I’ve argued before that AI governance is the gap your frameworks don’t cover — that the models we build on break the assumptions the old playbooks rely on, because they’re probabilistic, opaque, and drift over time. That post was the diagnosis. This one is the prescription, and at a size a real team can actually lift.
Because here’s the problem with the prescriptions on offer. The serious AI-governance frameworks — NIST’s AI Risk Management Framework, ISO/IEC 42001, the OECD principles — are good, and they are heavyweight. They’re written for organisations with a governance function, a compliance team, and the appetite to stand up a new process. Most of us have none of those. The temptation, faced with that gap, is to do nothing and hope — to ship the model and deal with the governance question when someone asks. That’s the actual risk: not that we adopt the wrong framework, but that we adopt none.
Don’t adopt a framework — reuse your tools
The proportionate move isn’t to import a governance programme. It’s to notice that most of what AI governance asks for, you already do for everything else — and to point those same tools at the model.
- A model deployment is an architecture decision record. What model, why this one, what it’s allowed to touch, what we traded away, and — the part that matters most for something that drifts — a date to look at it again. A model choice with no owner and no review date is exactly the shelfware failure mode ADRs exist to prevent.
- The go / no-go is the risk equation. The expected cost of it going wrong — a confident wrong answer acted on, a leaked prompt, a biased decision at scale — weighed against the value it adds, with the honest fourth option always on the table: don’t use a model here. Some tasks aren’t worth the exposure.
- Fairness, explainability, safety, and privacy are non-functional requirements. They’re the -ilities, where the real decisions live. Written as acceptance criteria they’re testable; left as aspirations they’re decoration. And the fairness question in particular is an ethics problem you own, not one the vendor solved for you.
- Drift is an observability problem. A model that was fine at launch degrades quietly as the world moves away from its training data. If you watch latency and error rates but not output quality, you’ll find out from a customer. Monitor the outputs like you monitor everything else.
Four questions instead of four functions
NIST’s framework has four functions — Govern, Map, Measure, Manage. You don’t need the framework to use the shape of it. Collapse it into four questions a single architect can ask in an afternoon:
- Who owns this model, and when do we review it? (Govern)
- What could it get wrong, and where does it touch the business? (Map)
- How will we know if it’s drifting or causing harm? (Measure)
- What do we do when it does? (Manage)
If you can answer those four honestly, you have governance — not a certificate, but the thing the certificate is meant to stand for. If you can’t answer them, no framework will save you; you’ll just have unanswered questions in a nicer template.
Know which side of the line you’re on
One question deserves a document of its own, because it’s the one most likely to change your obligations without anyone noticing: are you still the deployer of this system, or have you become its provider? The diagnosis set out the roles — provider, importer, distributor, deployer — and how a firm can drift from one of the lightest to the heaviest. In practice the drift is mundane. It isn’t a decision to build a model. It’s a fine-tune, a wrapper with your name on it, an off-the-shelf model pointed at a use its maker never intended. Each looks like a minor adjustment, and any one of them can move you across the line.
When I was asked to look at where AI could help with some customer-facing work, the most useful thing I wrote wasn’t the options paper. It was a short governance note for management that sat beside it: for each use case, the risks, and the point at which adapting or building it ourselves would take us from deployer to provider — and what that would mean for our obligations. It turned “build or buy?” from a technical question into a governance one, which is what it had been all along.
Getting it read is the hard part
Writing that note was the easy bit. Getting anyone to engage with it was harder, and I learned more from that. I had very little feedback. Partly that was timing: it was a strategy document for work that wasn’t on the delivery roadmap yet, and attention naturally goes to this year’s delivery before next year’s plan. Partly it was the team’s focus — deep in delivery, where anything shaped like governance reads as a brake and gets quietly discounted. And partly it was that the senior people whose attention would have given it weight had bigger things on their plate.
Next time I’d do two things differently. I’d find a senior backer before I start, not after: governance with a named sponsor gets read, and governance without one gets filed. And I’d build it with the compliance team rather than beside them, lining my AI risks up with the business’s own risk register so that a model becomes one more line in a process people already run, not a separate document they have to go looking for. That’s the proportionate approach applied to its own adoption. Don’t build a parallel AI-governance track; attach it to the governance that already has an audience.
The regulation is coming, and that’s fine
None of this is instead of the law. Regulators are converging on a risk-tiered approach — heavier obligations on higher-stakes uses — and the sensible stance is to manage AI under the regimes you already answer to rather than waiting for an AI-specific one. Data-protection, sector rules, and your existing risk processes already cover most of the ground; the model is a new kind of component inside them, not a new universe. Treat it that way and most of the compliance question answers itself.
The honest caveat
This is a floor, not a ceiling. If you’re deploying models where a wrong answer costs someone their loan, their claim, or their liberty, you need more than four questions and a review date — you need the full weight of the heavyweight frameworks, and probably people whose whole job this is. The proportionate version is for the vast middle: the teams quietly adding a model to a product with no governance function and no plan, for whom the realistic choice isn’t “lightweight versus rigorous” but “lightweight versus nothing.” It depends, as ever, on what you’re optimising for and what it costs to get wrong. But “we don’t have a governance team” has never been a good reason to have no governance at all.