How We Use AI.
A living document. Essays are fixed once published; this one changes as our practice does. It describes what we actually run, not what sounds good on a capabilities page.
We build with AI, and we sell AI systems to other businesses, so the honest version of this policy has to answer one question: who is accountable when the machine is wrong? The answer is a person, by name, every time. Everything below follows from that.
The four rules
1. A named human owns every deliverable. AI drafts, summarizes, extracts, and checks. It does not sign. Nothing reaches a client — a filing, a return, an audit finding, a contract — without a person who read it end to end and put their name on it. If we cannot point to the human who owns a piece of work, that work does not go out.
2. No autonomous client-facing action. Our systems propose; a human approves. Sending, filing, paying, submitting, or committing something on a client's behalf sits behind an approval gate by design, not by configuration. We build the gate first and the automation second — that order matters, and it is the opposite of how most agent projects get built.
3. Your data is not training data. Client data is processed to do the work you hired us for and for nothing else. It does not get used to train models, it does not get pasted into consumer AI tools, and it lives in environments we control rather than a vendor's shared playground. Access is scoped to the systems that need it, not granted account-wide.
4. You can switch it off. Tell us a piece of work should be handled entirely by people and it will be. That applies to a whole engagement, one system, or one document. No fee change, no argument, no upsell — it's your work.
Where we do use it
- • Drafting and restructuring writing, proposals, and internal documentation.
- • Research intake — reading, extracting, and summarizing public sources at volume.
- • Code we build and maintain, including the checks that catch our own mistakes.
- • Internal operations: routing, scheduling, reconciliation, and reporting.
- • Review passes that look for what a tired human would miss at 11pm.
Where we do not
- • Final judgment on a client's books, taxes, eligibility, or legal position.
- • Anything submitted to a government agency or a carrier without human review.
- • Client-identifying data inside third-party AI services.
- • Conversations that require advice. The system can brief; a person advises.
How we test our own claim
A policy nobody checks is decoration. Two habits keep this one honest. First, every system we ship passes a pre-launch security gate before it touches a real user — the checklist is public. Second, when an automation produces something wrong, we treat it as an operating failure of ours, not a curiosity about the model. The machine did not sign anything.
Hold us to it
If you work with us and any of the four rules above does not match what you are seeing, tell us directly and we will fix it or explain exactly why it happened: franklin@prospyr305.com. A policy you cannot challenge is a brochure.
This note is a living document and will be revised as our tooling and our standards change. When it changes, the date at the top changes with it.