AI in Architecture: How Buildings Are Designed Now

From one sketch to many tested options

In a traditional workflow, an architect draws one scheme, tests it, changes it, and tests it again. Each round takes time, so only a few versions get a full check. Generative design changes the order of work. The team first states the problem in clear terms: the site boundary, the number of rooms, the height limit, the need for daylight in each room, and the budget for materials.

Autodesk describes Generative Design in Revit as a way to generate and explore more options directly in the model, then rank, filter and sort them by parameters and constraints the team defines. The team can look for a balance across all factors, or pick the option that best meets one factor, such as the most daylight or the slimmest building shape. The chosen option can then be turned into model elements for further work.

This makes the trade-off visible. If a scheme gives every room a view but uses too much floor area, that shows up. If a compact scheme saves area but leaves a room dark, that also shows up. The architect still chooses; the software only shows what each choice costs in the terms the team set.

Image tools play a different role. RIBA notes that image generators can produce striking pictures, but they lack a full sense of space. A picture has no wall thickness, no structure, and no way out in a fire. Early pictures can help a client react to a mood or a material, but they must not be mistaken for a tested design.

Sunlight, wind and energy, tested before building

The earliest checks are about the site itself. Autodesk Forma is built for this stage. Its analysis tools let a team test a proposal during early planning, before detailed drawings exist. Sun hours analysis measures how long points on the ground and on buildings receive sunlight on a given date. Daylight potential analysis gives insight into daylight performance, based on a measure called Vertical Sky Component.

Wind analysis comes in two forms in Forma: a rapid analysis that uses AI for quick, directional results, and a more detailed simulation for documentation. Autodesk says the rapid wind analysis has been benchmarked against engineering firms and academic benchmarks. Noise analysis covers road and rail traffic near a site. Microclimate analysis combines sun, wind and daylight results with local weather and climate data to map how outdoor spaces may feel.

Energy checks sit beside these. Operational energy analysis predicts the energy needed to run the proposed building, including lighting, heating and cooling, and powering appliances. Solar energy analysis identifies which facades and roofs get the most sun, and helps estimate the annual electricity a solar panel system could produce. These are estimates that depend on the data fed in, and they must be checked again as the design firms up.

For a small firm, the value is practical. A test that once waited until late in a project can now shape the first massing model: the simple block form of a building on its site. Turning a block or moving an opening can be tested in the same session, while the client is still deciding what matters most.

The shared model: catching clashes early

After the early stage, the work moves into a shared digital model. Building Information Modelling, known as BIM, is a way of building a digital version of a building that carries more than shape. Walls, floors, pipes, cables and ducts sit in the same model, so a clash between them can be found on screen instead of on site.

Autodesk describes architectural simulation as a process that starts by importing a detailed design model, setting conditions such as climate, materials and occupancy, and running scenarios. The outputs include heat maps and sun path diagrams, and comparisons between design options. Simulations can cover energy, structure, daylight, acoustics and the flow of people through a building.

The benefit is timing. Finding that a duct crosses a beam, or that a corridor is too narrow for the expected flow of people, is far cheaper on screen than after steel and concrete are in place. Autodesk also notes that simulation suits small projects, such as a single-family home or a small office, as well as large ones.

A shared model also changes how teams work. Architects, engineers and builders can look at the same version, rather than passing separate drawings back and forth. That makes disagreement concrete, because each side can point to the same model and the same test result.

Small projects, existing buildings and tight budgets

Much of the talk about AI in architecture focuses on large towers and new districts. The everyday work is smaller: a house extension, a shop fit-out, a school classroom, a clinic waiting room. Here the same tools apply in a simpler form. A sun test can show whether a new extension will shade a neighbour's garden. A daylight check can show whether a deep room will need lights on all day. An energy estimate can show whether a larger window is worth its heating cost.

Existing buildings bring a second use. A survey model of an old building can be tested for light, heat and flow before work starts. That helps a team decide what to keep, what to change, and what to leave alone. It also helps with honesty about limits: a tool can test a design, but it cannot see a damp wall, a cracked beam, or a pipe hidden behind plaster. A site visit still matters.

Cost is a real limit. Software, training and time to check results all cost money, and a small practice must choose where the effort pays. RIBA's practice writing suggests a plain test: find where work slows down because a task repeats, match the tool to that task, start with internal work where the stakes are low, and measure honestly whether time is saved or clients get better options sooner.

