How do I know what you did and what AI did for you?
A prospect asked me that question recently. I thought I already had the answer. BARE has had an internal GenAI usage policy for some time, and a one-page summary that explains to clients how we use AI and how their data is handled. I walked them through it. It wasn't enough. The more I thought about it, the more I agreed with them.
A fair question
The policy told the prospect which rules I follow. It didn't tell them, for any given document, which parts came from a model and which parts came from me. That was the real question. Not "do you have rules?" but "when I read this, whose work am I reading?" It's a fair thing to ask. If your advisor uses the same tools you could use yourself, you want to know what you are paying for.
We never started from a blank page
Here is an admission that may surprise people outside almost any professional fields: security professionals have rarely written a policy, a risk register or a procedure from scratch. Long before generative AI, we started from templates, frameworks and our own previous work. AI has made that starting point faster, but it hasn't changed what the starting point is.
A template was never the deliverable. It is one piece of a larger whole: the processes it has to fit, the risk it is meant to address, the people who will have to follow it, the understanding of the organization's culture. Used without understanding what it is for, a template is a document that looks like security and does very little.
Right template, wrong fit
When AI is wrong, it rarely looks wrong. The output is fluent, well structured and plausible. The mistake sits in the fit: a control that doesn't match how the company actually works, or a policy that addresses a risk the business doesn't have and misses the one it does.
Also, AI can be wrong in a way that sounds right, because it might not understand, or might mistake, how that template fits into the bigger context of the problem to solve. This is where I come in: I know what problem we are trying to solve, and I can review the output in that context.
This isn't only a security problem. In early October (2026, for posterity), OpenAI published AI-generated full or partial solutions to hundreds of open mathematics problems. Alongside the questions about whether every proof held up, part of the debate among mathematicians was about what a solution is worth on its own. Terence Tao argued that the field has over-rewarded being first to solve a problem, and that much of the value lies in explaining results and opening new directions. An answer that nobody can place in the bigger picture is only part of the work.
A security template is the same: correct in general is not the same as right for this company.
The knowledge of a lifetime
In 1878 the painter James McNeill Whistler sued the critic John Ruskin for libel. In court, Ruskin's barrister questioned the price of a painting Whistler had finished in two days:
"The labour of two days is that for which you ask two hundred guineas? No. I ask it for the knowledge I have gained in the work of a lifetime."
Whistler won the case. The jury awarded him one farthing in damages, he had to share the legal costs, and the trial helped push him into bankruptcy. He had the better line. It didn't pay. Asserting the value of experience is not the same as showing it, and my prospect wasn't asking me to assert anything.
What changed
That conversation changed how I work.
Every artefact is now either AI-assisted (I wrote it, using AI as a tool) or AI-generated (a model produced the draft).
AI-generated documents reach the client with a disclaimer that says so. Every template ends with a review note: what I checked, what I changed, and my sign-off.
AI-assisted work carries no label, for the same reason a policy built from a template never cited the template: the thinking, the fit and the responsibility are mine.
None of this means there will never be a mistake. People miss things too. What it does is make my part visible, and put my name on it.
The answer to "how do I know?" should never be "trust me."
DISCLOSURE: This article started with a reflection of mine, then it was drafted with AI from my notes, then reviewed, edited and signed off by me.
