What is the difference between IBM watsonx.governance and Arcjet?
IBM watsonx.governance governs AI models and use cases through their life cycle: it records each model in a factsheet, evaluates quality, fairness and drift, and connects the results to risk and compliance workflows in IBM OpenPages. Arcjet builds the AI inventory from what devices and agents do, and enforces provider decisions and agent policies at runtime.
The two products govern different things. watsonx.governance is model-led. It follows a model or prompt from the request for a new use case through development, evaluation and production monitoring, and it serves data science, model risk and audit teams. Arcjet is runtime-led. It records which AI providers people reach from managed devices, decides each coding-agent action, and decides tool calls in the agents you build. If you train, tune or host models and need to show how each one was evaluated and approved, watsonx.governance does that work. If your risk sits in coding agents and third-party AI providers that you don't train, Arcjet supplies the inventory and the decisions. Organizations with both kinds of AI can use both.
Arcjet publishes this comparison. IBM details come from IBM's public product pages, documentation, release notes and announcements, reviewed on November 16, 2026. Some IBM pages are dated 2025, and products change, so check the linked sources before you decide.
How do IBM watsonx.governance and Arcjet compare at a glance?
watsonx.governance governs models through their life cycle, and Arcjet governs AI use at runtime. The following table compares the two products across 10 areas.
| Area | Arcjet | IBM watsonx.governance |
|---|---|---|
| Category | Runtime-led AI governance and AI agent runtime security | Model-led AI governance with an enterprise risk and compliance suite |
| Unit of governance | The AI system: an application or agent with the model providers and tools it uses | The AI use case, with a factsheet for each tracked model or prompt |
| How the inventory is built | Observed from devices, coding-agent hooks, software development kits (SDKs) and agent records, then confirmed by people | Use cases created by people; models evaluated in watsonx.governance; AI found by IBM Guardium AI Security |
| Model evaluation | Not covered | Quality, fairness, accuracy, drift and explainability for machine learning models; task metrics for foundation models |
| Employee AI use | Each connection to an AI provider from managed macOS and Windows devices, attributed to a person | Shadow AI detection, with Guardium AI Security sources named as cloud environments, code repositories and embedded systems |
| Coding agents | Hooks in Claude Code, GitHub Copilot, Cursor, OpenAI Codex and Muse Code decide each action | Not described in the pages reviewed |
| Runtime controls | Guard decides tool calls and prompts; the endpoint agent blocks prohibited providers | Thresholds and alerts on model inputs and outputs, and AI guardrails that use a Granite Guardian model as a filter |
| Risk and compliance | Risk tier and EU AI Act role per AI system; evidence page for the National Institute of Standards and Technology (NIST) AI Risk Management Framework (AI RMF), ISO/IEC 42001, the EU AI Act and your own frameworks | IBM OpenPages for risk, controls and compliance; EU AI Act, NIST and ISO 42001 named on the product page |
| Deployment | Arcjet cloud service, with the endpoint agent on devices | IBM Cloud, AWS, and self-managed software, including a version authorized under the Federal Risk and Authorization Management Program (FedRAMP) on AWS GovCloud |
| Pricing | Platform plans, with Discovery through a demo | Listed starting prices per plan; software priced by virtual processor cores |
What is IBM watsonx.governance?
IBM watsonx.governance is IBM's platform for governing AI models and use cases through their life cycle, with risk and compliance management through IBM OpenPages. IBM's product page describes a governance graph that maps AI systems to risks, controls and policies, automated mapping of obligations to requirements, capture of compliance evidence, continuous monitoring, and shadow AI detection that finds unmanaged and unapproved AI use. The page shows logos for the EU AI Act, NIST AI and ISO 42001 among the frameworks it supports.
The documentation organizes governance around the AI use case. You create a use case to track AI assets from request through production, and each tracked model or prompt gets a factsheet that records its details. For machine learning models, watsonx.governance evaluates quality, fairness and accuracy, and drift evaluations find changes in model output. For foundation models, it evaluates prompt templates for tasks such as summarization, generation and question answering. In production, it monitors fairness, performance and data consistency, and explains predictions.
The model governance page says watsonx.governance governs models built on third-party platforms, including Amazon Bedrock, Microsoft Azure and OpenAI. Preset thresholds monitor generative AI inputs and outputs and alert on toxic language, hate speech, abusive language and profanity. The page says you can deploy in the cloud or on premises, and links an announcement of a FedRAMP-authorized version on AWS GovCloud.
The release notes for the self-managed software add detail. Version 2.2.0 (June 2025) added an Amazon SageMaker integration to the governance console. Version 2.2.1 (August 2025) added a governed agentic catalog to find, add and compare AI tools and agents, and AI guardrails that use a Granite Guardian model as a filter.
For agents, IBM's agentic governance announcement describes more than 50 evaluation metrics that you add with a Python decorator on a tool node in a LangGraph application. Your application can use a computed metric to change its flow, for example by running a web search when retrieved context is irrelevant. The announcement lists production alerts on those metrics as an upcoming feature. The pages reviewed don't describe watsonx.governance itself blocking an agent's tool call.
