Where do I start with AI?

In the Credit Crisis of 2008, I spent the summer being a cycle rickshaw driver because there were no architecture jobs. And since then, I don’t think life has got much easier for architects.

Even though total value of fees is actually going up, and the number of architects is going down, the pressure on profit margins is brutal, and the effect on salaries is that in real terms they have fallen by 30% over 30 years.

So why is this? Because there is a huge amount more admin that falls on architects. Teams have got bigger, and compliance has got more complicated, and we seem to pick up the slack.

But here is my great hope… AI is the solution. But we are using it wrong

Most practices I speak to are using AI for rendering. It’s great for rendering, but I feel that this is not really where the emphasis of the profession should be. There are really great and inexpensive tools such as Visoid, Planlift and Gendo that have solved the problem for architects - you can use them to make CGIs from hand sketches, or hand sketches from CGIs - try a few of them, then pick one, roll it out across the office.

The other main use case people have been using it for is a kind of ‘Google search on Steriods’ research assistant.

Anyone who has used ChatGPT, Gemini or Claude has had it happen. You ask a question, you get a confident answer, and it’s total rubbish.

Here’s why. On its own, a model is just a very clever form of next word prediction, or large language model (LLM). It’s trained on a huge chunk of the internet, but that data sits in its memory like half remembered snippets. Enough to sound convincing, not enough to be right. So when it doesn’t know, it doesn’t stop. It fills the gap with whatever sounds plausible. That’s a hallucination.

The fix isn’t a smarter model, it’s giving the model context.

Next time you open up Claude, Chat GPT or Gemini, do the following. Firstly, write a one-page profile of you and your practice. Who you are, what you specialise in, etc. In Claude, these are .MD files. The LLM then reads these every time you ask it to do something.

Secondly, give it examples of your writing style and rules for how you want it to write - tell it to stop saying things like “that changes everything”. Finally, pull together your three most-used documents. Your meeting agenda document, your design and access statement structure, your stage report template. Put them in a folder location that you feel is safe, such as your desktop, and give Claude access to them.

That is your Context Layer. It takes an afternoon.

But that is just the start of building your intelligence layer. The power of AI is to connect together the data that sits within your business that is currently in silos. This is where AI has the power to liberate architects from grinding admin tasks that are eating up time that should be spent on design.

The most important concern that business owners have is that of security - If I give Claude access to my documents, are the models being trained on the data? Is the information now leaking onto the internet?

If you pay for Claude, it won’t do either. The entry level of £20/ month comes with a button in the settings that pre-sets it to not let the model train on your data. Information in paid versions such as this should be as secure as they are if you use Google Drive, Eygnite, or Dropbox to store files in.

The harder nut to crack is a company wide rollout. Connections to Excel, email, Word, even via an MCP to Revit/ Archicad have a relatively small blast radius on an individual basis. But connecting to the main file server has more risk, particularly if there are multiple people with different levels of authentification accessing the server, all with Claude or an equivalent on their desktops. Danny Ivatt, the AI expert we are working with uses the analogy of meatballs and spaghetti. The LLM effectively creates a series of connections between different users and different data sources that are hard to manage.

To continue the gastronomic metaphor - AI is only as good as the data you feed it. If it is looking at one Excel file, then it is super accurate, as it has no way to get confused. There is no retrieval before the LLM goes to work - it just works on the structured data of the excel.

But architects have a veritable larder of data - from 3D BIM models, CRMs, trackers, file servers. AI agents are very enthusiastic consumers of data, but given all the data on offer, the chances of them retrieving the same information every time they go in to the data buffet is unlikely. And this is the ultimate problem that AI implementation across a company needs to solve - we need to know reliably that if we ask the same question 100 times, we get the same answer.