If you run a practice, someone has already put a report through a chat window. Possibly you. Almost certainly one of your people, quietly, because it saved them an hour and nobody said not to.
That is the actual starting position for most firms. Not “should we adopt AI” but “AI is already in the workflow and nobody has looked at it.”
Here is what to ask. It applies to us, to every competitor, and to the free chat window your grad is pasting into right now.
Who is the author of record?
Start here, because everything else follows from it.
There is only one acceptable answer: a named person in your practice, who read the thing, and whose signature means what it has always meant. Any product whose pitch is that it produces a finished, sign-ready report is selling you something you cannot professionally accept. The value is in removing the typing, not the judgement.
Ask the vendor directly: does your product claim to produce a final report, or a draft? The answer tells you how seriously they have thought about who carries the liability.
Can you show your working?
If a line in your report came out of a machine, you need to be able to answer “where did this come from” months later, possibly to someone hostile.
The specific things to look for:
- Does every generated statement trace to a source? A photo, a frame, a timestamp, the words that were said.
- Can you get at that trace from the report, or only from the vendor's app?
- Is the trace preserved if you stop paying?
A tool that produces confident prose with no path back to the evidence is a liability generator. The prose is the easy part. The trace is the product.
What does your PI insurer need to see?
This is the question most firms discover late.
Talk to your broker before you roll anything out, not after. The professional indemnity market has moved on AI faster than most practices have, and the requirements clustering into policies are broadly consistent: a named human reviewed the output, there is a record that the review happened, and the client was told AI was used in the process.
The clearest institutional marker so far is on the UK side. RICS brought in a mandatory standard for AI use in professional work, effective March 2026, built on exactly those three things: named-surveyor review, written records for PI purposes, and disclosure to the client. Non-compliance can affect coverage.
Australia has no direct equivalent yet. But if you are wondering what “good practice” is going to be defined as, that is a reasonable preview, and building to it now costs you very little.
What are you disclosing to clients, and where?
Decide this deliberately rather than by accident. A short, plain line in your engagement terms and on the report itself, stating that AI assists in preparing drafts and that a named professional reviews and signs, is not a weakness. It is the same posture as saying you used a drone or a moisture meter.
The firms that will have a problem are the ones who never said anything and then get asked.
Where does the data go, and who trains on it?
Your walkthroughs contain other people's buildings, other people's defects, and sometimes other people's faces.
- Where is it stored, and in which country?
- Is it used to train the vendor's models? If the answer is anything other than a clear no, or a contractual opt-out, treat that as a finding.
- Who at the vendor can see it?
- What happens to it if you leave?
The free chat window is worth singling out. On consumer tiers, your prompt history is generally not covered by anything resembling a professional confidentiality arrangement. Pasting a client's condition report into one is a disclosure decision, whether or not anyone framed it that way.
What happens when you stop paying?
Ask it plainly, because the answer is rarely on the pricing page. If your subscription lapses, can you still read, search, and export the records you already made? A record you lose access to was never really yours.
Our answer, for what it is worth: processing pauses, the record stays viewable and searchable. We think anything else is indefensible for something that is meant to be a record.
How wrong does it get, and how would you know?
Every one of these systems is wrong sometimes. The vendors who are worth dealing with will tell you where.
Ask what it is bad at. If you get a clean answer about limits, that is a good sign. If you get “it is very accurate,” you are talking to sales, and you should assume the limits exist and are undocumented.
Then test it yourself, on a job you already finished. You know the answer, so you can grade it. Do that before it goes near live work.
A reasonable position
You do not have to be first and you do not have to abstain.
Pick the narrow, high-volume, low-judgement part of the work. For most practices that is the write-up, not the assessment. Put one tool in one workflow with one named reviewer, run it against jobs you have already completed, write down what you found, and tell your insurer what you are doing.
That is a defensible position that survives contact with a bad day. “We banned it” is not, because it is not true, it is just unsupervised.