For a client, the useful questions are simple. Ask what was tested, what data went in, and what the result changed. Ask to see two or three options side by side, with their trade-offs stated. A tested option, with its downsides shown, is a better basis for a decision.

Limits, copyright and who stays responsible

The limits of AI in architecture are not just technical. RIBA's Horizons 2034 scan sets out four questions to watch: whether AI can gain true professional knowledge, whether it can form human-like judgement, how it fits into a business, and where professional responsibility lies. On judgement, RIBA is plain: today's systems have only early ability in this area, and they are not close to the insight of an experienced architect dealing with uncertain information, culture and ethics.

Copyright is also unsettled. RIBA notes that the legal status of training AI systems on copyrighted architectural designs or buildings is not currently settled. Much practical knowledge in architecture sits in private firm files: contracts, sketches, floor plans, sections and models. That material is not public, and feeding it into a tool without care can raise questions about ownership and privacy. A practice needs a clear rule about what goes into a tool, and what stays private.

Responsibility is the sharpest limit. RIBA states that humans still have ultimate accountability for actions, even when AI systems show strong knowledge and ability. A practitioner writing for RIBA Journal puts it in working terms: AI should handle only a small share of the volume, while thinking, strategy and judgement stay with the professional who signs the work. An unchecked output, such as a wrong material specification, can carry serious consequences for safety and liability.

That is why the best use of AI in architecture looks dull from the outside. It is a check, a comparison, and a draft, reviewed by a person who can defend it. The architect remains the person a client can question and a builder can call when the drawing does not match the site.

Frequently asked questions

Does AI in architecture design a building by itself?

No. The team sets the site, the rooms, the limits and the goals, and the software produces and ranks options within those limits. A person still chooses, checks and signs off the design, and a picture from an image tool is not a tested building.

Can a small practice use these tools, or are they only for large firms?

Small practices can use them, and Autodesk notes that simulation can suit small projects such as homes and small offices. The sensible start is one repeated task, tested on internal work first, with an honest check on whether it saves time or improves options.

What should a client ask before approving an AI-tested design?

Ask what was tested, what data was used, and what changed because of the result. Ask to see options side by side, including their downsides, and ask who checked the output and takes responsibility for it.

AI in architecture is most useful when it is treated as a testing bench, not as an author. It can show many options, test sun, wind and energy early, and catch clashes in a shared model. It cannot visit a site, judge a cultural question, or take responsibility for a building.

If you are planning a building project, start with one clear question: what needs testing first? Bring that question, and the site facts behind it, to the team doing the design. If you need help turning an idea into a clear brief and a workable plan, take a look at our services, or get in touch through our contact page.

What to watch next

Watch whether RIBA's planned guidance on AI in practice gives firms a clearer rule for what may go into a tool and who must check the output. Watch whether rapid early tests become a normal part of small-project fees. And watch whether clients start asking for tested options, with trade-offs shown, as a matter of course.

Sources

1. Autodesk — Architectural Simulation: Refining Designs Before Construction — date checked: 2026-11-05 — https://www.autodesk.com/solutions/architectural-simulation
2. Autodesk — Analysis in Forma Site Design — date checked: 2026-11-05 — https://www.autodesk.com/learn/ondemand/tutorial/analysis-in-forma
3. Autodesk — Generative Design in Revit Now Available — date checked: 2026-11-05 — https://www.autodesk.com/blogs/aec/2020/04/08/generative-design-in-revit/
4. RIBA — Architecture in the Age of AI: Four Signposts to Watch — date checked: 2026-11-05 — https://www.riba.org/work/insights-and-resources/riba-horizons-2034/horizons-2034-technological-innovation/horizons-2034-architecture-in-the-age-of-ai/
5. RIBA Journal — AI-Augmented Practice From the Inside — date checked: 2026-11-05 — https://www.ribaj.com/intelligence/ai-augmented-architectural-practice-from-the-inside/
6. RIBA — Second RIBA AI Report Shows Surge in Usage Among UK Architects — date checked: 2026-11-05 — https://www.riba.org/news/second-riba-ai-report-shows-surge-in-usage-among-uk-architects/