Artificial intelligence is already useful in an architecture practice, just not where most people expect. Today it saves the most time on documentation and administration: site visit reports, technical reports, specifications, responses to authority comments, quantities and email. It doesn’t replace professional judgement or a signature. The safest way to start is to pick one repetitive task, test it on real examples from the practice, and always keep a person who reviews before anything goes out.

This guide covers what AI can do in a practice, what to keep away from it, how to look after data, and what each way of getting started costs.

Where the hours go in a practice

Anyone who has worked in a practice knows it: a large part of the week isn’t spent designing. It goes on tasks that repeat with small variations.

  • Writing up notes and photos from a site visit.
  • Drafting certificates, reports and minutes in the practice’s own format.
  • Adapting the specification of a previous project to a new one.
  • Answering comments from the planning authority or building control, point by point.
  • Preparing quantities and cost estimates from spreadsheets.
  • Replying to emails from clients, contractors and consultants.

These are necessary tasks with a known format and a lot of text. That is exactly the kind of work today’s AI is good at.

What AI can do in an architecture practice today

Site documentation

This is the clearest case. From the notes taken on site, a voice memo or the photos, AI prepares the minutes or the report in the practice’s structure and tone: progress, issues, instructions given and who is responsible. The architect reviews it, corrects what’s needed and signs.

What matters is that the result comes out in the practice’s own format, not as generic text. For that, the AI has to work from documents the practice has already signed. That is what NotaFiel, our product for certificates, site records and technical reports, does. For close-out, we explain a complete method in completion certificates and handover files.

Specifications and project texts

Given the practice’s earlier documents and the data of the new project, AI writes good first drafts of design statements, specifications or competition texts. It saves starting from a blank page or adapting an old document by hand.

What it doesn’t do well is fill gaps responsibly. If a fact isn’t in the source material, it may supply something that sounds right and isn’t. So every draft is read in full, and every figure and every regulatory reference is checked. We go into this in specifications and design statements with AI.

Authority comments

When a request for information arrives with ten points, AI can split it into numbered items, find where the project deals with each one (or flag that it doesn’t), and draft a response in the practice’s format. The architect decides what to answer and what to change.

Quantities and cost estimates

From a quantities spreadsheet, AI can sort items, flag inconsistent units and prepare a draft estimate using the practice’s own rates. It doesn’t replace the estimating software: it gets the file ready for a person to check and import.

Competitions and tenders

Tender documents are long and similar to each other. AI summarises deadlines, award criteria, required experience and the documents to submit on one page, and lists what’s missing. Deciding whether to bid, and at what price, stays with the practice.

Everyday communication

Meeting summaries, draft replies to clients, translations for international clients. Small savings, but they add up every day.

Images and design

Image generation tools are useful for exploring ideas, building mood boards or trying variations of a material. They don’t replace the project model or the final visualisation, and their terms of use are worth checking before showing results to a client.

What to keep away from AI

AI is brilliant at some tasks and clumsy at others that look easier. Knowing which is which is half the work. These are the ones we don’t recommend automating:

  • The signature and professional responsibility. A signed certificate or report is the responsibility of whoever signs it, whoever drafted it.
  • Deciding whether a project complies with regulations. AI can help find clauses and prepare checklists, but it sometimes misquotes or mixes versions of a regulation. Always check against the official text.
  • Structural or building services calculations. There is dedicated, verifiable software for that.
  • Design decisions and the client relationship. That’s what the client pays for, and what sets one practice apart from another.

An AI that is almost always right is still sometimes wrong. The question isn’t whether it makes mistakes, but who checks. That’s why in everything we build the final decision always belongs to a person, and every piece of data shows where it came from.

Three ways to start

From least to most effort:

1. Assistants set up with the practice’s templates

A business plan for ChatGPT, Claude, Gemini or Copilot, with saved instructions that include the practice’s templates, tone and standards. It’s the cheapest way to start: licences cost a few tens of euros per user per month. It works well for drafting and reviewing text, but each person copies and pastes by hand, and the result depends on how it’s used. To get started, there are instructions ready to copy in 15 AI prompts for architects and engineers.