Shadow AI detection connects to IBM Guardium AI Security. IBM's integration announcement says Guardium detects new AI use cases in cloud environments, code repositories and embedded systems and brings them into watsonx.governance, where risk and compliance controls apply.
The pricing page lists indicative starting prices: model management from $0.64, Risk and Compliance Basic on IBM Cloud from $3,500 per month, Risk and Compliance Advanced as enterprise software from $6,450 per month, and an AWS plan from $42,000 that includes 5 AI use cases, 12,000 evaluations and 25 concurrent users. Self-managed software is priced by virtual processor cores.
What is Arcjet?
Arcjet is an AI agent runtime security platform that records governance on top of its runtime layer. It builds the AI inventory from what it observes, records who owns each system, enforces provider decisions and agent policies, and maps the records to framework evidence.
An AI system is an application or agent together with the model providers and tools it uses. Arcjet suggests AI systems by grouping an actor with the providers and Model Context Protocol (MCP) servers it uses, and a person confirms each grouping. The observations come from the Discovery endpoint agent on managed macOS and Windows devices, which records each connection to an AI provider and attributes it to a person; from coding agent hooks in Claude Code, GitHub Copilot, Cursor, OpenAI Codex and Muse Code; and from the agents you build through Arcjet's SDKs, the Claude Compliance API and OpenTelemetry.
Each AI system carries an owner, who must be a member of the team, a purpose, a review every 180 days, a risk tier and its EU AI Act role. An owner who leaves the team reads as no owner. Only team administrators edit these fields, and each change goes to the WorkOS audit log. Retiring an actor revokes its keys and keeps the record.
Provider decisions approve or prohibit an AI provider, or one client of it, such as the ChatGPT app as opposed to the OpenAI API. The Discovery agent enforces the decision on managed devices, and an observation of a prohibited provider at or after the decision time is a violation attributed to a person. Purpose drift compares the data classes and tools an owner declared with sensitive-information findings and captured tool calls, and notifies the owner. A weekly digest sends each owner their overdue reviews, violations and drift.
Arcjet Guard decides tool calls and prompts in the agents you build, with dry run, and a rule must pass a test before it goes live. Publishing a policy can require a second approver, and Hold decisions go to the system's owner. Sensitive-information detection covers personally identifiable information (PII), secrets and protected health information (PHI).
The evidence page maps these records to NIST AI RMF 1.0, ISO/IEC 42001 and the EU AI Act, plus frameworks you define, and marks every requirement as supported or not covered. You export it as JSON or CSV or push it to Vanta or Drata. Long-term decision retention is opt-in, for frameworks with log-keeping duties.
How do IBM watsonx.governance and Arcjet differ?
watsonx.governance and Arcjet differ in what each treats as the thing being governed, what each measures, and whether each decides at runtime. The following sections cover each difference, then the areas where watsonx.governance goes further.
Models you build compared with AI you use
watsonx.governance is built around models. A use case holds models and prompts, a factsheet records how each was built and evaluated, and monitoring measures how each behaves in production. That fits an organization that trains or tunes models, or that has to show a regulator how a credit, hiring or claims model was validated.
Arcjet is built around use. Most of the AI that a software company or an enterprise IT team needs to govern is third-party: coding agents, chat assistants and model APIs. You can't evaluate the fairness of a model you don't host, but you can record who uses it, decide whether it is allowed, and decide what an agent may do with it. Arcjet records those decisions on the AI system and enforces them.
Evaluation metrics compared with decisions
watsonx.governance measures models: fairness across features, drift in output, quality against ground truth, and task metrics for generative AI and agents. Those metrics tell a model owner whether a model still behaves as validated, and alerts fire when a threshold is crossed. For agents, the announcement reviewed describes metrics your own code can act on, with production alerts listed as upcoming.
Arcjet decides actions. When a coding agent is about to read a credential file, call an MCP server outside your allowlist or send data to a host outside your egress allowlist, the hook asks Arcjet and the policy allows or denies the call before it runs. Arcjet doesn't score model quality or fairness, and watsonx.governance, from the pages reviewed, doesn't decide a coding agent's tool calls.
Discovery sources
watsonx.governance finds AI through the use cases people create, the models evaluated in it, and Guardium AI Security, which looks in cloud environments, code repositories and embedded systems. Those sources find AI that is deployed or committed in code.
Arcjet finds AI from managed devices and agents: each connection to an AI provider, attributed to a person, and each coding-agent session. That covers the ChatGPT app a finance team uses and the Claude Code session a developer runs, which leave nothing in a cloud account or repository. Neither set of sources covers everything, and the evidence page states Arcjet's coverage so you can see the gap. For more on the tradeoff, see observed vs survey-based AI inventory.
