How we use AI.

1. In our own work

We use AI coding assistants, language models and generation tools in the ordinary course of our work: drafting and reviewing code, writing tests, summarising documents, exploring design directions, and producing first drafts of text. A person decides what goes to a client, reviews it, and is accountable for it. Nothing is delivered because a model produced it; it is delivered because an engineer or designer judged it correct.

Client material is never used to train a model, ours or anyone else's. Where we use hosted AI services on client work, we do so under settings that exclude training and limit retention, and we tell the client which services those are so they can appear in the client's own records.

2. On this website

The service, industry and technology pages you are reading are written and edited by us, with AI tools used in research, drafting and revision. Every factual claim, figure and description of our work is checked by a person, and the site is tested against a written set of rules before it ships — no invented metrics, no credentials we do not hold, no unenforceable guarantees, and no claim that a system eliminates error. Those tests run in our build, so a page that breaks a rule fails rather than publishes. Our own logo and the diagrams are hand-built. Screenshots on the work pages are of the live sites, not renders.

Vincent Valentine is the editor responsible for this site’s pages and articles. No change reaches the live site until Vincent Valentine has checked it against its sources and approved that exact revision. We improve existing pages gradually, a section at a time: each change is previewed on a private copy of the site, measured over the weeks that follow, and reverted if it does not help readers.

The engineering blog is written with AI assistance. A human editor supplies or verifies the source material, checks the claims against the cited sources, edits the finished article and its illustrations, and approves the exact revision before it enters the publication schedule. No article publishes solely because a model produced it or gave it a high score. Articles approved this way carry Vincent Valentine’s byline and link to a profile page; Vincent Valentine answers for them.

When a model helps draft a post or a page change, a second pass reviews it against the same written rules the rest of the site is tested against, and mechanical checks run alongside it: banned phrases, invented figures, links that do not resolve, changes larger than the edit asked for, and articles too thin to be worth a reader's time. Those checks are evidence for the editor; they are not permission to publish.

News posts begin with source material and links. We check the sources, edit the article, generate any illustrative images, and review the complete package. Approved articles are placed into spaced publication slots so preparing several pieces does not make them all appear at once.

A machine-drafted post that has not had that named editor’s review carries an “AI-drafted” marker on the post itself and in its structured data. The illustrations on our posts are generated, and they are abstract by instruction: we do not generate photographs of people, screenshots of software that does not exist, or charts of data we have not measured. If you find an error in any of it, tell us and we will correct or remove the post. Nothing here is too valuable to delete.

Two things this site does not do: there is no AI chat assistant on it, and there is no automated profiling of visitors. Every enquiry is checked automatically for spam, by rules that run on our own server. Anything that looks like spam is held for a person to review, never deleted automatically. The contact form is read by a person, and the multiple-choice answers on it are there to make the first call shorter, not to score you.

Each week we ask AI search tools — Perplexity, ChatGPT search, Claude and Google’s AI Overviews — a fixed set of questions about the work we do, to see whether their answers cite this site, and we look up search volumes and results pages for the services we offer. Those questions are written by us; no visitor or enquirer data is involved.

3. What we tell clients

• Which models and providers a system uses, so the client can name them in its own privacy notice and vendor records.
• Where the model can be wrong, with an evaluation set the client agrees and results they can inspect, rather than a headline accuracy figure.
• What the system is allowed to do without a person, and where a human decision point sits, especially when an output could affect someone's legal position, finances, health or safety.
• Where a system talks to the public, that it must identify itself as automated where the law requires it. Article 50 of the EU AI Act makes that a duty for the operator of many such systems, and we design the disclosure in rather than leaving it to a policy document.
• What it costs to run, including model usage, so there are no surprises after handover.

4. What we will not build

We decline work whose purpose is to deceive people about whether they are dealing with a machine, to manipulate people against their interests, to carry out surveillance that the law does not permit, or to make consequential decisions about individuals with no route to a human. The EU AI Act prohibits several of these outright; we treat the rest as a matter of professional judgement, and we would rather lose the project.

5. How we talk about AI

We do not claim that a system eliminates error, guarantees an outcome, or replaces a professional's judgement. Retrieval makes model errors traceable to a source; evaluation makes them measurable; monitoring makes them visible. Those are the honest claims, and they are the ones we make. If you find a sentence on this site that overreaches, tell us at [email protected] and we will fix it.