From Policy to Practice: What It Takes to Keep AI Governance Working

Ask a room of business leaders who uses AI and around two thirds of the hands go up. Alasdair Walker, Managing Director of Optimum Path Financial Planning, asks that question whenever he speaks, and the number grows every time. To the rest he says: "Trust me, there are people using AI in your business."

That was the starting point for Beyond the Framework, our webinar on keeping AI governance working once the policy's written. Alasdair joined our CTO James Ridgway and Sam Newton, co-founder of Sheffield firm Gravitate Accounting. Both run businesses built on sensitive client data, and both already use AI.

Banning AI doesn't stop people using it, the panel agreed, and a policy left alone for a year goes stale. The governance that works gets revisited every few months and matches its controls to the risk of each task. It also gives people somewhere safe to start.

Banning AI doesn't stop people using it

No business gets to opt out. Once you have more than two or three people, Alasdair says, someone is asking a chatbot to sort their diary or putting work questions into the free version of ChatGPT. Even a business that doesn't officially use AI needs a policy that says so.

James calls this shadow AI, the same problem as shadow IT, where people adopt tools the business never chose. At Gravitate, client data is banned from AI tools without the right process, and Sam still hears the odd remark about running an email through one.

James thinks a blanket ban leaves you more exposed than a safe, sanctioned way to use AI. We've worked with organisations where staff were banned from certain tools and carried on using them on personal accounts, often responsibly. Policy has to be grounded in what's sensible, or it becomes a blocker to delivering value.

The second problem is time. When Alasdair reviewed his firm's AI policy, it was about a year old and no longer reflected how the team used AI at all. Most policies get an annual review. The panel's view was that AI needs one every three to six months, because the tools change and the risk changes with them.

AI also arrives without being invited. Sam pointed out that partners such as Xero are building it into software his team already uses. Our own policy had moved on too. When we reviewed it around August and September, it allowed a range of tools while we worked out what we wanted. We've since decided to narrow that to an approved set.

If the blank page is the hard part, Alasdair's tip is to let AI help. Tell it where you work, say you need an AI policy, and ask it to interview you to make the answer better.

Give people a safe place to start

Both firms started small, with a safe tool and a quick win. Alasdair's firm already worked in Microsoft 365, so it gave everyone a Copilot licence as a sandbox. Copilot could already reach client data through Outlook and SharePoint, so it added no new exposure.

Their IT provider's warning is worth passing on. Check who can see what before you switch Copilot on. Set it up wrong and it will surface anything to anyone who asks, salary spreadsheet included.

The quick win was meeting recording. Note-taking in client planning meetings used to vary between planners, and Alasdair reckons he was the worst. Now he listens and looks clients in the eye, and the notes are better than any he wrote.

Gravitate also chose Copilot, largely because security was people's biggest fear. The directors tested AI themselves first. At the annual strategy day, the team had a training session, then small groups came up with ideas for their own roles before signing off the policy.

James's experience is that around 95% of people want to use these tools responsibly. They'd rather help shape the rules than have them handed down. Training reassures people they're doing it right, and working groups surface what works so it can feed into the policy. We test new uses on our own internal projects first, where a mistake makes a mess but hurts nobody.

Not every use carries the same risk, either. Tidying an email or organising a calendar exposes far less client information than doing the core work with AI. Some of our legal contacts give people wider access for productivity and keep tighter control, with proper training, on casework.

Let AI check its work, but keep a person in the loop

Alasdair's firm uses Claude for many internal processes and increasingly asks it to audit its own work. With instant access to the regulator's compliance manuals, it has a strong working knowledge of the rules. A person still reviews the result. "We have the AI checking itself, which is potentially dangerous, but we've always got a human in that loop to check the checking."

The checker needs to know what good looks like. Alasdair pointed to the legal cases where made-up case law found its way into real work. Once an invented case is cited in genuine case law, the next AI tool to search for it will treat it as real. He calls it AI eating itself.

Skills matter here too. Sam trained before Xero and Sage, so he learnt what each transaction meant. People now click buttons without that understanding, and leaning too hard on AI could make it worse. One law firm mentioned in the session holds AI back from legal work until people are more senior, and Gravitate is applying similar logic. James's concern is that newly qualified people won't have the hands-on experience to spot when work is heading down the wrong path.

None of this is a case against AI. James's answer to "but AI isn't perfect" is a question: "Are you telling me you run an organisation with perfect people?" Given guardrails, AI can catch things people miss or tire of. A blend of the two beats ruling it out because it's new.

Choose your risk deliberately

Alasdair thinks many leaders are worrying about the wrong thing. He remembers businesses being told email wasn't secure, often by people selling an alternative, and hears the same pitch made to SMEs about AI now. His view is to accept "a level of reasonable risk" in return for frontier tools that can change a business.

His firm tells clients how and where their data is processed, mitigates the risks, and uses models such as Claude, Gemini and GPT on commercial terms. As he reads it, his main regulator, the FCA, is less concerned with the AI itself because the Senior Managers regime already holds named people accountable.

He offered a motorway as a way to think about it. Some drivers sit at 55mph behind a lorry. Some do 95mph in the outside lane and brake for every speed camera. The middle lane accepts some risk to move faster. No lane is wrong in itself, but the choice should be deliberate. The real danger is believing you don't do AI while people in your business are doing 120mph.

James takes the same pragmatic line. Compliance too often becomes something people do because they've been told to. He'd rather start from first principles: what are the risks, what controls can you put in place, and are you happy with them?

Sam's version is knowing your limits. When Gravitate built a tax planning tool in Lovable and realised it would hold client data, the team stopped. It kept breaking and they couldn't be sure the data was secure, so they brought James in. His advice: if you won't learn where your data lives, work with someone who will.

James went further. Let policy, legislation and fear block progress, he argued, and the inertia they create becomes the biggest risk of all.

Ignore the noise and budget for the change

James expects AI to change most people's work more than it replaces them. The bigger threat is someone in your own profession who uses it well. That makes it worth benchmarking against your industry, and others, to see what's becoming table stakes.

Don't use LinkedIn as the benchmark, though. Alasdair says it makes every industry look both immune to AI and destroyed by it, depending on the day. Sam admitted it's given him something close to AI paralysis, though what he's seen come out of it is nowhere near what was promised.

Alasdair's final point was about the cost of change. Someone working in AI at Salesforce once told him to budget £5 on change management and training for every £1 spent on technology. He laughed then. Now he thinks it may be too low. When one of his team came back from a year's maternity leave, the firm's processes were completely different. And in AI Lab, the financial services AI forum he runs, he often hears of big rollouts nobody uses, or that just get bolted onto old processes.