comes from its catalog roots and is captured “from source datasets through model training, inference, deployment and usage” collibra.com. On October 5, 2026 Collibra announced it had acquired trail ML “to automate AI governance from policy to production” collibra.com; trail ML is a separate company, and our governance explainer covers the name clash. Ask how the two products will merge before a multi-year deal.

Atlan

Atlan registers “AI models and applications in a single source of truth”, routes new AI assets and change requests through approval workflows tied to Jira and ServiceNow, and runs policies “for model drift, ethical AI, and other risks”. Three features on the module page carry a beta label: automatic intake, AI asset versioning and compliance readiness atlan.com. Its agent story sits on the homepage instead: domain experts “certify what’s production-ready” before data context reaches agents over MCP atlan.com. What gets certified there is data context such as definitions, not the agent itself.

When are platform-native agent controls enough?

They are enough when nearly all your agents run on that platform. Microsoft registers agents from other vendors, but its Purview compliance coverage for them is narrower, and UiPath’s policies cover agents in its own tenant.

Microsoft Agent 365 and Purview

Microsoft Agent 365 registers agents built on Microsoft platforms, partner agents and agents you register yourself. It controls “which users, data, and tools or MCP servers agents can use”, and its audit features add “logging, reporting, and audit capabilities” for agent actions. It costs $15.00 per user per month, paid yearly microsoft.com. Data compliance runs through Purview, and Purview’s AI agents table is the place to check scope. Copilot Studio agents get both information protection and compliance management. ChatGPT Enterprise agents get data classification and Insider Risk Management but no compliance management. Claude Enterprise agents are marked unsupported in both columns. Microsoft also notes that this “might require you to enable pay-as-you-go billing” learn.microsoft.com.

UiPath

UiPath governs agents through Automation Ops policies. Guardrails on agent LLM calls can be set once for every agent in a tenant or per agent in Studio Web, and a separate policy sets the limits an agent has to meet before it is published. “A policy takes effect in the tenants where it is deployed,” so agents built on other platforms sit outside it docs.uipath.com. Its 2025.10 release blog says the AI Trust Layer “pseudonymizes personally identifiable information before it reaches a model” uipath.com.

Where a rules layer like Trail sits in an AI governance program

The eight tools above govern the AI: what exists, how risky it is, what it may call. Trail governs the business rules an agent applies on a case, such as an AP tolerance or a discount approval limit, and records which approved version it read.

For a governance team, this layer holds evidence the other tools ask about but do not keep. Trail drafts rules from docs, SOPs and systems into its Context Graph, and each extracted piece “keeps the document, page and line it came from” trailhq.com. No draft reaches an agent until someone on your team signs it off; the Rule Engine page calls that review gate “the product, not a setting”. The approver’s name and time, earlier versions and every agent read are recorded, and access is scoped by role for agents as well as people trailhq.com.

Agents cannot edit rules. Feedback from a run comes back as a proposed change that a person approves, edits or rejects, and your team picks the confidence level below which an agent has to stop and ask trailhq.com. Agents connect over an MCP server, native retrievers or REST/GraphQL, and deployment can be managed, single-tenant or inside your VPC trailhq.com.

Look elsewhere if the gap is an AI inventory, EU AI Act risk classification, bias testing or shadow AI discovery; we found none of those on trailhq.com. Developer docs are not public yet: trailhq.com/docs reads “Coming soon” trailhq.com. After a Free plan, paid plans run from $20k to $150k a year trailhq.com. Rule engines built for back-office agents are compared in AI rule engines for finance and operations agents.

What should you check against NIST AI RMF and the EU AI Act?

Map each requirement to a named control and ask the vendor to show it working on your own agents. The NIST subcategories below are voluntary; the EU AI Act articles apply to high-risk AI systems as the Act defines them, so check with counsel whether any of your agents falls under them.

