Every firm has an AI policy. Few can prove anyone follows it

Picture an ordinary Tuesday afternoon at a UK professional services firm. A deadline is close, and an employee pastes a client contract into a free AI chatbot to summarise it faster. There is no malice and no alarm. The firm has an AI policy, approved by the board and filed on the intranet. That policy has just been breached, and nothing stopped it.

The scene is hypothetical. The behaviour behind it is not. Microsoft found that 71% of UK employees surveyed had used unapproved consumer AI tools for work, and 51% continued to do so every week. Okta’s research highlights the confidence gap: 96% of UK executives said they were confident in their organisation’s visibility into AI tools, while 55% of UK knowledge workers reported using unapproved AI. These are self-reported surveys, not audits, but they point to a substantial gap between executive confidence and everyday practice.

Research from the Ponemon Institute and DTEX reinforces the concern. It found that 73% of organisations worry unauthorised AI use is creating invisible routes for data to leave the business, yet only 18% have fully integrated AI governance into their insider risk programmes. Worry is widespread. Preparation is not.

The government’s own figures tell a similar story. The Cyber Security Breaches Survey 2025/26 asked about AI for the first time. Among businesses using, adopting or actively considering AI, only 24% reported having cyber security practices or processes to manage its risks. Almost a third had no plans to introduce them.

The legal obligations already exist. An employee disclosing personal data to an unapproved AI service can cause a personal data breach just as a cyber attack can. Under UK GDPR, organisations must notify the ICO without undue delay and within 72 hours of becoming aware of a breach, unless it is unlikely to pose a risk to individuals’ rights and freedoms. That duty depends on recognising what has happened. When AI tools are used out of sight, firms may struggle to identify the disclosure, assess the risk or respond promptly.

A policy is not a control

A policy is a statement of intent. It tells people what they should do and establishes the standards an organisation expects. On its own, it does not step in when an employee, an AI agent or an automated integration tries to do something it shouldn’t. Visibility helps organisations identify problems. Enforcement determines whether an action is allowed to proceed.

The problem is also one of complexity. An organisation may need to apply internal policies alongside client restrictions, contractual obligations and legal requirements across several jurisdictions. The answer can depend on who is acting, what data they are using, which matter they are working on and what they are authorised to do. Nobody can reliably hold all of that in their head at the moment they act. That is not a criticism of staff. It is a design flaw in leaving enforcement to memory and goodwill.

Where this is heading

Four pressures are likely to make that gap harder to ignore.

The first is evidence. Having a policy and demonstrating that it governed a particular action are different things. If a client, auditor or regulator asks which rule applied, what decision was made and whether it took effect before data was shared, a policy document alone cannot answer. Existing logs may help reconstruct events, but reliable evidence is much easier to produce when the policy decision is recorded as it happens.

The second is contractual. As clients introduce AI clauses into their contracts, one careless paste could breach a confidentiality obligation or an agreed restriction on AI use, even where no personal data is involved. The consequence may be the loss of a client relationship that took years to build.

The third is speed. The risk already extends beyond a person pasting text into a chatbot. An agent acting on delegated authority can attempt a payment beyond its limit, retrieve information outside its remit or move sensitive data to an unapproved system, faster than a human can review each step. Controls need to operate before those actions take effect.

The fourth is cost. As AI becomes embedded in workflows, consumption can accumulate across tools, models and repeated agent actions. A budget written into a policy will not stop an automated process exceeding it. Organisations need a way to attribute spending and apply limits while the work is happening.

Enforcement at the point of action

None of this is an argument for banning AI. Staff use these tools because they work, and restrictions without practical alternatives can push that use out of sight. It is an argument for moving enforcement to the point of action, which means three things.

First, translate the documents an organisation already has, from client contract clauses to internal guidelines, into practical rules, reviewed and validated by the people responsible for them.

Second, apply those rules across human use, AI agents and automated workflows. That means governing what enters a model, how its output is handled and what actions a connected system is authorised to take, including which data it can access and how much it can spend.

Third, record each policy decision as it is made, including the rule applied, the outcome and the time, so the evidence exists when someone asks for it.

That is the approach we have taken at Ephiria, but the principle matters more than the product. Britain has spent the past two years writing AI policies. The next two will be judged on whether anyone can show they worked.

Featured Photo by Vitaly Gariev on Unsplash.