Explained Simply
The person who signs off on an AI system is rarely the person who can read its architecture diagram. They're a practice principal, a compliance lead, a partner whose name is on the advice. They need to understand what the thing does well enough to be accountable for it — and a technical document written for engineers does not give them that. So I wrote eight short guides that do.
Why the plain-language version is the real deliverable
Most AI explanations fail in one of two directions. They're either marketing — confident, vague, and impossible to disagree with — or they're technical documentation, correct but written for someone who already knows what a context window is.
Neither helps the person who has to decide whether client data goes anywhere near a model.
What that person needs is the mechanism, not the vocabulary. Not "we use enterprise-grade PII protection" but: the model sees [PERSON_1], the real name never leaves your server, and here is the one place it gets swapped back. Once someone understands the mechanism, they can ask good adversarial questions about it — which is the point. A client who can interrogate the design is a client who can genuinely approve it.
That's a hard thing to write and an easy thing to skip. These guides are the attempt.
The collection
Three guides cover how AI actually behaves, and what that means for personal details and for money work:
- How an AI chat really works — five ideas that explain almost everything: messages, tokens, the context window, tools, and agents. Learn these and the model stops being magic and starts being a tool with knowable limits.
- Personal information and the cloud AI — what counts as a personal detail, where it goes wrong, and the two protections people constantly mix up: a disguise and a promise. You want both, for different reasons.
- Finance and the dice-rolling AI — why a model that gives different answers to the same question can still be trusted with money work. The short version: keep it out of the seat where sameness matters.
Five more cover the Document Substrate — those ideas put to work on real client documents:
- How Document Substrate works — the whole trip in eight pictures, upload to finished workpaper.
- How it's put together — the pieces, the three shelves of memory, and a question's journey through them.
- How privacy works — the promise, the three nets, the vault, and the honest edges.
- Teaching it a new job — what it takes to add a new checking job to the system.
- How answers stay good — the quality loop and the exam books that keep the output honest as the system changes.
How they're written
One idea per section, one picture per idea, and no jargon that isn't unpacked on the spot. Nearly every section carries an Under the hood box — collapsed by default — with the technical version of what you just read: the table names, the model calls, the actual guarantees. Plain language on the surface, precision one click away, and no pair of separate documents to drift apart.
The awkward parts are in there too. The privacy guide ends on "the honest edges" — the three places real client details genuinely do travel, and the one thing deleting a document can't reach. A guide that only lists strengths isn't an explanation, it's a brochure.
Each guide stands on its own, so you can start anywhere — though if AI is still largely magic to you, start with how an AI chat really works and the rest reads easily.