Your pilot is stuck in the security questionnaire.
In five business days you get the threat model, the deployment topology, and a written answer to every question your security team has asked — for the AI system you already have, whether we built it or not. Enough to either clear the review or know exactly why it cannot be cleared.
At a glance
- Price
- $3,500
- Turnaround
- 5 business days
- Access to your data
- None
- Obligation to continue
- None
- Must be our system
- No
Why this exists
Most AI projects do not fail technically.
They work in the demo, and then they sit for four months while nobody can produce a document that satisfies the security review. The engineering was never the blocker. The evidence was.
The people who can answer are not the people being asked
The questionnaire lands with whoever built the pilot. They are an ML engineer. It asks about egress policy, tenant isolation, key custody, and audit reconstruction — and the honest answer to most of it is "I would have to find out."
Nobody wants to be the one who signs it
A reviewer asked to approve something they cannot inspect will default to no, and the default is invisible — it looks like the request is still being considered. It is not. It is waiting for evidence that nobody has been asked to produce.
Sometimes the honest answer is that it cannot ship
If the architecture genuinely cannot meet the requirement, five days and $3,500 is a cheap way to learn it. That outcome is not a failed engagement — it is the point. Knowing in a week beats discovering it in month six.
The deliverable
Four documents, written for your reviewer.
Not a slide deck and not a recommendation. Documents your security team can read, mark up, and act on without a meeting.
01
Threat model
Trust boundaries, the assets on each side of them, and the specific attacks the architecture is and is not exposed to — prompt injection, tool abuse, data exfiltration through the model, and privilege escalation via agent-invoked tools. Written against your system, not a generic template.
02
Deployment topology
A diagram your reviewers have already seen the like of: where execution happens, what crosses which boundary, where credentials live, what egress is permitted, and which components hold data at rest. The document that ends the "but where does the data actually go" conversation.
03
Questionnaire answer set
Your actual vendor assessment, answered line by line in writing — or, if nobody has sent you one yet, the questions your reviewer is going to ask, answered before they ask them. Each answer states what is true, what is not, and how to verify it.
The questions themselves are published free — all twenty-four of them. Work through them yourself first; if they are answerable without us, you have saved $3,500.
04
The verdict
Two pages. Whether this can clear review as built, what has to change if it cannot, roughly what that change costs, and whether it is worth doing at all. Written plainly enough to forward to an executive without translation.
How the week runs
Two hours of your time, total.
The constraint on a five-day engagement is your calendar, not ours. It is structured so that nothing waits on a meeting that has not been scheduled yet.
DAY 1
Ninety minutes, architecture level
You walk us through what was built and what the reviewer objected to. Bring whoever wrote it. If you have the questionnaire, send it beforehand — we would rather arrive having read it.
DAYS 2–4
We write, you carry on
No standups, no status calls, no access requests. Occasionally one clarifying question by email. This is the part that does not need you.
DAY 5
Documents, then thirty minutes
The four documents arrive first so you can read them cold. The call afterwards is for the questions they raise. Invite your reviewer — that conversation is usually worth more than the documents.
The first three are free, and here is why
Not a discount and not a trial. Every client we have worked with has asked us not to use their name, which is the right instinct and we have honoured it — but it leaves this firm arguing for itself with no outside voice. We would rather earn three references than keep asking people to take our word for it.
So the first three are free, and the only thing asked in return is agreed before we start: if the work is good, you say so on the record — a named reference, a quotation we can attribute, or an introduction. If the work is not good, you owe us nothing and we will say so ourselves.
After the third, it is $3,500. Stated here so that nobody discovers the price after the fact, and so that the three are actually three.
Disqualifiers
When not to buy this.
A five-day memo is the wrong instrument for several common situations. These are the ones we see most.
You need someone to sign off on the risk. This is an engineering assessment, not an audit and not an attestation. It informs your reviewer's decision; it cannot replace it.
There is no system yet. If nothing has been built and nothing is being bought, there is nothing to assess. Book a briefing instead — it is free and it is the right shape of conversation.
You want a penetration test. Different discipline, different firms, and several of them are very good. This assesses architecture and evidence, not a running deployment's exploitability.
The decision has already been made. If the project is being cancelled and this would be documentation for the file, say so. We will decline, and save you $3,500.
Three sentences is enough to start.
What the system is, who is blocking it, and how long it has been sitting there. We will tell you within a day whether a memo is the right instrument — and if it is not, which one is.