Where watsonx.governance goes further
watsonx.governance does several things that Arcjet doesn't. It evaluates and monitors models for quality, fairness, drift and explainability, records each model's history in a factsheet, and governs models built on Amazon Bedrock, Microsoft Azure, OpenAI and Amazon SageMaker. Through OpenPages, it runs enterprise risk, controls and compliance workflows for finance, risk and audit teams. It runs as self-managed software for organizations that keep governance on premises, and it has a FedRAMP-authorized version on AWS GovCloud.
When should you choose IBM watsonx.governance?
Choose IBM watsonx.governance when you build, tune or host models and need to show how each one was evaluated, approved and monitored. watsonx.governance fits when any of the following apply:
- You train or tune machine learning or foundation models and need fairness, drift and quality metrics.
- You need a factsheet for each model that follows it from request to production.
- Your risk and audit teams already use IBM OpenPages.
- You need on-premises deployment or a FedRAMP-authorized service.
- You want to evaluate agents built with LangGraph using IBM's agent metrics.
When should you choose Arcjet over IBM watsonx.governance?
Choose Arcjet when the AI you need to govern is mostly third-party and used by people and agents, and you need decisions enforced at runtime. Arcjet fits when any of the following apply:
- You need to know which people reach which AI providers from managed macOS and Windows devices.
- You want provider decisions enforced on devices, with violations attributed to people.
- Your developers use Claude Code, GitHub Copilot, Cursor, OpenAI Codex or Muse Code, and you want each action decided before it runs.
- You build agents and want tool calls and prompts decided at runtime, with dry run and tested rules.
- You want an owner, purpose and review on each AI system, with evidence exported as JSON or CSV or pushed to Vanta or Drata.
Can you use IBM watsonx.governance and Arcjet together?
Yes. The two products cover different parts of an AI estate, so they overlap little. watsonx.governance holds the models you build, their evaluations and their risk records. Arcjet holds the coding agents, third-party providers and agents in use, with their owners, decisions and runtime records.
An agent you build can sit in both: watsonx.governance evaluates the model and prompt behind it, and Arcjet Guard decides its tool calls in production. Arcjet's evidence export, in JSON or CSV, can feed the AI systems it observes into your watsonx.governance use cases or OpenPages records. Neither product documents a direct integration with the other, so plan for an import your team maintains, and keep each system's owner and purpose in one place.
For the layering in general, see AI governance vs AI runtime security.
What are the alternatives to IBM watsonx.governance?
The alternatives to IBM watsonx.governance differ in whether they start from a register, from models or from runtime records. The following products cover overlapping ground:
- Arcjet builds an observed AI inventory from devices, coding agents and the agents you build, enforces decisions at runtime, and maps the records to framework evidence.
- Credo AI runs an AI registry, a risk assessment library and policy packs for the EU AI Act, NIST AI RMF, ISO/IEC 42001 and SOC 2. For more information, see Arcjet vs Credo AI.
- OneTrust AI Governance inventories AI from Amazon Bedrock, Microsoft Foundry, Google Vertex and Databricks, and runs assessments in OneTrust's privacy and risk platform. For more information, see Arcjet vs OneTrust AI Governance.
For the categories side by side, see AI governance platforms compared.
Frequently asked questions
How is IBM watsonx.governance different from Arcjet?
IBM watsonx.governance is model-led: it tracks AI use cases and models with factsheets, evaluates quality, fairness and drift, and connects to IBM OpenPages for risk and compliance. Arcjet builds the AI inventory from Discovery on managed macOS and Windows devices, coding-agent hooks and the agents you build, records an owner, purpose and 180-day review per AI system, enforces provider decisions on devices and agent policies at runtime, and maps the records to NIST AI RMF 1.0, ISO/IEC 42001 and the EU AI Act on its evidence page. watsonx.governance measures the models you build; Arcjet decides what people and agents do with AI.
Is Arcjet an alternative to IBM watsonx.governance?
For governing third-party AI that people and coding agents use, with decisions enforced on devices and on agent actions, yes. For evaluating and monitoring models you train or host, keeping a factsheet per model, or running enterprise risk workflows in OpenPages, watsonx.governance does work Arcjet doesn't, and the two can run side by side.
Does watsonx.governance block AI agent actions?
The IBM pages reviewed on November 16, 2026 describe evaluation metrics for agents that your application code can act on, thresholds and alerts on model inputs and outputs, and AI guardrails that use a Granite Guardian model as a filter. They don't describe watsonx.governance itself blocking an agent's tool call. Arcjet decides each coding-agent and Guard-protected tool call before it runs.
How does watsonx.governance find shadow AI?
IBM says watsonx.governance detects unmanaged and unapproved AI use, and that IBM Guardium AI Security detects new AI use cases in cloud environments, code repositories and embedded systems and brings them into watsonx.governance. Arcjet finds AI from connections on managed macOS and Windows devices, attributed to a person, and from coding-agent hooks.
Can watsonx.governance run on premises?
Yes. IBM says you can deploy watsonx.governance in the cloud or on premises, prices self-managed software by virtual processor cores, offers plans on IBM Cloud and AWS, and has announced a FedRAMP-authorized version on AWS GovCloud. Arcjet runs as a cloud service, with its endpoint agent on managed devices.
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.