AI governance programmes fail not when a rule is broken, but when a decision arrives and nobody can move without asking permission first.
The boundary exists. It has been written, reviewed, and given an owner. In some organisations it has even been enforced before. On paper, AI governance is present. But at the moment an employee needs to decide whether to use an AI tool for a client task, whether to connect a new platform to company data, whether a particular use case is within the spirit of the policy, the boundary offers no answer. Not because it is wrong, but because it was written for the policy document, not for the decision in front of the person.
This is where governance friction actually begins. Not in enforcement. In indecision at the moment judgment is required.
The Boundary Exists. The Intent Doesn't Travel With It.
An AI acceptable use policy is the most common starting point for enterprise risk management around AI tools. Most mid-market organisations now have one, or are in the process of writing one. It lists approved tools, prohibited categories, data handling requirements, and escalation procedures.
What it rarely carries is the reasoning behind each boundary. Why is this tool approved and that one not? What data sensitivity threshold triggers the escalation path? If a tool is not on the approved list but solves a genuine business problem, is the default answer no, or is it start a conversation with IT?
Without that context, the boundary becomes a document that employees consult and then put down, still uncertain. A policy that requires interpretation at every point of application has not reduced the cognitive load of deciding, it has transferred that load from the governance programme to the individual, and then hoped the individual makes the right call.
At low volumes, this works. Individual decisions get made, questions get asked, exceptions get approved. But as AI tool adoption accelerates, and it is accelerating faster than most governance programmes anticipated, the volume of decisions that require interpretation outgrows the capacity of the people being asked to interpret. The policy was written for a steady state that no longer exists.
Why Decisions Escalate Even When Nothing is Broken
Escalation is often read as risk aversion or excessive caution. In practice it is what happens when boundaries fail to do their job.
When intent is not visible at the point of decision, responsibility shifts upward. Teams do not escalate because they lack confidence in their judgement. They escalate because the governance system gives them no clear signal that deciding locally is safe. So the decision moves up, not because it is exceptional, but because it is ambiguous.
Over time this creates a quiet inversion. CISOs and IT Directors become the default mechanism for resolving uncertainty about AI tool usage, not because they want to be the bottleneck, but because no boundary clearly says: this is within intent, proceed. Every week there are more AI tools, more edge cases, more employees who have found something useful and want to know if they can use it. The governance programme was not designed to process that volume of individual decisions. But it has become responsible for all of them by default.
The cost is not just the CISO's time, though that is real. It is decision latency across the organisation. Enterprise risk management frameworks are designed to reduce exposure, not to function as the primary decision engine for routine tool adoption questions. When they start doing the latter, both functions suffer.
The cost never announces itself. A CISO who gets asked whether a new AI writing tool is permitted does not think of that conversation as a governance failure. It sounds reasonable. But the same question comes in from a second team two days later and then from a third team. Each conversation is short and none of it feels like a problem. But run that pattern across a quarter and the entire picture changes. What looked like a mature governance process from the outside, now from the inside, is a queue. And the queue grows faster than anyone expected, because the boundary was written for the policy document, not for the decision in front of the person asking. - Ankur Arora, Co-Founder, Montro |
How Ambiguous Boundaries Quietly Centralise Authority
As more AI governance decisions require interpretation, authority concentrates, not through policy, but through necessity.
When boundaries do not carry intent forward, organisations compensate by leaning on the individuals who can explain the history, assess the nuance, or arbitrate the trade-off. These people; usually the CISO, the DPO, or a senior IT lead, become informal choke points. Not because they seek control, but because the governance system keeps routing decisions to them.
This centralisation feels temporary at first. A reasonable response to a period of rapid change. But the period of rapid change in AI tool adoption is not temporary. The number of AI tools available to employees grew by 393% in a single year according to internal discovery data from enterprise audits . The centralised arbitration model does not scale to that rate of change.
What starts as sensible oversight gradually reshapes how the organisation operates. Decisions that could have been distributed become escalated by default. Autonomy narrows. Security governance stops being a framework that enables safe movement and becomes a queue that movement has to pass through.
Shadow AI is What Happens When Governance Loses the Employee
When AI governance boundaries are ambiguous, employees make a predictable calculation: asking permission takes time, the tool solves a real problem, and the risk of asking and being told no is higher than the risk of quietly using the tool without raising the question.
This is not recklessness. It is a rational response to a governance system that has made the cost of compliance higher than the perceived cost of non-compliance. The employee does not think of it as shadow AI. They think of it as getting their work done.
Harmonic Security's 2025 Annual AI Usage Index, drawing on over 22 million enterprise prompts, identified 665 distinct AI tools in active use across enterprise environments, the vast majority without central IT oversight. A Fishbowl survey of over 11,700 professionals found that 70% of those using generative AI at work had not disclosed that usage to their employer. These are not edge cases. They are the aggregate outcome of governance programmes that made asking permission feel harder than finding a workaround.
The AI compliance consequence is direct. An organisation whose employees have concluded that the governance system is not worth engaging with cannot build a meaningful AI inventory, cannot satisfy GDPR Article 30 processing records, and cannot demonstrate the human oversight required by EU AI Act Article 26 for deployer obligations. The governance programme exists. The employees have simply routed around it.
What Boundaries that Carry Intent Actually Look Like
Governance without friction does not mean fewer constraints. It means fewer moments where a legitimate decision stalls because the intent behind the constraint is unclear.
