Something a client said to me a few weeks ago has stuck. We were talking about Copilot and he said, more or less, "we've got a policy, so we're fine". I didn't argue with him at the time. But I've been turning it over since, because I think it's where most firms are right now, and I think it's wrong.
The policy isn't the problem. It's just that a policy tells you what people are supposed to do with AI. It doesn't tell you what you've already given away.
Fragments
For most of my career the job was putting intelligence pictures together out of bits. You rarely got the whole thing. You got a name in one place, a date in another, a document that on its own meant nothing much, and you worked out what they added up to. That was the skill, really. Not finding the one big secret, which almost never existed, but noticing that twenty small things everybody thought were harmless were a complete answer once you laid them side by side.
I can't help looking at how businesses are using AI through that lens, and what I see worries me a bit.
Think about what goes in over a normal month. Someone pastes a bid summary into ChatGPT to tighten the language. Someone else uploads a client org chart because they want help drafting an email to the right person in it. Copilot has been reading every Teams meeting transcript since the day it was switched on, including the one where you talked about who you're going to poach from a competitor. A junior engineer has connected a notes tool to the shared drive because the summaries are good, and nobody mentioned it to IT because why would they.
None of that is a leak in the way we normally mean it. But between them those things describe your bids, your clients, your people and what you're capable of. And they describe all of it to whoever can get at the vendor, the account, or the person using it.
AI does not leak your secrets one document at a time. It assembles them.
If you're a defence supplier this stops being an interesting thought experiment fairly quickly. Primes have started asking where that data goes. The MOD is asking. Your own customers are asking. Most firms I talk to can't answer, and the ones that try tend to answer with the policy, which is where I came in. The policy isn't the answer. The answer is what has actually gone in, and what a competent adversary could do with it.
So we built something for it
We've called it EXPERTIS APERTURE. It's a threat-led assessment of what your use of AI exposes. It isn't a privacy audit and it isn't a policy review, and I'm labouring that because the market is filling up with things that look similar on the surface. We look at Copilot, ChatGPT, the AI that has quietly turned up inside software you already pay for, and the tools your people are using that you don't know about, and we work out what the lot of it gives away.
The reason I'm comfortable saying we're the right people to do this is simple enough. Dave, Les and I spent a long time on the collection side. We know how a hostile state goes about building a picture of an organisation because, for a good chunk of our careers, that was the job. So when we look at your AI footprint we're not asking whether it's compliant. We're asking what we could get out of it if we were on the other side. Those are different questions and they get different answers.
What it looks like in practice
It runs for four to six weeks, remote or on site or a mix, and it's fairly light on your team's time because we've been doing this sort of interrogating for a long while and know where to dig.
We start by framing it with you. Who wants what you hold, what you'd least want them to have, which clients and programmes carry the most risk. Then we run structured sessions across twelve themes with the board, the security lead, the function heads and, importantly, the people actually using the tools day to day. If I had to pick the single most useful thing in the whole engagement it would be the gap between what the board thinks is happening and what the users tell us is happening. That gap is nearly always where the finding is.
We score each theme and roll it up to a level for the organisation. The scoring is weighted towards aggregation risk, adversarial risk and contractual risk, and it's gated, so a good score in nine themes can't paper over a bad one in the tenth. Then we present it to the board, in person, with a roadmap and timescales.
You walk away with a threat model of your AI exposure, a maturity level you can stand behind, and your AI use mapped against DEFCON 658, your security aspects letters, CSM, DCC and UK GDPR. You also get the thing I suspect most people will actually use, which is a one-page AI sovereignty statement you can hand to a prime when they ask. If you want one we'll write you an acceptable use policy too, but we write it at the end, once we know what your people are really doing, rather than at the start.
Who it's for
Tier 2 and 3 defence suppliers mostly, plus national security consultancies and dual-use technology firms. Really it's anyone handling client-confidential, export controlled or OFFICIAL-SENSITIVE material who has let AI anywhere near it, and at this point I'd be surprised if that wasn't you.
If you'd like to know what your organisation looks like from the outside, book a conversation. We'll show you what they see.