What AI does here, what a person checks, and where the record is.
A consultancy that sells governed AI should be able to say exactly how it uses AI itself. This page does: what the machines do, what a person checks, and where the record lives. It is the shortest proof we own.
The disclosure
Six places AI works inside ProjxAI. For each: what the machine does, what a person does, and where you could check.
| Where | What the AI does | What a person does | Where the record lives |
|---|---|---|---|
| Articles under Insights | Drafted by the content engine from an approved editorial theme and our own facts; house style applied in code before anything is recorded. | The review gate passes or holds every piece; the theme definitions were approved by a person one at a time; any article can be edited by hand and the edit is what the training signal is built from. | content_items, with body as generated and body as published kept separately; approval records with a hash of what was approved. |
| Website copy | Drafted with an AI assistant in a session Michael ran, page by page, from a written brief and the research documents in the repository. | Michael reads and edits every page; nothing goes live on this site without a person having looked at it in a browser at desktop and phone widths. | The repository’s commit history, which names what changed and why, and the session notes that record each decision. |
| Decisions about our own business | Anything that would cost real money goes through the Verdict Engine: criteria first, a panel of competing models under opposing roles, cross examination. | A person reads the panel and gives the verdict. The first verdict the process ever gave was STOP, on this business’s own founding project. | A dated audit trail per run; the client version is the decision pack. |
| Client work | Scoping, drafting, integration code and measurement pages are prepared with AI assistance inside the systems we run. | A named person is accountable for every scope and every pilot, signs the pre commit, and checks the number before it is reported. | The Command Centre’s immutable review events; the client portal shows only what a policy allows. |
| Our own machines | ProjxAI Attest’s Recorder configured them so the evidence exists, and its Ledger records AI attributable activity. | The monthly attestation is read by Michael; the verifier checks the ledger without the product. | The ledger and the attestation, per machine, per month. |
| This page | Drafted with an AI assistant from the records above. | Read by Michael before publishing, like every page. | This commit. |
The runner, in its own words
The content engine’s own log from its most recent scheduled run on this website. This is what unattended publishing looks like when it is governed.
Every article on this site since July has arrived this way. The one thing the log does not show is the person: the theme it wrote under was approved by a person, the review policy it passed was set by a person, and a person can hold, edit or unpublish it at any time from the portal we built.
What we will not do with AI
Publish a figure without its source. Name a client without written permission. Let an AI decide anything a client would expect a person to decide. Put personal information into a model without pseudonymising it first. Present anonymised work as named work. Claim a capability of our own products that is not built.
Where we have got one of these wrong, it is recorded in the repository and corrected in place rather than quietly deleted. That is the standard we ask of clients and it would be odd not to hold it ourselves.
Want the same disclosure for your own business?
A one page statement of what AI does in your business, what a person checks and where the record lives is a good first piece of governance, and we will help you write it.