Transparency statement

How we use AI

We use AI tools every day, and we say so. This page states what they do in our work, what they never do, and how every output is checked before it reaches you.

Why we say so

Most practices now use these tools, and many hide it, because careless use has given AI a bad name. Hiding it would contradict the one rule this practice runs on: every claim carries its evidence. So the use is stated plainly here, together with the controls that make it safe. Used with those controls, AI lets a small practice carry the documentation load of a much larger one without lowering the standard. Used without them, it lowers the standard while looking busy.

What we use it for

• Drafting and revising documentation: permit narratives, sizing justifications, installation and operating instructions.
• Searching, extracting and summarising vendor documentation, certificates and standards, so that the right clause is found and quoted.
• Writing and testing the control software that runs on the system controller.
• Arithmetic, unit checks, and consistency checks across a package: does every number on every sheet agree with the bill of materials and the sizing chain.
• Transcribing recorded calls, which are announced before recording begins.
• Keeping the project record: indexes, cross-references, change logs, and the verification trail described below.

What it never does

• It does not make design decisions. It does not select equipment, set a rating, or choose a topology. A person does, and that person is named on the drawing.
• It does not sign anything, and nothing it produces reaches you without the principal's review against the primary source.
• It is never trusted on the strength of its own output. One AI's output is never checked by another AI's agreement; agreement between models is treated as a lead, not as confirmation.
• It is never given your material outside the terms of the engagement agreement.

How the work is checked

1. Every load-bearing claim carries its source and an evidence tier: direct (readable on a document, or measured), inference (reasoned from a cited source), or assumption (needs confirmation). The tier travels with the claim. An assumption is never written in the voice of a fact.
2. AI output is verified against the thing that could prove it wrong: the standard's text, the manufacturer's document, the measurement, the running machine. Never against a summary, and never against the AI's own report of what it did.
3. Where a claim can be checked by a machine, it is re-run. Scripts recompute counts and sizing chains. Control software is re-run against known-good cases before it is deployed. Hash sums confirm that evidence files are the files that were captured.
4. Where a claim cannot be checked by a machine, a person checks it, and the checking effort goes where errors are hardest to see, because that is where they survive.
5. Changes to the record are gated. Canonical documents change only through a review step, enforced mechanically at the point of commit, and every exception names the person who approved it.
6. Mistakes are recorded, not erased. When a reading is overturned, the correction is written as a dated addition that points at the evidence, so the record shows what was believed, what changed it, and when.
7. Independent review sits outside the practice: a licensed professional engineer where a seal is required, the third-party evaluation body, and the authority having jurisdiction. The method is built so that their review is fast and their questions have answers on file.

Your material

Client material is handled under the engagement's confidentiality terms. Which AI tools and providers process it, and under what data-retention and training terms, is stated in the engagement agreement, and we will answer that question directly at any time. Calls to and from the business line are recorded and transcribed after an announcement; a transcript is treated as a lossy record and corrected against the people who were on the call before anything in it is relied on.

What you can ask for

• Which tools touched your package, and for what.
• The verification record for any claim in it.
• Any number traced to its source. Usually that takes minutes, because the trail is kept as the work is done, not reconstructed afterwards.
• A copy of the drafting standard the drawings follow.

AI output is plausible by construction. That is exactly why none of it is taken on its own word.