Here's how AI actually arrives in most traditional practices. Not through a strategy day. Someone — often the youngest person in the office — pastes their site notes into ChatGPT, asks it to "write this up professionally", and tidies the result into the report template. It half-works. So it spreads.
Half-working is the trap. The output reads well enough to keep doing it, and falls short in ways that only show up when it matters — a wrong detail that survived because it was fluently written, a paragraph a client challenges that nobody can trace back to anything.
If your firm signs its work — architects, building consultants, surveyors, engineers — it's worth being precise about what's actually going on in that loop, because the problem isn't the AI. It's the shape of the usage.
The brute-force loop, named
The pattern is always the same: retype the context, paste the material, hope. Four things are wrong with it, and none of them are fixed by better prompts.
The model never saw the site. It's polishing your words, not checking them. It can make thin notes sound thorough — which is precisely the failure mode a professional document can't afford. Fluency isn't evidence.
The context evaporates. Every chat starts from zero. Your firm's actual knowledge — the photos, the footage, the drawings, the last three reports on that building — never enters the room, and whatever you laboriously typed in this time is gone next time.
The data path is unexamined. Client material pasted into a personal chat account sits wherever that account puts it, under terms your engagement letter never contemplated. For a practice with confidentiality obligations, that's not a hypothetical.
There's no working to show. When a line in the report is challenged, you can't show where it came from or what was machine versus judgement. For RICS members that stopped being a style question in March 2026, when the first mandatory professional standard on AI use came into force — judgement at the centre, transparency with clients, show how the output was produced.
What AI is genuinely good at, today
None of this means the technology is oversold. Used with the grain, it's remarkable, and a traditional practice has more to gain than most — because so much of the work is producing documents from evidence.
Honest list: transcription is essentially solved, even through site noise. Summarising long material — a day of capture, a stack of past reports — is solved. First-pass drafting is excellent when the draft is produced from real inputs rather than from vibes. Retrieval across your own records — "what did we say about this wall last year?" — is now possible in a way it simply wasn't.
Notice the shape of that list: it's the admin layer, not the judgement layer. AI is very good before the signature, and doesn't belong instead of it.
Four rules for a firm that signs
If you're bringing AI into a practice like this, we'd suggest holding every tool — ours included — to four rules:
- AI drafts; a named person signs. No output goes to a client without a professional owning it.
- Keep the working. Every generated line should trace to its source. If you can't click through to the evidence, you're proofreading fiction.
- Bring the tool to your evidence, not your evidence to a chat window. The context should live in the system, accumulate, and never need retyping.
- Know where client material goes. Where it's stored, who can see it, what happens when you leave. Get it in writing.
These four rules are the standard we build Sully against — they're why every summary is a cited draft, why the record accumulates instead of evaporating, and why your data stays yours. But they're not really about us. They're the test that separates AI that belongs in a professional practice from AI that just sounds like it does.
Ask any vendor to show you the working. The good ones will enjoy the question.