AI for architects saves the most time on writing, analysis and early visualisation, while the design and the professional judgement stay with the architect. A language model drafts a plan of approach, breaks down a design brief or summarises a meeting, but it decides nothing about the building.
If you use AI in a practice, you are answering two questions at once: how the output is checked, and which confidential project information goes in. Below you will find the applications, the typical mistakes, what the provider does with your input, who holds the rights and which rules apply to architects.
What AI changes in an architect’s work
What AI changes in an architect’s work is mainly the pace of preparatory and writing tasks. The four disciplines protected by the Dutch Architects’ Title Act (Wet op de architectentitel) notice this equally: architect, interior architect, landscape architect and urban planner. Who creates the design and signs off on it does not change.
Which work gets noticeably faster
Work that gets noticeably faster with AI has three features: there is a lot of text, the result is a draft and you supply the source material yourself. Structuring a tender submission, reducing a sixty-page design brief to its hard requirements and turning notes from a building team meeting into minutes all fall into this group.
Work without those features gains little from AI. A design decision, weighing the wishes of the client, the municipality and the neighbourhood, or choosing a structural principle cannot be outsourced to a model that predicts text patterns.
What stays with the architect
Responsibility for every design, every drawing and every piece of advice stays with the architect. A model has not seen the plot, has not spoken to the client and does not weigh spatial quality; it produces a proposal, not a judgement.
From my years at major banks, I bring the four-eyes principle to my AI work: nothing goes out of the door without a second check. In an architecture practice, that means a second reading of everything that goes to a client, a contracting authority or the municipality.
What types of AI tools do architects use?
Architects use three types of AI tools: language models, image generators and AI features in BIM software. Each type has its own risk profile, which is why each needs a different kind of check.
Language models for text and analysis
Language models such as ChatGPT and Claude work with text and documents: they summarise, rewrite, structure and answer questions about material you supply. A language model recognises patterns in language and uses them to generate the most probable continuation, without checking whether it is correct.
Image generators for concept and atmosphere
Image generators produce mood images and moodboards from a description within seconds, which is useful for exploring design options at an early stage or opening a conversation with a client. The way AI image generation works also explains its limit: the model combines patterns from example images and has no notion of dimensions, structure or regulations.
A generated image is therefore not to scale and not necessarily buildable. Present it as a mood image, not as a design, so that a client does not read a commitment into it.
AI features in BIM software
AI features in BIM software, the software in which a building takes shape as a digital information model, help with generating variants, detecting clashes between disciplines and analysing energy performance and daylight. The result of such a feature is only as reliable as the model and the data behind it.
An energy or daylight analysis from an incomplete model gives a precise-looking figure that says little. So check the input first, before the result ends up in a sustainability section or a conversation with the client.
What do architects use a language model for?
As an architect, you use a language model for work with a lot of text and few design choices. Three applications fit that well: tender texts, analysing a design brief and the writing around a project.
Writing tender texts
Writing tender texts goes faster when the language model supplies structure and wording and you supply the content. Give it the award criteria from the tender guidelines and have the model propose a structure for the plan of approach for each criterion, or have it sharpen your own draft for readability and fit with the question.
Always supply reference projects, key figures and names yourself. If a model does not get them, it fills the text with plausible but made-up projects, and a submission with an incorrect reference is a risk to the whole practice. Which tender information the model should and should not see is covered further down, under protecting tender information.
Analysing a design brief
Analysing a design brief with a language model works best with a targeted instruction rather than an open question. Ask for an overview of hard requirements, wishes, internal contradictions and open questions for the client, each with the page reference in the document.
Ask it explicitly to include only what is actually in the document. Without that instruction, a model fills in missing requirements with what is typical, and you end up reading a summary of a brief that is not in front of you. With the page reference, you can check every point in the original.
Writing around the project
Writing around the project is where a language model works most reliably, because you supply all the content. Notes from a building team meeting become minutes with action points, a technical explanation becomes a letter a private client understands, and a completed project gets a project description for the website.
With private clients, much of the writing revolves around fee proposals, quotation texts and the emails that go with them. A language model sets out a fixed structure for a fee proposal, rewrites an explanation of phasing or additional work in plain language and drafts a reply to a client’s question; you fill in the amounts and agreements yourself.
Give the model a role, the target audience and a sample text in the practice’s house style. For specifications, the model works from your own specification texts, not from what it generates about materials on its own.
Where AI goes wrong in architectural work
In architectural work, AI rarely goes wrong on language and often goes wrong on facts, rules and numbers. The common thread is false certainty: the text sounds expert, even where the content is wrong.
