To govern AI across agents, copilots, and AI-generated code, an organization aligns to external rules such as the EU AI Act and runs five internal jobs, each with an owner. Leadership sets the principles, security inventories every AI system and controls data and access, process owners decide which actions need a human, and the teams running each workflow monitor and test it. Each job produces its own evidence, so programs that cover only two or three can look complete on paper while missing the places where incidents start.
What governing AI actually means
AI governance is the set of commitments, controls, and evidence an organization uses to decide which AI systems run, what they can touch, who answers for them, and how anyone proves those decisions afterward. It covers commercial models, internal agents, and the scripts an analyst generated last Tuesday.
Governance is the internal work. Regulation is the external floor. Externally, laws such as the EU AI Act and frameworks such as the NIST AI Risk Management Framework (AI RMF) set the targets.
Many organizations are still building the internal part. ISACA's 2026 AI Pulse Poll found that 38% of organizations have a "formal, comprehensive AI policy," up from 28% a year earlier. A policy also covers only one of the five jobs.
Breaking down AI governance
Governing AI breaks into five distinct jobs:
Setting the principles
Knowing what AI is running
Controlling data, access, and evidence
Keeping humans in the loop
Each job produces a different kind of evidence. Principles produce policy documents, inventory produces a register, access control produces logs, oversight produces approval records, and monitoring produces test results and drift data. Teams can audit coverage quickly by asking, for each job, which artifact they could hand over today.
1. Setting the principles
Five commitments set the terms every later control has to honor:
Accountability: A named person answers for each AI system's behavior.
Transparency: People know when AI is involved and what it is doing.
Fairness: Outputs are checked for systematic bias.
Privacy: Personal data is used only as permitted.
Explainability: A decision can be traced back to its inputs.
Principles have to be specific enough to settle a dispute. "We value privacy" settles nothing. "No customer personally identifiable information (PII) leaves the region or reaches a model that trains on it" lets an engineer turn down a request without escalating it.
The evidence for the principles job is the written policy and its review history. An executive steering group typically owns the principles job, with security and legal drafting the text.
2. Knowing what AI is running
Governance starts with finding every AI system in the environment, because each later control depends on the inventory. Much of what turns up is shadow AI (or “shady” AI — approved tools behaving in unexpected ways), tools and agents teams adopted without IT ever approving them. A Cloud Security Alliance (CSA) survey of 418 IT and security professionals found that 82% of organizations discovered previously unknown AI agents in the past year. Internal automation or scripting environments topped the list at 51%, and 40% of organizations found unknown agents in developer-created workflows.
Inventory now has to cover more than models and SaaS assistants. Developers in Sonar's 2026 State of Code survey reported that 42% of the code they commit is AI-generated or assisted, and 65.9% of Stack Overflow's 2026 respondents use AI coding assistants or agents at work.
Agent sprawl rarely starts with a bad decision. A team automates a lookup, an analyst asks a coding assistant for a script, someone wires an agent to a shared mailbox, and nobody registers the lookup, the script, or the agent anywhere.
Discovery is half of the inventory job. The other half is risk tiering, which sorts the discovered systems by what each one can reach and decide. An agent that summarizes public documentation and an agent that can change permissions in an identity provider require different levels of review.
The evidence is a current register with owners and tiers attached. A sensible split has security running discovery and each business owner keeping their entries accurate.
3. Controlling data, access, and evidence
Access control decides what an AI system can see, what it can send out, and what it can change. In Cyberhaven's endpoint telemetry, 39.7% of AI interactions involve sensitive data, which includes prompt text, pasted content, and file uploads. Data classification sets the rules, and egress limits enforce what leaves.
Access is where agents differ most from chat tools. Of the organizations in a CSA survey, 53% have had AI agents exceed their intended permissions. OWASP's guidance on excessive agency is to grant each extension only the permissions it needs. The same guidance tells teams to "implement authorization in downstream systems rather than relying on an LLM to decide if an action is allowed or not."
Take a workflow that a webhook triggers when Okta flags a risky sign-in. An AI agent summarizes the context, and the workflow opens a ticket in ServiceNow.
In Tines 3B, the complete environment for enterprise AI, a credential proxy injects the Okta and ServiceNow secrets into outbound requests at runtime, so workflow code, builders, and the AI never hold them. Role-based access control (RBAC) scoped per space, connector, and network sets who can change the workflow, and audit logs record every user and system action.