The simplest version of this is making the reasoning visible at the point where the boundary is encountered. Not just what is permitted or prohibited, but why the line is drawn where it is. An employee who understands the reasoning behind a boundary can apply it to situations the policy document did not anticipate. An employee who only knows the rule cannot, and so they either escalate or they guess.
Beyond that, the consequence of crossing a boundary needs to be proportionate and legible. A governance system where the penalty for every infraction is the same, regardless of whether someone connected a low-risk writing tool to their work email or uploaded customer data to an unvetted AI platform, teaches employees that the system cannot distinguish between severity levels. They learn to treat all risk as equivalent, which means they treat all governance as equally burdensome.
And when a situation falls outside what the boundary covers, the path forward needs to be defined rather than improvised. The current default in most organisations is some version of 'ask your manager, who asks IT, who schedules a review.' That process is not designed to handle the volume of edge cases that modern AI adoption generates. The organisations that have reduced governance friction have replaced that improvised path with a defined one, a lightweight intake process, a risk classification that routes automatically, a decision that happens in days rather than weeks.
For AI governance specifically, this means moving beyond approved lists. An approved list tells an employee which tools are sanctioned today. It does not tell them how to think about a tool that is not on the list, what data categories trigger which level of review, or what the organisation's actual risk posture is. A well-designed ai governance framework answers those questions at the moment of decision, not after the employee has already made the call.
Compliance automation is where this becomes operational. When governance decisions are embedded into workflows, an AI tool request that routes to a risk assessment, a data sensitivity flag that triggers the right approval path, an offboarding process that revokes AI tool access automatically, the boundary becomes part of how work happens. Not a document employees consult. A process they move through.
Why Adding More Rules Makes this Worse, Not Better
When AI governance friction becomes visible, when CISOs start tracking the hours spent on ad hoc tool approval questions, when shadow AI audits reveal that employees have built entire workflows around unapproved tools, when a regulatory review exposes gaps between the written policy and actual practice, the instinct is to tighten governance.
More rules. More checkpoints. More specificity in the policy. A larger list of prohibited tools.
These interventions feel corrective. They rarely address the actual problem. More rules without embedded intent simply create more moments where interpretation is required. Each new boundary is another surface where the intent can fail to travel. Each new checkpoint is another point where a legitimate decision can stall. The governance programme becomes denser without becoming more useful.
The organisations that have made AI governance work in 2025 did not get there by writing better policies. They got there by treating governance as a design problem rather than a documentation problem. The boundary is encountered at the point of action. It carries enough context to guide the decision. It has a defined path for the cases it does not cover. The policy document is the output of that design process, not the process itself.
Governance does not slow teams down, poorly designed boundaries do. The bill is rarely paid when the rule is written. It is paid six months later, when a discovery audit identifies an AI tool that has been running on customer data since the boundary failed to provide the employee with a clear answer at the moment they needed one. That is when the poorly designed boundary presents its bill - not as a policy violation. As a gap that has to be explained to a regulator. - Ankur Arora, Co-founder, Montro |
Frequently Asked Questions
Why do AI governance policies create friction even when they are well-written?
Because a well-written policy and a boundary that carries intent are not the same thing. A policy document tells employees what is permitted or prohibited at the time it was written. It does not tell them how to think about situations the policy did not anticipate, what the reasoning behind each boundary is, or what to do when their specific situation sits in the gap between two rules. At the moment of decision, those gaps produce friction, not because the policy is wrong, but because it was designed to be read, not to guide judgment in real time.
What is shadow AI and why does it happen in organisations with governance programmes?
Shadow AI is the use of AI tools outside an organisation's sanctioned inventory, tools that employees adopted without going through IT review, that AI features activated inside existing SaaS tools without procurement decisions, or that arrived through personal accounts. It happens in organisations with governance programmes when the cost of compliance; asking permission, waiting for approval, risking a no, feels higher than the perceived cost of quietly using the tool. The governance programme did not fail through enforcement. It failed at the point of adoption by making the sanctioned path harder than the unsanctioned one.
How does poor AI governance create enterprise risk management exposure?
Poor AI governance creates exposure in two directions simultaneously. Operationally: decisions that should happen at team level escalate upward, CISO and IT Director time is consumed by routine tool approval questions, and the organisation moves slower than its competitors on AI adoption. Regulatorily: an AI inventory that is incomplete because employees have routed around the governance programme cannot satisfy GDPR Article 30 record-keeping requirements, DORA Article 28.3 ICT register obligations, or EU AI Act Article 26 deployer oversight requirements. The enterprise risk management exposure compounds poor governance produces both an operational drag and a compliance gap simultaneously.
What does compliance automation do for AI governance?
It moves the governance decision from a document employees have to remember to consult into the workflow they are already using. When an AI tool request automatically routes to a risk classification, the employee does not need to know the policy, the process applies it for them. When a data sensitivity flag triggers the right approval path rather than a generic 'ask IT' instruction, the decision happens faster and more consistently than any manual process could produce. When offboarding automatically revokes access across all AI tools rather than relying on IT tickets that get missed, the governance boundary enforces itself. The intent travels with the process. It does not depend on the employee's interpretation of a document they may not have read recently.
How do you build AI governance that scales with the rate of AI tool adoption?
Stop treating governance as a list of approved tools and start treating it as a risk classification framework. An approved list tells employees what is sanctioned today, it does not scale when the number of AI tools available doubles every year. A risk classification framework tells employees how to think about any tool, regardless of whether it has been reviewed before: what data categories it can and cannot process, what level of review a given use case requires, and what the defined path is when a tool does not fit neatly into existing categories. That framework scales. A list does not.