Made-up regulations
Made-up regulations are the best-known mistake: a language model cites article numbers and requirements that do not exist or are out of date. In the Netherlands, for example, the technical building rules have been in the Buildings (Living Environment) Decree (Besluit bouwwerken leefomgeving, Bbl) since 1 January 2024, the successor to the Buildings Decree 2012 (Bouwbesluit 2012), and a model trained on older texts still cites articles from the Bouwbesluit.
The solution is to give the model the regulatory text itself rather than asking about it. Paste the text of the building code article or the section of the local zoning plan into the conversation and add three instructions: answer only on the basis of the text supplied, cite the article or paragraph each answer relies on, and say “not in the text supplied” when the answer is not there.
With that source rule, a building code question becomes a reading task rather than a memory task. You can check every reference in the answer in a few seconds against the text you pasted yourself, and an answer without a source stands out immediately.
Calculations
Calculations are error-prone for a language model, because it treats numbers as text instead of calculating them. A floor area, a percentage or a sum in an answer looks just as confident when it is wrong as when it is right.
Floor areas, daylight access, energy performance and structural values therefore belong in calculation software, with a language model at most for explaining the result.
What the provider may do with your input
What the provider may do with your input depends on the account type and the settings, not on the tool itself. The same question to ChatGPT or Claude is covered by different terms on a personal account and on a business account.
Why the account type decides
The account type determines whether the provider uses your input for training and whether there is a data processing agreement. With ChatGPT, OpenAI uses conversations on Free, Go, Plus and Pro for training by default, unless you switch off “Improve the model for everyone” under Data controls; on Business, Enterprise and the API that does not happen by default. With Claude, the “Help improve our AI models” setting determines whether Anthropic uses conversations on Free, Pro, Max and Ultra for training, while Team, Enterprise and the API stay out of training by default.
With Claude, switching it off also applies to previously stored chats and coding sessions: Anthropic no longer uses them in future training runs, while training already under way and models already trained are not affected. Feedback with a thumbs up or down can still make the whole conversation available for training. Both providers offer a data processing agreement only on business accounts.
An employee who pastes project documents into their own free account is using shadow AI, because the practice then has no visibility of, and no agreements about, where that data ends up. For confidential project work, a business account with a data processing agreement is therefore the minimum.
AI in the practice’s office software
AI in the practice’s office software falls under the agreements the practice already has with Microsoft or Google, provided you work with the organisation’s work account. According to Microsoft, prompts, responses and data that Microsoft 365 Copilot accesses through Microsoft Graph are not used to train the underlying language models, and for European customers Copilot is a service within the EU Data Boundary. That boundary does not apply to the Anthropic models in Copilot: they fall outside the EU Data Boundary. Microsoft uses two separate settings for this. Claude in Copilot, such as in Copilot Cowork and Claude in Researcher, is off by default in the EU, EFTA and the UK until an administrator turns it on, while the setting for Claude in Word, Excel and PowerPoint is on by default for tenants in that region created after 25 March 2026. For Gemini in Workspace, Google states that customer data is not used for training without prior permission and that interactions stay within the organisation, with prompts covered by the data processing terms of the Workspace agreement.
The line is drawn at the account. Copilot with a personal Microsoft account or the Gemini app with a private Gmail address falls under the consumer terms, not under the practice’s agreements. A practice that already runs on Microsoft 365 or Google Workspace therefore has a business option within existing contracts, provided everyone logs in with their work account.
Protecting tender information
Protecting tender information starts with the question of which data is not intended for third parties. In the Netherlands, the DNR 2025 is the set of standard terms governing the relationship between client and architect or engineer. According to the explanatory notes of the Royal Institute of Dutch Architects (BNA), it treats as confidential all data the parties exchange in their collaboration that is not intended for third parties. The client’s documents are therefore included, as is your own bidding strategy.
Work with a business account and do not share conversations via a link, because a shared link is a snapshot that anyone with that link can open. For one-off questions, use the temporary chat in ChatGPT or the incognito chat in Claude, which do not appear in your history and are not used for training. In ChatGPT, also delete uploaded files separately from the file library (replaced by Space on Pro, Business and Enterprise since 29 September 2026), because deleting a conversation does not remove them.
Leave out bid prices, competitors’ names and your own award strategy, or replace them with a description. The model does not need that data for structure and wording.
Who holds the rights to the design and the output
You keep the rights to your own design and other input, and the provider assigns any rights in the output to you. Whether that output is also protected by copyright is a second question with a different answer.
Your own design as input
Your own design as input remains yours: the terms of both OpenAI and Anthropic leave the rights to what you enter with the user. In return, you declare that you hold the rights to that input.
