Your AI agents do not read your AI policy

Agentic workflows
Content

A written artificial intelligence (AI) policy governs people. Agents do not necessarily read it. An agent that has not been told a rule will arbitrarily break it, and the consequences arrive later. It shows up as a rejected release, a failed deployment, or a change nobody reviewed. We met this problem in our own engineering work. Our conventions were written down, so we built each one into the tools the work passes through. Today, changes in our gated code repositories are reviewed by AI agents, and the merge waits on the result. Risk rules still decide when a person must approve. This post describes four controls we run today, what each one replaced, and a test for agent controls that a leadership team can apply to its own organisation.

An AI agent is software that carries out a task on its own, using the same systems an employee would when applied in a business context. It might write code, draft a policy or open a change request. A guardrail is a control that limits what an agent can or cannot do. Until recently, people did the production work, so a written policy and a human reviewer were enough. That has abruptly changed. In our own trials, agents made prototyping five to seven times faster, but reviewing their output now takes more of our people’s time. The bottleneck moved from producing work to checking it. The commercial stakes sit on both sides. Without controls, a mistake can reach a customer, a regulator, or the brand before anyone sees it. With controls that depend on people reading every change, delivery is capped by the number of reviewers and the amount of time they can spend reviewing. The speed gain paid for by the organisation is lost. This matters to the executives who own the risk and the budget owners funding the tools. It matters equally to the security leaders who must account for every credential, and to the operations leaders whose teams now work beside agents. The four controls below share a common design: each makes the safe route the default route.

A written rule depends on someone having read it

If the safe route has to be the default route, the first question is why written rules do not hold. The answer is plain. A written rule works only when whoever does the work has read it and chooses to follow it. Our engineering rules were all documented. Commits must be signed, changes must be reviewed before they merge, and releases follow a set process. An agent that has not been given those rules may break them without even noticing.

The same pattern appears outside our walls. In an EY survey of 202 senior AI executives at large United States companies, 98% said they have an AI governance policy. Yet 47% admitted they had bypassed it for urgent deployments. A further 49% had not updated it to address AI agents. Even though the policies may have existed, they did not hold when the deadlines arrived.

So we stopped relying on the documented policies themselves, opting instead to integrate each policy’s rules into a shared set of tools that every agent loads the moment it does work. The principle behind those tools fits in one sentence. Agents propose, people approve, and automation carries out the change under sandboxed credentials that expire. That principle raises the next question. Who checks the proposal?

“Our agent control checklist turns this principle into ten questions.”

Every agent change waits for a review before it merges

The first control answers that question with a gate. In our gated code repositories, AI reviewers examine a proposed change, and the merge waits on the result. The gate does not replace human review. Risk rules decide when a person must approve, and some changes always go to a person, regardless of what the agentic reviewers say.

This matters because review is where the cost of agents now lands. In BairesDev’s survey of 705 developers in the third quarter of 2026, 67% said they review more AI-generated code than a year ago. Among the chief technology officers in that survey, 78% said they spend more on review, testing, and validation. If people must read every line an agent writes, the time saved in writing is spent in checking and reviewing.

A gate changes that arithmetic, because people read only what the rules send to them. In one recent case, the reviewer left findings on a set of changes, and the fixes were released the next day. A review covers what an agent wrote. It does not cover what the agent could reach while writing it.

An agent that never holds a credential cannot leak one

The second and third controls deal with that reach. They limit what an agent can touch and what it can publish.

The easy way to run an agent unattended is on an employee’s machine, under that employee’s credentials. The agent can then conveniently reach everything the employee can, for as long as it runs. We replaced that with a sandbox, a sealed and temporary workspace created for each specific task. Each session has a named owner, a time limit, and a record of what it did. The credentials, meaning the keys and passwords that grant the agents access to connected tools and services, stay outside the agent’s reach. They are injected into its requests on their way out. This is a common weak point. In a Gravitee survey of 900 technology leaders in October 2026, 45% relied mainly on shared accounts, borrowed human credentials, or fixed keys.

The third control applies the same idea beyond engineering. The agent that helps draft our human resources policies cannot publish. It can only open a draft, and the policy formally changes only when a person approves and publishes that draft.

All three controls check what an agent did or could do. None of them checks whether it actually behaved well while doing it.

“Who holds your agents’ credentials? Questions 3 to 5 of the agent control checklist cover access.”

Passing checks do not prove an agent behaves correctly

The fourth control exists because the first three can all pass while the agent still gets the work wrong. We learned this directly. One of our internal tools interviews project leaders and records what they say. Every automated check on its files passed. Meanwhile, three real defects reached a user. The faults lay in how the conversation went, and no file check could see any of the conversations.

So now we test behaviour separately. A suite of eight realistic scenarios exercises the tool and scores what it did. A full run costs about $11. Since then, about 13 fixes to our shared tools have come from tests of this kind.

Others report the same gap. VentureBeat surveyed 157 enterprise respondents in June 2026. Half had shipped an AI agent or feature that passed internal evaluations and then failed a customer. Only 5% said they fully trust automated evaluations. The sample was small and self-selected, so the figures are directional. Testing behaviour is not a guarantee; it is how we find the faults that every other check is unable to see. That leaves the question of what a leadership team should do with all this.

Ask what your agents are unable to do

The four controls point to one test that any executive can run. For each agent working in your organisation, ask what it is unable to do, and what enforces that limit. If the answer is a policy document, the limit depends on the agent having read it. If the answer is a control in the route the work takes, the limit holds.

Regulators are moving the same way. Singapore’s Infocomm Media Development Authority published a governance framework for agentic AI in January 2026. It recommends limiting the tools and permissions an agent has, requiring human approval at defined checkpoints, and testing before deployment. The framework is voluntary, which makes worked examples more useful than rules.

Controls carry a cost, and it is fair to say so. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027. CIO magazine reported the forecast in February 2026. Some will be cancelled because oversight by people erased the gains. That is the case for building controls into the route, so that people are tapped to review only what carries real risk.

Where to start

A policy tells people what should happen. A control built into the work decides what an agent is able to do. Our agents now work inside four controls. Their changes are reviewed, their credentials are managed and injected for them, their high-stakes work is limited to drafts, and their behaviour is tested. The aim is speed as much as safety. With the checks built into the route to production, agent work can ship at the pace it is produced. A named person stays accountable for the results. One decision remains open for us, which is exactly where a person must always approve, and we expect to keep adjusting it. To see where your own organisation stands, download the agent control checklist. It sets out ten questions your leadership team should be able to answer before your agents touch production.

Recommended articles