When AI can do the drawing

At a glance TL;DR

• A practitioner experiment shows a general AI model producing substantial architectural documentation under expert direction.
• The drawing production that supports a practice's staffing and fees can change before AI can deliver an entire architectural commission.
• If that capability becomes dependable, the advantage shifts towards organisations that can turn it into a coherent service and better buildings.

An architect does not have to be replaced for an architectural practice to change profoundly. If a substantial part of drawing production becomes something a small team can delegate to AI, the staffing, fees and divisions of work built around that production become open to revision. The profession can remain essential while familiar ways of earning a living within it become much less secure.

That is the significance of the experiment Alex Wissner-Gross highlighted in his 2 November dispatch. The original work is by architect Stephen Gallagher, who compared thirteen models on a house-documentation task. He supplied concept sketches and a previous drawing set. Claude Opus 5.5 performed best in his assessment, with the most feedback: written comments and two rounds of redlines. He reports twelve hours of model work and a few hundred dollars in compute. The model generated vector drawings through Python, without operating existing CAD software.

The released PDF makes the claim tangible. Its 23 sheets include floor plans, framing plans, elevations, sections, schedules and construction details. The first-floor plan is dimensioned and tagged; the wall sections describe assemblies and their junctions. This is a substantial attempt to communicate a building for construction. Looking at those sheets establishes the character of the output, not the correctness of every instruction they contain.

Gallagher’s judgment that the set approaches permitting quality needs to travel with his qualifications. He identifies a missing Massachusetts egress provision, and reports that changing the building length in the generated code moved some dimensions without moving the corresponding plans. The comparison used unequal human assistance. No permit approval is demonstrated.

The result is nevertheless consequential. Architecture’s exposure to AI now includes the production of its technical documents. The interesting question is what happens to a practice when that work requires much less human effort.

The unit of change is a service

Consider a small practice delivering a house. It has to understand the client and site, develop the design, resolve the building with others, prepare information and respond as construction proceeds. The RIBA’s project-stage checklists span briefing, design, construction, handover and use. Producing drawings sits within a longer relationship with a project.

Imagine that the practice gains a dependable agent capable of developing much of the documentation from an evolving design. People and agents can investigate assemblies, propose resolutions and carry decisions into the set. The principal still needs a team capable of understanding the project, but the amount of production labour needed to sustain that team changes.

There are several possible responses. The same people could undertake more projects. They could spend more time resolving difficult buildings. A principal could choose to remain small while handling work that previously required a larger office. A competing firm could offer a narrower, cheaper service. These are different businesses, even if all use the same model.

The technical advance therefore does not dictate a single employment outcome. It changes the feasible arrangements. Which arrangement wins depends on demand, the quality of the resulting work and who can persuade a client to pay for it. A practice that saves effort has gained capacity; turning that capacity into income still requires a market.

This is why reassurance that architects do more than draw is incomplete. It is true, but a business can be disrupted through one of its substantial activities. The surviving responsibilities need not support the same number of people, the same fees or the same organisational structure.

What is left between the output and the outcome?

The difficult boundary becomes visible when a design changes. Suppose a client decides to lower a floor to improve the relationship with a garden. A drawing can be updated quickly. A satisfactory response also has to work through drainage, thresholds, structure, access and the consequences for other decisions. Some information may be missing. Some answers will change the design again.

A system that handles this well does more than generate another sheet. It develops a consistent account of the proposed building, knows which assumptions support that account and can explain what remains unresolved. It can also challenge the instruction: lowering the floor may be the wrong response to the client’s underlying concern.

There is no sound reason to assume these capabilities will remain exclusively human. Agents may become very good at identifying dependencies, testing details and proposing better solutions. But the relevant test is a project that survives successive changes and scrutiny. A persuasive set produced at one moment is a smaller achievement.

For a practice, the valuable capability becomes delivering that continuity. If production grows cheaper while coordination remains expensive, savings will be limited by the remaining work. If agents also reduce the coordination burden, the range of projects a small team can carry expands much further. Our earlier essay on keeping a design project connected explores that second step. The organisational consequences depend on how far the whole process advances, rather than the speed of its most impressive operation.

This also changes what a useful demonstration would show. A revised brief, an awkward junction, a consultant’s conflicting requirement and a late substitution would reveal more about dependable delivery than another polished opening set. The necessary evidence is progress through a project, with the cost of corrections included.

Who gets to assemble the practice?

If those capabilities become reliable, established architectural offices will have opportunities. They already hold project knowledge, working relationships and experience of how buildings come together. An agent that can use those resources could make a small practice unusually capable.

The opportunity will also be available elsewhere. A contractor could bring more design development into its own organisation. A specialist could combine deep knowledge of one building type with a much smaller documentation team. A new practice could begin with assumptions about staffing and software quite different from those of an established office.

These possibilities do not require one company to own every activity. They require someone to organise a credible service around the new capability. Consultants, designers, builders and clients may remain separate while their working relationships change. The competitive question becomes who can bring the necessary knowledge together and reliably carry it into a building.

For clients, lower production costs create room for better service, but do not guarantee it. A fee reduction accompanied by more unresolved information can simply transfer work to a contractor or owner. Conversely, a practice may retain some of its efficiency gain while delivering much more investigation for the same fee. The relevant comparison is the whole commission and its consequences, including work pushed onto other people.

A healthy outcome would make expertise accessible on projects where it is currently too costly to apply thoroughly. An ordinary alteration might receive careful investigation of several difficult junctions. A small organisation might afford serious comparison of ways to adapt its premises. The benefit is greater attention to the building, rather than an expanding volume of documents for its own sake.

Expertise needs a route into the next generation

Gallagher also raises the training problem: the experience needed to supervise drawing production has historically been developed by doing it. That tension deserves more than a promise that graduates will immediately become directors of AI.

In a practice that delegates much of production, learning has to be designed into the work. A junior colleague could develop a junction, compare it with an agent’s alternative and explain the difference at review. They could follow the detail onto site and see what the drawing failed to communicate. An agent could generate variations and expose contradictions, giving a learner more cases to examine than one project would normally provide.

That could be an excellent education. It still requires time, access to real work and someone able to assess understanding. A practice that expects expertise to arrive fully formed from elsewhere is consuming a resource it is doing little to replenish. If routine production becomes less useful as the economic support for learning, practices and educators will need a more deliberate arrangement.

The aim would be to develop people who understand buildings and can work critically with increasingly capable systems. Neither preserving every old task nor skipping experience altogether achieves that.

Gallagher’s experiment is a reason to take the production shift seriously now, while testing how far it extends. The deeper change is that drawing capacity may cease to determine the size and shape of an architectural business. Once that constraint loosens, the profession’s future depends increasingly on the services it builds around the new capacity: what it undertakes, how it develops expertise and whether the buildings become better as the drawings become cheaper.

Sources

1. Alex Wissner-Gross: Welcome to November 2, 2026 Retrieved 3 November 2026
2. Stephen Gallagher: AI Can Draw Blueprints Now, 28 October 2026 Retrieved 3 November 2026
3. Gallagher experiment: Claude Opus 5.5 drawing set Retrieved 3 November 2026
4. RIBA Job Book: project-stage checklists and forms Retrieved 3 November 2026