AI for Architects in South Africa: What's Actually Working in Small Practices

If you run a small practice, you already know where the week goes. A brief comes in and has to be turned into a project. Site information has to be found, checked and drawn up. A fee proposal has to go out before anyone can start. Somewhere in there, most of us are also trying to draw. AI gets talked about as the answer to all of it, but most of what’s written about “AI for architects” is either a generic list of chatbot tips or a sales pitch for one plug-in. Neither tells you what actually changes in a working practice.

I run AB Architecture, a residential practice in Knysna and Cape Town that also works under the joint brand AB+HvD Architects. That is where these systems actually run, not a demo built to sell software. I’ve worked in the built environment for close to 20 years, across South Africa, Norway and the USA. I’ve also used AI daily since ChatGPT launched in late 2022, and I’ve spent years building software with AI coding tools alongside the design work. Over that time I’ve rebuilt a fair amount of my own office around it — not as an experiment, but because the admin and repetition were eating time I’d rather spend on the work itself. This post is what that’s actually looked like: what’s in daily use, what’s still early, and where I’ve deliberately kept a person checking the output.

Where AI actually saves time in a small practice

The useful version of AI in a practice like mine isn’t a single tool — it’s a handful of narrow systems, each doing one job, feeding off the same project information instead of retyping it. The pattern that’s worked for me is: gather the information once, keep it in a proper project record, then let AI do the first pass on the repetitive parts — drafting, filling, summarising — while a person checks the result before anything goes out.

That’s a different idea to “AI will design your building” or “AI will run your office.” It doesn’t. What it does is remove the parts of the week that were never really about architecture in the first place: retyping the same property details into three different documents, starting a fee proposal from a blank page, or hunting through email for what was agreed on a call two months ago.

What I’ve built and use every day

A few examples, described honestly rather than as a highlight reel:

• Project setup. When a new job starts, a workflow creates the folder, searches my email and files for anything relevant, and fills in a standard project record — client, contact details, property, key dates. It reads the title deed and stores the conditions that matter, and suggests structural engineers and other consultants who suit the job and the area. See the project setup case study for the fuller description.
• Site and zoning research. I built Zonely, a web app that turns Cape Town’s zoning rules into plain answers for a property — plot size, what you can build, in seconds instead of a long by-law read. The same site data flows straight into Revit, already set up with boundaries, building lines, setbacks and coverage. That’s a Cape Town–specific tool right now; the underlying idea — get accurate site data into your drawing software before you draw a line — applies anywhere.
• Scan to draft drawings. A laser scan of an existing building becomes a draft Revit model and as-built drawing set, which I then check and correct. I go into this in detail, including honest numbers from my first live project, in a separate post and at the scan-to-drawings case study.
• Fee proposals. I describe the client and the job out loud, and a system drafts the proposal in the practice’s house style, using the current SACAP fee guideline, our standard terms and what we’ve charged on similar jobs. More on this at the fee proposal case study and in a dedicated post.
• Council forms. Municipal and building-regulation forms ask for the same details repeatedly. A tool reads the project’s own documents, fills each form, and lists what’s still missing — anything it can’t find is flagged for me rather than guessed.

None of this runs unattended. Facts get checked against more than one source where that’s possible, and anything uncertain gets flagged rather than filled in — that rule matters more than any of the tools themselves.

A practical checklist: where to start in your own practice

If you’re weighing up where AI could actually help rather than just generate more busywork, this is roughly the order I’d work through:

1. Write down where a normal week actually goes, not where you think it goes. Most people underestimate admin time until they track it for a week.
2. Find the thing you retype most often. Property details, standard terms, the same three paragraphs in every proposal — that’s usually the first candidate, because it’s low-risk and the payoff is immediate.
3. Pick one workflow, not five. A single tool with one clearly defined job is easier to trust, test and fix than a general-purpose assistant with access to everything.
4. Decide what “checked” means before you build it. Will a person review every output, spot-check a sample, or just sanity-check the first few weeks? Decide this up front, not after something goes out wrong.
5. Keep drafts as drafts. Nothing — an email, an invoice, a proposal — should leave the office without a person approving it first.

Limits — what still needs a person

I’d rather be specific here than vague. AI makes mistakes, and I see them every week in my own systems. A drafted Revit model still needs review and correction before it’s usable. A fee proposal still needs a read-through against the actual brief, not just the guideline. A filled council form still needs someone who understands the submission to check what’s flagged as missing. None of the figures I use publicly — the 5–6× faster fee proposals, the 30–45 minute first-draft scan-to-model time — are industry benchmarks; they’re my own estimates from my own practice, and they don’t include review and correction time, which is real and separate.

The other limit is data. If a system touches client information, that’s something to think through properly — which AI service sees what, where files are stored, how long anything is kept — before you build it, not after.

What this could look like for your firm

I started with architecture because I know the work from the inside — but the pattern (gather once, keep one record, let AI do the repetitive first pass, a person checks before anything leaves) isn’t architecture-specific. Engineers, land surveyors, interior designers and property managers run into the same shape of problem: information scattered across email, documents and one person’s memory, and admin that repeats itself every week.

If any of this sounds like your practice, the way I’d suggest starting is the same way I’d want to start if I were in your seat: a short conversation, not a sales pitch. Book a free 30-minute call and tell me what eats your week — I’ll tell you honestly where AI can help and where it can’t yet. See Services for how the free call, the paid check-up and the build stage fit together.