What AI is good at, and what it isn't

The AI tools most architects meet are large language models: ChatGPT, Microsoft Copilot, Google Gemini, Claude and the assistants built into other software. They all work in a similar way. They predict likely text based on what you give them and what they learned in training.

That one fact explains most of their strengths and weaknesses. They are very good with language. They are not databases, they do not “know” things in the way a reference book does, and they have never stood on your site.

Where AI genuinely helps

Drafting

A first draft of a routine email, a covering letter for a planning submission, a fee proposal introduction or a job advert. You give it the facts and the tone, it gives you something to edit. Editing a decent draft is quicker than staring at a blank page, especially at the end of a long day.

Summarising

Paste in a long email thread and ask what has been agreed, what is still open and who is waiting on whom. Give it a 40-page consultant report and ask for the points that affect your design. Summaries are where most people feel the time saving first.

Meeting notes

Rough notes typed during a design team meeting, or a transcript from a Teams call, can be turned into tidy minutes with decisions and actions pulled out. Tools like Pulse do this inside the project record, so the decisions are filed with the job rather than left in a document somewhere.

Research starting points

AI is useful for getting your bearings on an unfamiliar subject: what to look for in a heritage statement, the general considerations for a timber frame on a sloping site, or questions to ask a client about accessibility needs. It is a starting point for your own reading, not the answer.

Specification and planning-policy research

This is the area to be most careful with. AI can help you structure a search, explain a term, or summarise a policy document that you have given it. It is far less reliable when asked to recall a specific policy, product or standard from memory. It may name a local plan policy that does not exist, or quote the wrong version of a document.

The safe pattern is simple: find the source document yourself, then ask the AI to help you read it.

Where AI lets you down

Task	Why it goes wrong
Quoting regulations, policy numbers or standards from memory	It produces plausible text, not checked facts, and its training data may be out of date
Arithmetic, areas, fee calculations	Language models are not calculators, and errors look just as confident as correct answers
Anything about your specific site	It has no knowledge of your site, your client or your local authority unless you tell it
Judgement calls	It cannot weigh a client relationship, a planning officer’s mood or a contractor’s track record
Current events and recent changes	Training data has a cut-off, and some tools do not search the web

The most important thing to understand is that AI does not signal uncertainty well. A wrong answer reads exactly like a right one. That is why the later lessons on checking output matter so much.

A useful way to think about it

Think of an AI tool as a very fast, very well-read junior who has never worked in practice and sometimes makes things up to avoid saying “I don’t know”. You would happily ask that person to tidy your notes or draft a letter. You would check everything before it went out. You would not let them sign off a drawing.

Sorting your own tasks

Here is a simple exercise. List ten things you did last week that involved writing or reading. Then sort them.

Good candidates for AI (low risk, mostly language):

• Tidying meeting notes into minutes
• Drafting routine emails and chasers
• Summarising long threads or reports
• Rewording something for a client who is not an architect
• First drafts of standard letters and proposal text

Use with care (the facts need checking):

• Planning-policy summaries, using documents you supply
• Specification research and product comparisons
• Design and access statement structure

Keep for yourself:

• Fee calculations and anything numerical
• Professional judgement and advice to clients
• Anything you would be uncomfortable explaining to the ARB

Most practices find the first list covers more of the week than they expected.

Try this week

• Take one long email thread from a live project, remove names and sensitive details, and ask an AI tool to summarise what has been agreed and what is still open. Compare it with your own understanding.
• Write down your ten reading and writing tasks from last week and sort them into the three lists above.
• Notice one moment where an AI answer sounded confident and turned out to be wrong. That is the habit of mind this course builds on.