You ask AI to write the client a note about the delay.
What comes back is a good note. Well judged, the right length, apologetic without being weak, and better written than the one you'd have produced at the end of a long day. It's also wrong in every way that matters. It doesn't know the delay is the contractor's, not yours. It doesn't know this client has already been told once and will notice if the reason changes. It doesn't know that the thing actually holding the project up is a drawing the engineer has had for three weeks. So you spend eleven minutes fixing it, which is roughly what writing it from scratch would have taken.
That's not a shortcoming of the tool, and it's worth separating the two things carefully, because they get muddled constantly.
Ask the same tool a general question and it's genuinely strong. What the practical limitations of a particular cladding system are. How a fire regulation tends to be interpreted. What questions are worth asking a structural engineer at concept stage. For anything that belongs to the profession rather than to your office, it has read more than you have and it recalls it instantly. Architects who dismiss this are going to look silly.
Now look at what the admin half of the week actually consists of. The fee proposal. The meeting notes. The client email. The consultant chase. The project documentation. The answer to the question somebody asked about what stage the CCC’s project is at.
Not one of those is a general question. Every single one needs facts that exist only inside your practice. The proposal needs what you charged for the last three projects of that type and how badly they ran over. The meeting notes need to know who was in the room and what the position was before the meeting started. The chase needs to know what was sent, when, and what came back. The email about the delay needed to know about March.
Which puts the limit somewhere other than where most people look for it. The familiar version of this conversation is that the tools will get more useful as the models improve. On the admin half, that isn't really the constraint. The model is already capable enough to draft a decent fee proposal. What it lacks is your fee history, and your fee history isn't anywhere it can read, because it's in eleven signed PDFs, a chat thread, and the principal's recollection of which jobs hurt.
Nobody set out to make it that way. It's just what happens when a practice grows by doing work rather than by designing itself. Each project found a home for its information, usually the nearest available one, and the nearest available one was a WhatsApp group and a folder that made sense at the time.
And it's worth noticing that this is the same reason a new hire takes four months to be useful. The context isn't unavailable to AI specifically. It's unavailable to anything that isn't the person who was there. A new architect, a stand-in while someone's away, a second office, an AI assistant: all four hit exactly the same wall, which is that the practice's own information has never been written down anywhere addressable.
So the sequence most people assume is backwards. The assumption is that you adopt the tools, and getting organised is a tidy side effect that follows. In practice it's the other way round. Organised information is the thing that makes any of this worth having on the admin half, and without it you get a very articulate assistant who has to be told everything first, every time, which is most of the work.
There's a second-order version of this once you get to running the practice rather than running a project. The questions that matter at that level are which projects made money, where the time actually went, which clients are slow to pay, whether the fee structure still works. No tool can answer any of those from general knowledge. They're arithmetic on your own records, and they're unanswerable for most studios not because the arithmetic is hard but because the records were never kept in a form that can be added up.
The honest limits deserve stating too. Anything drafted for a client has to be read properly before it goes, because a confident wrong sentence about a project is worse than a clumsy right one. Information that was true two months ago gets repeated as though it's current. Client material and commercial terms are not things to paste into whatever's open in a browser tab. And none of this is oversight. A tool that can tell you what stage a project is at has not thereby taken responsibility for the project, and the judgment stays exactly where it was.
There's something worth doing here that needs no tools at all. Take the three questions you get asked most often about a live project. Where is it. What's outstanding. What did we agree. Then count how many places you have to open to answer them. If the answer is four, that's the actual job, and it's the same job that would make any assistant, human or otherwise, useful on day one instead of in month four.
Projectsmate holds that record in one place as the work happens, so those three questions have somewhere to be answered from that isn't somebody's memory. Whatever a practice eventually layers on top of its own information, this is the part that has to exist first.
The interesting question was never whether AI can help an architecture practice. It's that the half of the week everybody wants help with runs entirely on information most practices have never written down. The model is already good enough. What it doesn't have is your practice.