2. Automating one specific process

One recurring process (for example, from site notes to minutes as a PDF, or from an authority comment to a draft response) connected to the tools the practice already uses: email, WhatsApp, spreadsheets, the project folder. The AI prepares and proposes; a person approves. The savings are bigger here because the copying and pasting disappears.

3. A custom tool or a specialised product

For a complete workflow, or for documents that carry a lot of responsibility, a dedicated tool or a product built for the sector pays off. It takes more investment, but the result is more reliable and doesn’t depend on everyone writing good instructions.

Data: what to look after before you start

It’s the question we’re asked most, and rightly so. A practice handles clients’ personal data, project documentation and sometimes confidential information. Here’s the summary; ChatGPT and data protection in a technical practice covers it in detail.

  • Use business plans, not free personal accounts. In many free or personal versions, what you type can be used to improve the model unless you turn it off. Business plans generally don’t use it for training and come with a data processing agreement.
  • Check where data is hosted. If a tool processes your clients’ personal data, the simplest option is hosting within the European Union and a data processing agreement with the provider.
  • Don’t upload what isn’t needed. Writing site minutes doesn’t require the client’s ID number.
  • Write a one-page usage policy. Which tools are used, for what, which data is never uploaded, and who reviews. One page is enough.
  • Train the team. Article 4 of the EU AI Act requires companies that use AI to take measures to support their staff’s AI literacy: knowing what the tool does well, what it doesn’t, and how to check what it produces. Since the 2026 amendment it’s an obligation to make an effort, not to guarantee a specific level. We explain it in our guide to the EU AI Act for architecture practices and builders.

A five-step method for your first project

  1. Pick a single task. Frequent, with a known format, and done by hand today. Site visit minutes are often a good candidate. If you work in a construction company, what to automate first in a construction company explains how to score the candidates.
  2. Measure how long it takes now. Hours per month and who spends them. Without this number you won’t know whether it was worth it.
  3. Collect five to ten real examples. Signed documents, with the notes they started from. They’re the best specification there is.
  4. Test on real cases with a person reviewing. For a few weeks, compare the result with what the practice would have done by hand.
  5. Decide with data. After a month, look at the time saved, how many corrections were needed and whether the team uses it. If it works, move on to the next task.

What it costs

It depends on how you start. For reference, these are the prices we charge at SYBRANE (excluding VAT):

Way to start Indicative cost Timeframe
Assistants on business licences A few tens of euros per user per month Days
Assessment: what to automate first From €500 1–2 days
Automating one process From €1,000 1–2 weeks
Custom tool From €5,000 4–8 weeks

To find out whether it pays off, do the sum with your own figures. If a task takes 20 hours a month and 70% of it is automated, the practice gets back 14 hours a month. Multiply that by the real cost of an hour of your team’s time and you’ll know how many months the investment takes to pay back. If the numbers don’t work, it isn’t worth it, and it’s better to know that first. We go into more detail in how much AI automation costs.

Frequently asked questions

Will AI replace architects?

No. It makes routine drafting and administration cheaper, not judgement, responsibility or the client relationship. As our founder puts it: “AI makes documents cheap. It doesn’t make the judgement of the person who signs them cheap.” What does change is the competition between practices that use it and practices that don’t.

Can I use free ChatGPT in the practice?

For trying things out, with text that contains no client data, yes. For daily work with project documentation, a business plan and a clear usage policy are better.

Do I need to know how to code?

Not for assistants with templates. To connect several tools and automate a whole process, you need someone with technical experience, either on your team or an outside provider.

Can AI check whether my project complies with building regulations?

It can help find clauses, prepare checklists and point out things to review. It shouldn’t be used to decide compliance: it makes mistakes when quoting regulations, and the responsibility stays with the architect.

Where do I start if I have no time for any of this?

With a single task: the one that takes the most hours every week. If you’re not sure which one it is, a twenty-minute conversation is usually enough to find it.