A field guide for IT and security
AI governance.
From policy to practice.
Decide who can use AI, with which data, and through which providers. Then make those decisions part of how work gets done.
An operating model, not a document
What is enterprise AI governance?
AI governance is the combination of accountability, policy, and controls for using AI responsibly. For employee-facing tools, that means reviewing providers, classifying data, granting access, checking outputs, and knowing how to respond when something goes wrong.
Enterprise AI governance connects those decisions across IT, security, legal, procurement, and business teams. A policy without a usable approved tool can leave shadow AI untouched. A tool without an owner leaves important decisions to chance.
This guide focuses on employee tools and agents. Developing or deploying your own AI models may also require model-risk assessment, evaluation, and other obligations beyond this scope.
Your AI usage policy
Answer the questions employees actually have.
Use these prompts as a working checklist, not a legal template. Have your security and legal owners adapt the policy to your data, contracts, and obligations.
- Which tools can I use?
- List approved clients, accounts, models, and providers. Explain how to request a new tool and who reviews it.
- What data can I include?
- Define permitted data classes and prohibited material. Cover prompts, file attachments, agent context, connected services, and secrets.
- Who gets access?
- Name the access owner, the approval route, and how access changes when employees move teams or leave.
- What needs human review?
- Set review expectations for AI output, code, customer-facing work, and actions an agent takes on someone's behalf.
- What happens when things go wrong?
- Give employees an incident reporting channel, an exception process, and a clear owner for investigation.
- When do we revisit this?
- Choose a review cadence and update triggers, including new providers, tools, data categories, and contractual requirements.
Evaluate the platform, not the promise
What to look for in an AI governance platform.
Start with the controls your policy needs. Ask vendors to demonstrate each one with a real account and a non-sensitive task. Separate shipped behavior from roadmap commitments.
- Identity and offboarding. Test company SSO, provisioning, and access removal—not just the sign-in screen.
- Model and provider control. Verify permissions and provider choice, such as AWS Bedrock, for each client you intend to approve.
- Visibility with a defined scope. Understand what usage is recorded, who can see it, and what traffic is outside the platform.
- A clear data boundary. Trace prompts, file context, credentials, logs, and connected-service traffic. Review retention and processing terms separately.
- A practical employee experience. Check installation, onboarding, familiar tool support, and a clear path for exceptions.
Where OpenWork fits
A governed path for everyday AI work.
OpenWork combines a local-first desktop workspace with organization-managed model access. The AI gateway provides company sign-in, permissions, provider choice, and visibility into routed model usage.
OpenWork and OpenCode with the OpenWork plugin support company AI gateway models. Claude Code, Codex, and Cursor can use the MCP gateway for shared skills and connections; AI gateway model integration for those clients is in progress. Connecting MCP alone does not route model traffic.
Files stay on the user's PC in local desktop mode, but prompts and file context sent to a remote model leave the device. Analytics, diagnostics, updates, connected services, and optional cloud workers have separate flows. OpenWork is not a network-wide discovery or data-loss prevention tool.
Start small, review, then expand
Make secure AI for employees a repeatable practice.
Discover existing use with team interviews, software inventories, and permitted endpoint or network signals. Set policy before connecting sensitive data. Pilot a sanctioned path with a small group, and test access, provider routing, output review, and offboarding.
Monitor adoption and exceptions, not just spend. Usage through a gateway tells you about that path; it does not prove nobody is using another tool. Keep your broader discovery and incident processes in place.
Use the four-step shadow AI frameworkAI governance FAQ
What is AI governance?
AI governance is the set of owners, policies, reviews, and technical controls that determines how an organization uses AI. For employee tools, it covers approved providers, permitted data, access, human review, usage visibility, and incident handling. It is an ongoing operating practice, not a certification or a single product.
What does an AI governance platform do?
An AI governance platform can centralize parts of the operating model, such as identity, permissions, approved model providers, and usage records. Evaluate the scope of each product: model access controls are different from network-wide discovery, data-loss prevention, model-risk assessments, and regulatory reporting. OpenWork focuses on governed access for supported agents and model traffic routed through its gateway.
How is enterprise AI governance different from an AI usage policy?
The AI usage policy records what is allowed and who is accountable. Enterprise AI governance makes that policy operational through onboarding, provider review, access controls, employee training, exception handling, and recurring reviews. Both need named owners and an escalation path.
How can we provide secure AI for employees?
Start by reviewing the work employees want to do and the data involved. Offer approved providers and tools, company sign-in, least-privilege access, clear data rules, and output review. Test the data flows, access removal, and incident process in a pilot. No platform can guarantee security independently of its configuration, providers, and employee practices.
Does local-first mean no company data reaches a model?
No. The local desktop keeps files and workspace processing on the user's PC, but prompts and file context supplied to a remote model leave the device. In the model-gateway setup, workspace processing sends SSO configuration and model traffic rather than a workspace copy. Analytics, diagnostics, updates, connected services, and optional cloud workers have separate data flows.
Is an AI gateway a complete enterprise AI governance solution?
No. A gateway can control supported model traffic routed through it. Pair it with a policy, provider and legal review, training, access reviews, and endpoint or network controls where needed. OpenWork does not automatically discover or block every AI tool outside the approved path.
From policy to practice