Checks to run against your shortlist

Checklist
  • Inventory (NIST GOVERN 1.6)NIST asks for mechanisms to inventory AI systems. Ask each vendor to discover one agent you did not register and show its owner and risk level.
  • Human oversight (NIST MAP 3.5, EU AI Act Art. 14 and 26)Article 26(2) says deployers of high-risk AI systems shall assign human oversight to natural persons with the necessary competence, training and authority, and the necessary support. Ask where in the tool a named person approves, and what the agent does while it waits.
  • Logs you can keep (EU AI Act Art. 12 and 26)Article 26(6) says deployers of high-risk AI systems shall keep the logs the system generates automatically, to the extent those logs are under their control. The period is at least six months unless Union or national law provides otherwise. Check retention, export and whether logs cover agent tool calls or only model calls.
  • Post-deployment monitoring (NIST MANAGE 4.1)Ask which runtimes the monitoring connects to. Several vendors name specific clouds; agents elsewhere may sit outside it.
  • Decision traceabilityPick one agent decision from last month and ask the tool to show the rule or policy it applied, the version of that rule and who approved it.

Article 12 says high-risk systems “shall technically allow for the automatic recording of events (logs) over the lifetime of the system” eur-lex.europa.eu. For the same questions applied to agent memory, see our guide to governing a memory layer for LLM agents.

Which AI governance tool should your team start with?

Start from who raised the problem, because each group buys a different layer.

A decision rule for choosing an AI governance tool

Step list
  1. Risk or compliance needs an AI register and framework evidenceIBM watsonx.governance if you want public prices and GRC; Credo AI for regulatory policy packs and an agent registry; Holistic AI if testing and red teaming are in scope.
  2. Privacy already runs on OneTrustAdd the OneTrust AI module and confirm your agent runtimes are on its monitoring list.
  3. The data team owns governanceCollibra for agent contracts and lineage to the agent; Atlan for certified data context, noting the beta features.
  4. Most agents run in one platformMicrosoft Agent 365 with Purview, or UiPath Automation Ops. Check coverage for agents built elsewhere.
  5. The business owner needs agents to follow written policyTrail, or another rules layer with per-rule approval, citations to the source line and version history, next to one of the tools above.

How we chose these nine AI governance tools

On 19 October 2026 we read each vendor’s product, pricing and documentation pages with curl and checked every quote against the fetched text. We picked tools to cover the five kinds of product buyers call AI governance, favoring those that publish enough detail to compare. The criteria were agent registration, runtime action on agents, framework templates, approval workflows and published prices; silence on a page is reported as “not found”. We read ModelOp’s homepage and left it out because it overlaps IBM and Credo AI on lifecycle governance; ServiceNow, Google Cloud and AWS controls were not checked. We ran no product trials. Trail publishes this page and is one of the nine.

Frequently asked questions

Best AI governance tools?

It depends on who owns the problem. Risk and compliance teams usually start with a lifecycle platform such as IBM watsonx.governance, Credo AI or Holistic AI; privacy teams already on OneTrust can add its AI module; data teams can extend Collibra or Atlan; and Microsoft or UiPath shops get controls inside their own platform. Check agent coverage, runtime enforcement and framework templates before you choose.

Which AI governance tools cover AI agents?

On their own pages, Credo AI has an agent registry with agent cards, Holistic AI has Guardian Agents whose Sentinels watch each tool call, OneTrust enforces MCP policies with audit logs, Collibra has agent contracts and guardian agents, Microsoft Agent 365 registers agents across the organization, and UiPath sets publish limits for its agents. A rules layer such as Trail covers a different part: the approved, cited rules an agent reads before it decides a case.

How do AI governance tools map to the NIST AI RMF?

The NIST AI RMF has four functions: Govern, Map, Measure and Manage. Inventories support GOVERN 1.6, approval workflows support MAP 3.5 on human oversight, and runtime monitoring supports MANAGE 4.1 on post-deployment monitoring. The framework is voluntary, so the mapping is a planning aid, not a certification.

Do AI governance tools publish prices?

Some do. IBM lists watsonx.governance Risk and Compliance Basic from $3,500 a month for one instance, one module and one concurrent user, and Microsoft lists Agent 365 at $15.00 per user per month, paid yearly. We found no price list for Credo AI, Holistic AI, OneTrust, Collibra or Atlan; their sites point to a demo request or a sales form.

Where should an AI governance rollout start?

Start with an inventory: every later control depends on knowing which AI systems and agents exist and who owns them, and NIST's GOVERN 1.6 asks for one. Add approval steps, run-time monitoring and decision traces as your audit or risk findings call for them.

Is the NIST AI RMF changing?

Yes. NIST says AI RMF 1.0, released on January 26, 2023, is being revised as part of the White House AI Action Plan. Ask vendors how their NIST templates will follow the revision.

Sources checked

Show 22 sources