Three Layers That Stop AI Policy Failing

Most businesses I walk into have an AI policy. It's a PDF on a shared drive. Somebody in legal wrote it. Nobody on the floor has read it.

Then the same leadership team asks me why half the company is pasting client data into personal ChatGPT accounts.

The policy isn't the problem. The structure underneath it is. A policy tells people what they can't do. It never gives them a fast, safe way to do the thing they were trying to do in the first place. The workaround wins instead. That's shadow AI, and it's the clearest signal you have that the governance was built the wrong way round.

Here's the structure I put in instead. Three layers. It comes out of the G in C-H-A-N-G-E, the cultural operating system created by Nufar Gaspar, and it's the part of a build I won't skip.

Layer one: the sandbox

The sandbox is permission, in writing, to experiment.

Approved tools, non-sensitive data, no form, no manager sign-off. Someone wants to draft a supplier email or pull the gist out of a long report, they just go.

This is where the wow moment lives. The first time an ops manager turns 40 minutes of drafting into 40 seconds, something shifts in how they see the whole thing. Miss this layer and you never get that shift. You get a policy and a queue.

The sandbox still has an edge to it. Approved tools, non-sensitive data. That edge is what keeps experimentation from turning into exposure.

Layer two: guardrails

The middle layer is for the work that touches sensitive data, customers, or money.

One page. Pre-agreed criteria: privacy, data handling, vendor risk, how the output gets checked. The clock on it is three days. Not three weeks. Not "we'll raise it at the next steering meeting," which everyone in the room knows lands after quarter-end.

Three days, because the most common way I watch governance kill adoption is slow. Someone has a good idea. It goes in a queue. A fortnight passes. They've found a workaround and quietly stopped trusting the process. The guardrail exists to clear the right things fast.

Most of your team's AI work lives here. Client reports, CRM summaries, marketing copy, internal docs. Medium risk, high value, fine to run with a clear path through.

Layer three: the steering group

The top layer is for the decisions that can actually hurt you.

New vendors. Customer-facing automation. Anything that could dent the brand, create legal exposure, or move real money. These meet on a known rhythm, every two weeks, and they get proper human judgement on them. Human in the loop, where it earns its keep.

This is also where the strategy gets revisited. What's working, what's stuck, where the roadmap bends next. It's the layer that keeps AI pointed at the business instead of at its own paperwork.

Why three, not one

With one layer you've either locked everything down or thrown it all open. Both fail. Lock it down and people go rogue. Throw it open and nobody's accountable when something breaks.

Three layers hands you the three things most companies are missing at the same time: permission to play, a fast lane for the everyday work, and a slow, deliberate lane for the decisions that deserve it.

BDO's 2025 research put a number on it. Simple, clear guardrails cut shadow AI use by around 60%, by making the official route quick enough that going around it stops being worth the bother.

What shadow AI is actually telling you

WRITER found 94% of employees are using AI tools while only 4% of managers know about it. People call that a discipline problem. Look closer. Your team has already voted with their behaviour, and the vote says the sanctioned route is too slow or doesn't exist.

The three layers answer that with a faster route, not a tighter leash.

Where it goes in

We build the three layers before a single tool goes live, in the ENGAGE phase. We agree what sits in the sandbox, write the one-page guardrail, set the steering group and its rhythm. Governance bolted on after rollout is a seatbelt fitted after the crash.

One problem, one pilot, one measurable outcome. The layers are what make that repeatable instead of a one-off.

Skills that stay when we go.

Build the layers before the tools

Our governance work sets your sandbox, your one-page guardrail, and your steering group rhythm, so the official route is faster than the workaround.

See our Governance work

Human and technical governance • ISO 42001 aligned

Common questions

What are the three layers of AI governance?

The three layers are a sandbox, guardrails and a steering group. The sandbox is a defined space where people can experiment with AI on non-sensitive work without asking permission each time. Guardrails are the standing rules about what data may be used, what must never be entered, and which decisions require a human. The steering group is the small standing body that reviews what is happening and changes the rules as the picture changes. Most AI policies fail because they attempt only the middle layer, which produces rules nobody can follow and no mechanism to revise them.

Why do written AI policies fail in practice?

Because a policy with no sandbox gives people nowhere legitimate to experiment, so they experiment anyway without telling anyone. A policy with no steering group cannot be updated as tools change, so it becomes inaccurate within months and staff learn to ignore it. Rules alone describe the desired behaviour without creating the conditions for it. All three layers are needed together.

What is shadow AI and what does it indicate?

Shadow AI is employees using AI tools that the organisation has not sanctioned, usually through personal accounts. It is best read as a signal rather than a discipline problem: it shows exactly where people find their work slow or repetitive, and where the sanctioned toolset is failing them. Organisations that respond by banning tools lose that signal and drive the behaviour further out of sight.