For an architect, that declaration is not a formality. Drawings by a structural engineer, a design by a previous practice or material from the client are not automatically yours. Moral rights always stay with the creator, even when the exploitation rights have been transferred, as the explanatory notes to the DNR 2025 emphasise.
Copyright in AI-generated work
Copyright in AI-generated work is uncertain, because under the standard set by the Court of Justice of the EU a work must be the author’s own intellectual creation. An image or text that comes entirely from a prompt largely lacks such creative choices, and OpenAI states in its own terms that output is not exclusive to you: another user may get almost the same result.
So use AI images and texts as a starting point. The design you want to be able to claim copyright in takes shape in your own development work, and you record that work in your own sketches and files.
Which rules apply to architects using AI?
There is no separate professional rule for architects using AI. In the Netherlands, however, the statutory continuing education requirement with its associated duty to inform applies, as does the AI Act at the level of the practice. Contractually, according to NLingenieurs, the DNR 2025 contains new provisions on artificial intelligence, so a practice contracting under the DNR 2025 takes those provisions into account in its AI agreements.
Continuing professional development under the Dutch Architects’ Title Act
Continuing professional development is mandatory under the Dutch Architects’ Title Act: according to the Dutch Architects Register (Architectenregister), every registrant spends at least 16 hours a year on continuing professional development, per registration. The register explicitly names technological developments as a factor changing the profession, alongside sustainability and circular construction.
Registration also comes with a duty to inform: a registrant informs clients about their own expertise and competence, including the continuing professional development they have completed. Training in AI shows a client that the use of AI in the work rests on knowledge. The register’s continuing professional development offering also includes LearnLLM’s e-learnings on AI literacy, ChatGPT and Claude.
What Article 4 of the AI Act asks of a practice
Since 2 February 2025, Article 4 of the EU AI Act has required a practice that uses AI to take measures that support the AI literacy of its employees. Since 27 July 2026, under Regulation (EU) 2026/1744, this has been a best-efforts obligation, with no guaranteed level per employee and no certificate requirement.
The obligation rests with the practice as an organisation, not with the individual architect. What AI literacy covers determines which measures fit, and with a best-efforts obligation, what the practice can show is the only evidence of that effort.
Frequently asked questions about AI for architects
When it comes to AI for architects, the AI Act plays a role in more places than the literacy requirement alone.
Is AI in an architecture practice a high-risk system?
AI in an architecture practice is generally not a high-risk system for design and writing work, because those applications are not listed in Annex III of the AI Act. That annex names areas such as biometrics, critical infrastructure, education and employment. If the practice uses AI to filter job applications or monitor employee performance, that application does fall under employment and workers management. Such systems face stricter requirements than a language model that drafts minutes.
Do you have to label an AI mood image as AI-generated?
You have to label an AI mood image as AI-generated when it meets the definition of a deepfake. The AI Act defines this as AI-generated or manipulated imagery that resembles existing persons, objects, places or events and would falsely appear to be authentic. A realistic image of a new building on a recognisable existing site can fall under this, and Article 50(4) then requires you to disclose that the image was artificially created. For an evidently creative work, a disclosure that does not hamper the display of the work is enough.
Setting up AI responsibly in an architecture practice
Setting up AI responsibly in an architecture practice means moving from one employee’s ad hoc prompts to shared agreements. A savvy colleague saves time for themselves; only when the whole practice works to the same agreements does the result become verifiable.
An AI policy for the practice
An AI policy for the practice records three things: which tools and account types are allowed, which project data may go in and which check applies to each type of task. Internal minutes have a different risk profile from a tender submission or a letter to a client, and a policy that ignores that distinction is equally strict everywhere and is ignored as a result.
Also record in the policy who maintains it. Providers’ terms change, and an agreement on account types is only as good as its last review.
Starting small with one recurring process
Starting small with one recurring process works more reliably than a practice-wide AI strategy. Choose a process with high frequency and little variation, such as tender submissions or meeting minutes, because that is where time savings show up fastest and a mistake is easiest to catch.
During a no-obligation conversation, Dennis van de Velde gave us a clear picture of what AI can do for our firm. What stood out most was his practical view of our sector: he explained clearly how to align an AI policy with the DNR 2025, the Dutch standard terms for architects.
He also showed in practical terms how AI can help with complex questions under the Bbl, the Dutch building regulations. It was a valuable and clear advisory conversation.
AI for architects only works when responsibility visibly stays with the architect and confidential project information stays out of the wrong account. LearnLLM’s AI Literacy Course covers how a language model arrives at an answer, how to assess output and which input to protect, and the ChatGPT Course and the Claude Course apply this to each tool, each with a lesson on confidential information and intellectual property. All three e-learnings are included in the Dutch Architects Register’s continuing professional development offering.