Those logs are the evidence for the access control job. When someone asks after an incident who could reach what, the audit log is the record that answers it. Security and identity teams are the natural owners here.
4. Keeping humans in the loop
Human review is the right control for actions that are irreversible, high-impact, or ambiguous, such as disabling an account, changing production, or sending something outside the company. It isn’t necessarily the right control for everything else, because people who approve at volume may stop reading closely.
Approval screens can also be manipulated. CSO Online described an AI coding tool whose approval dialog asked a person to approve an edit to a config file while the tool was writing to SSH keys. A human sees only what the tool's developers designed the dialog to show. Similar weak controls surface in other real AI governance failures.
Automated policy enforcement takes over where human review is too slow to keep pace. In the CSA survey cited earlier, when agents exceeded their scope, only 11% of organizations automatically blocked the action. Most relied on approval (38%) or logging (24%). Gartner analyst Shiva Varma said, "Enterprises are treating AI agent governance as binary, either locked down or fully trusted, and that is the root cause of failure."
The workable pattern uses the full spectrum of execution, with three tiers:
Deterministic: Rules block actions that are never allowed, with no model deciding.
Agentic: Agents act alone on low-risk tasks, within tightly scoped permissions.
Human-in-the-loop: A person approves the small set of decisions with real consequences.
The evidence is the approval record together with the policy that routed each action to it. Process owners typically decide which actions need a human, and security writes the blocking rules.
5. Monitoring and testing over time
AI systems change after deployment, even when no one edits them. In a 2025 postmortem, Anthropic reported that three infrastructure bugs intermittently degraded Claude's response quality between August and early September, and that early user reports were hard to distinguish from normal variation. Because nothing throws an error, drift tends to show up as a workflow that runs on schedule and returns worse answers.
Testing has to happen both before and after release. Stanford HAI's 2026 AI Index found hallucination rates across 26 top models ranging from 22% to 94% on a new accuracy benchmark. Monitoring covers four practices.
Bias testing on outputs that affect people
Hallucination checks against known answers
Red-teaming for prompt injection and misuse
Drift tracking against a baseline after each model or prompt change
Take a cron-scheduled workflow that pulls CrowdStrike detections and asks an agent to classify them. Through Tines 3B, teams monitor that workflow on its own, with a dashboard covering spend, performance, and value. When a run fails, Autofix diagnoses it and proposes a tested fix on an isolated branch, and the team decides whether the fix goes live.
The evidence for the monitoring job is test results, monitoring history, and the change log. The team that runs the workflow owns monitoring, with security reviewing.
Where regulation and frameworks fit
Regulations and frameworks are alignment targets for the five jobs. The table below maps each job to its evidence and to the framework hooks that ask for it.
The EU AI Act's high-risk obligations now apply from December 2, 2027 for Annex III systems and from August 2, 2028 for AI embedded in regulated products, after the Digital Omnibus pushed back the original dates.
The NIST AI RMF organizes the work into Govern, Map, Measure, and Manage, and treats Govern as cross-cutting. ISO/IEC 42001 turns the jobs into a certifiable management system with an Annex A control set.
Enabling innovation without losing control
Blanket bans can feel like the safest choice, and the instinct behind them is reasonable when data is leaving through consumer chat apps. When leaders ban AI tools, the organization can lose visibility while employees continue using those tools. Gartner found that 69% of organizations suspect or have evidence that employees use prohibited public GenAI. In a TrustedTech survey, nearly one-third of employees said they would keep using AI tools even if rules prohibited them.
Teams govern AI more effectively through a sanctioned path that brings most of the five jobs into the same environment. Inventory is easier to keep current when agents and automations run in a known environment. Access and evidence come from the runtime instead of a later reconstruction. Monitoring stays attached to each workflow.
The governed path also has to be faster than going around it. In an EY survey of senior AI decision-makers at US companies with $1B or more in annual revenue, 47% said their organization had skipped its AI governance process for urgent deployments.
A governed environment doesn't write the principles or decide which actions need a human, so those two jobs stay with the organization's own teams. It does make the other three hard to skip. Tines 3B is that kind of environment for AI-generated code, agents, and automations. A credential proxy keeps secrets out of workflow code, changes are tested on an isolated branch before going live, audit logs record every user and system action, and each workflow is monitored on its own.
Teams can map the five jobs onto their own workflows in the free version of the platform, Tines 3B Explore Edition.
