AI governance must move from static policies to living guardrails that evolve in real time, flag drift, and keep work safe.
Sydney Scott
Editorial Strategist, AI
Workday
A few years ago, an AI policy could live in a PDF waiting for annual committee reviews. That slow approach worked well enough when AI just meant basic chatbots answering simple questions. Today, that static setup doesn’t work.
A rule written months ago cannot manage software that changes every week. Vague corporate paperwork will not stop an active agent from pulling a bad file or sending secret data to the wrong person. The old way relied on workers remembering a list of rules. The new approach builds those rules directly into the system.
This shift replaces old policy documents with active, living guardrails embedded right into enterprise software. These automated controls work like digital speed checks and emergency brakes. They watch the software in real time to limit what an agent can see, trigger human approval for risky steps, or stop dangerous actions completely. Continuous safety becomes a part of the daily flow of work, rather than a yearly event.
The old way relied on workers remembering a list of rules. The new approach builds those rules directly into the system.
Report
Some companies still treat AI like a basic desktop app: pitch an idea, drag it through legal reviews, get a signature, and pray the code behaves. That worked when AI was just a tool people used. It does not work when AI becomes an agent that can take action.
Agents do not just write answers. They can read data, start workflows, update records, draft messages, and move work across systems. That means policy has to do more than say what people should do. It has to control what agents are allowed to do.
A chatbot may need rules for tone, accuracy, and disclosure, but an agent needs rules for access, approval, escalation, and accountability.
As Workday VP, AI Platform Dean Arnold explains, “Organizations wouldn’t hire thousands of employees without an HR system to manage them. The same discipline is now required for AI agents.”
McKee Foods is already putting this idea into practice. The company gives each agent its own digital identity and credentials. That allows teams to verify which agent is acting and limit what it can access across systems. For Floyd Walterhouse, the company’s vice president and CIO, this is a necessary response to agents acting with delegated authority. The rule does not live in a binder telling the agent to stay in its lane. The lane is built into the system.
That is the new job of AI policy. It cannot sit on a shelf. It has to run inside the work itself.
Living guardrails are active controls that keep AI agents useful, safe, and accountable after launch.
A basic guardrail says: the system must follow this rule. A living guardrail goes further: the rule can be enforced, measured, tested, and improved as the agent operates. That is what makes it “living.” It learns from real behavior, catches drift, and adapts when workflows, data, users, or risks change.
And without necessary guardrails, they’re what Workday CTO Gabe Monroy refers to as “lawless agents.”
“They’re omniscient,” he says. “They can see all the things, they can touch and do all the things through lots of tools, but they're fundamentally lawless.”
Living guardrails create governance.
Some guardrails are built before an agent goes live. They define the agent’s boundaries: what data it can access, which tools it can use, what it must never do, and when a human approver is required.
Other guardrails are generated in the moment. These runtime controls check what is happening while the agent is in use. They can verify permissions, detect sensitive data, block a risky action, or route a decision to a manager before the agent moves forward.
Then comes the feedback layer: monitoring. This is what keeps governance alive and agile. Monitoring watches what agents actually do, compares behavior against policy, creates evidence for audits, and flags drift. Drift is when the agent slowly moves away from what was approved because the data, process, or environment changed.
A launch checklist can tell you an agent was ready on day one. It cannot prove the agent is still behaving as intended on day 100. The National Institute of Standards and Technology (NIST) and the EU AI Act are pointing towards this shift: AI risk management does not stop at deployment. Regulators are making the signal clear. “We checked it once” will not be enough.
Living guardrails are active controls that keep AI agents useful, safe, and accountable after launch.
Guardrails only work if they keep up with the agent. That becomes harder as AI systems expand. Prompts change. Models get upgraded. New data sources come online. Tools get added. Each change can alter what an agent can see, decide, or do.
A payroll assistant that answers policy questions carries one level of risk. A payroll assistant that can change bank details carries another. The controls should change with the stakes.
Workday CIO Rani Johnson puts it well: “When it’s time to do something deterministic, where you’re doing a mission critical service, that has to absolutely guarantee the right result. You can’t YOLO that type of work.”
Low-risk exploration and high-consequence action cannot be managed the same way. That is why testing has to become routine. Before an agent gets more authority, it should be challenged with real scenarios. Can it spot sensitive data? Does it respect access rules? Does it know when to escalate? What happens when someone asks AI to break a rule?
Those checks should run again when the agent changes. A new prompt, model, data source, or tool should trigger another look. Otherwise, leaders may be governing an older version of the system.
Certification turns that evidence into a decision. It defines the lane an agent is allowed to operate in. Monitoring keeps the lane visible. Important actions should show who asked, what data was used, what the agent decided, which tool it called, and whether a person approved it.
That trail helps with audits, but it also helps teams learn. Governance gets stronger when it produces evidence, feedback, and sharper decisions.
Guardrails only work when they fit the way people actually work.
That is the lesson leaders are learning fast. Vanessa Candela, chief legal and trust officer at Celonis, puts it plainly: “We can’t sit around and wait for the regulators to sort of figure it out,” she says. “We’re embracing it, we’re putting common sense guardrails around it and just pushing forward.”
Celonis backs that mindset with a cross-functional AI governance council that brings legal, engineering, product, security, compliance, and ethics together early. The point is to make governance a runway, not a roadblock.
And the model has to be shared. Central teams set the baseline: Security sees the threats, Legal sees the duties, and Risk sees the controls. But, Business teams shape the rules for each workflow because they know where work gets messy. They know which steps matter, which checks slow people down, and where a human needs to step in.
This is also why Workday Chief Responsible AI Officer Kelly Trindel argues responsible AI has become a core design discipline. As agents begin to plan and act for a business, leaders need guardrails built in from the start, not added after trust is already at risk.
When it’s time to do something deterministic, where you’re doing a mission critical service, that has to absolutely guarantee the right result. You can’t YOLO that type of work
Rani Johnson
CIO, Workday
AI doesn’t wait for the next review cycle. Neither can compliance.
Good governance has to show up in the messy middle of real work—when a model changes, when a team adds a new data source, or when a workflow expands from a pilot to daily use. When this happens, teams know where the boundaries are. Testing becomes part of shipping. Monitoring becomes part of operating. Business leaders get more confidence because the rules are visible, current, and enforceable.
But to get here, leaders need to stop asking “did we approve this use case?” and instead ask “can we see if it’s still safe, useful, and compliant right now?”
Governance isn’t the brake on AI. It is what lets teams move with speed and control. One-time approvals helped companies get started. Living guardrails are what help them scale.
Report
A gap between effort and impact is the real ROI problem. Discover the solution.