Initiative 05 — AI-assisted workflows, with a person at the end of the line

01What is already live on this site

Two things, and they are deliberately different. The project-brief tool on the homepage sends your description to a model and returns a short brief with recommended next steps; it runs only when you press the button, and it is labelled as an AI-generated starting point rather than a quote. The X assistant in the corner answers questions about this website from knowledge we wrote ourselves — no model, no conversation stored, nothing sent anywhere.

Two tools, two honest labels. One uses a model and says so. The other does not use a model and says that too.

02Where AI helps us internally

The useful places are unglamorous and measurable. They shorten the part of a project nobody enjoys, and none of them ship to a client unreviewed.

1. Turning a long discovery call into a structured summary a client can correct.
2. Drafting the first pass of repetitive copy — headings, meta descriptions, alt text — for a person to rewrite.
3. Summarising a competitor's site structure during research.
4. Explaining an unfamiliar error message faster than a search would.
5. Converting an agreed scope into a checklist the client approves.

03The human-review rule

This is the rule that decides whether a feature is offered at all. If the output can be wrong and nobody checks it, it does not ship.

Every AI output that reaches a client passes a person first. Anything that cannot be reviewed by a person before the client sees it is not built.

04The limits, stated plainly

Model output is fluent, which is the problem. It sounds as confident when it is wrong as when it is right, and a client cannot always tell the difference — that is our job, not theirs.

1. A model will state a fact it invented, in the same tone as one it checked.
2. It will answer in the client's language and miss a nuance a native copywriter would catch.
3. It has no memory of a decision made last Tuesday unless the system gives it one.
4. It cannot be held responsible for a wrong claim published on a client's site.

05Guardrails we will build to

The guardrails are not decoration — they are the reason a client can accept an AI feature without risking their own reputation.

1. Grounded answers: the model works from approved company facts and is told to say it does not know.
2. No invented prices, delivery dates or achievements, in any output.
3. Cost and availability handled honestly — a failed request says so instead of showing a fake result.
4. Nothing saved that was not meant to be saved; no conversation history kept unless a client asks for it.
5. A visible label wherever a model wrote the words.

06How a client would start with this

The right first AI feature is narrow and boring: something that saves an hour a week and can be checked in a minute. Bring us a task you would rather not do and describe the result you would accept, and we will tell you whether a model is the right tool or whether a form and a spreadsheet would serve you better.