‘No machine designed your space’: generative AI as a tool for exploration, not authorship

Start with the business problem. Why does the speed of concept visualization matter commercially for a studio like yours?

Concept visualization is where projects are won and where they stall. For hospitality and workplace clients, the pitch and the early schematic phase are where a developer or a facilities lead decides whether they trust your reading of the space. Traditionally a single photorealistic render took a visualization artist days, and that cost forced a hard choice. We could show a client two or three polished directions, not twelve, so exploration got rationed. We would commit to a direction early because reworking the imagery was costly, which meant the client's first real reaction to a scheme often came late, after we had already narrowed the field. The commercial cost is twofold. Pitches are won on the strength and range of what you can show, and revision cycles eat margin on fixed-fee work. If a client cannot see a direction, they cannot approve it, and every week of ambiguity is a week the studio carries risk. Compressing that first look changes the pace of the whole engagement.

Rendering tools have existed for years. What specifically changed to make this viable now rather than five years ago?

Two things converged. The first is the quality of diffusion-based image models. Five years ago AI imagery was suggestive at best, useful for a mood board and little more. Current models produce photoreal interiors with coherent lighting, material, and depth from a text prompt or a rough sketch in seconds, at a fidelity that reads as a considered scheme rather than a curiosity. The second and more important shift is control. Early tools gave you a slot machine: you typed a prompt and accepted whatever came back. Newer methods let us condition the output on our own inputs, whether a hand sketch, a massing model, or a plan, so the image reflects our spatial intent rather than the model's guess. We can mask a single zone of a render and regenerate only that part, keeping the rest fixed, and we can tune a model toward a specific material palette or house style. Adoption tracked that maturity. By recent counts, roughly 44 percent of architects use AI for concept imagery. It has moved from novelty to a normal part of the early-phase toolkit.

Walk me through how the system actually works on a live project, step by step.

It sits inside the early design phases, not the documentation set. We usually begin from the brief and the plan. A designer sketches a spatial idea, the seating logic of a restaurant floor or the rhythm of a workplace neighborhood, and we use that sketch as a conditioning input so the generated image holds the geometry we drew. From there we run variations on material, palette, and mood: warmer timber, a darker ceiling, a different lighting temperature. Where a client likes most of an image but not one element, we mask that region and regenerate only it, which lets us tune a scheme without starting over. We pull the strongest directions into a curated board, and the ones that survive internal review get built properly in our 3D and BIM tools for accurate, buildable visualization. The AI output is a thinking and communication device, not a deliverable. It is deliberately upstream of the drawings. Nothing generated by a model goes to a contractor. It shapes the conversation about intent, and then human-authored, dimensionally correct work carries the project forward.

What about data governance, client confidentiality, and the copyright status of AI-generated imagery? Those are live concerns.

They are, and we treat them as three separate problems. On confidentiality, our rule is that client-identifying material, unreleased brand work, or a floor plan tied to a named site does not go into consumer tools that may train on inputs. We use enterprise arrangements with data-retention controls, or we abstract the prompt so it describes the design problem without exposing the client. On copyright, the position in the United States is that a purely AI-generated image, produced from a prompt alone, is not protectable because it lacks human authorship. That matters when a client expects to own what we deliver. Our answer is that a model's imagery is never the final deliverable; the drawings, the specification, and the built work carry substantial human authorship, and that is what we license to the client. We are also candid that model providers generally disclaim any warranty that outputs do not resemble existing work. So we keep AI in the exploratory phase, review outputs for anything that looks derivative of an identifiable source, and document that review.

Where do these models fail, and how do you validate output before it influences a decision?

The failure mode people underestimate is confident plausibility. A model will generate a beautiful room that cannot be built: a stair with no headroom, a cantilever with nothing holding it, a material used in a way that would fail code or a fire rating. The image looks resolved, so the error hides in plain sight. Some of this is a form of overfitting, where the output carries artifacts unrelated to the actual prompt. Our validation is deliberately human and staged. A designer treats every generated image as a proposition, not a fact, and checks it against the plan, the dimensions, and the real constraints of the site before it goes anywhere near a client. Nothing from a model is presented as feasible until someone accountable has confirmed it can be built and specified. Anything that survives is rebuilt in our accurate 3D environment, where the geometry is true. The accountable person is always the project designer, not the tool. We would rather lose a striking image than let a client fall for a room we cannot deliver.

How did the team react? Designers can be protective of craft. What went wrong first?

There was real skepticism, and it was legitimate. The fear was that generative tools would flatten the studio's voice, that we would all end up producing the same glossy, faintly generic interior the models gravitate toward. What went wrong first was exactly that. Early on, people leaned on default prompts and the work came back competent and anonymous, with none of the specificity that makes a Foldbrook project ours. The fix was not more technology. It was treating the tool as an instrument that needs a point of view. We trained the team to drive it with their own sketches and references rather than accept its defaults, and we made clear that taste, editing, and knowing which of forty options is actually right remain the job. That reframing changed the mood. Designers stopped seeing a threat to their craft and started using the tool to test more of their own ideas per hour. We also redesigned the workflow so AI sat clearly in the concept phase, with a firm handoff into human-authored documentation, rather than blurring into the deliverables.

What have you actually measured, and how should a reader weigh those numbers?

I will give figures with the caveat that they come from our own projects and are directional, not audited. The clearest change is time to a first credible visual. Work that took a visualization artist the better part of a day, we can now explore in minutes for the concept phase, though final buildable renders still take real time. Because the marginal cost of another option fell, the number of directions we put in front of a client early has risen substantially; we explore more rather than committing prematurely. And decisions come faster. When a client can see and react to a range of directions in the first meeting rather than the third, sign-off on the design direction tends to arrive sooner, which pulls the whole schedule forward. I want to be careful here. These gains are concentrated in early exploration and communication. They do not speed up documentation, engineering coordination, or construction, and treating a fast concept as a finished design is the mistake I most want to avoid.

What would you say to a skeptical peer, or to a client who is nervous that their space was designed by a machine?

To a peer, I would say the risk is not that the tool is weak but that it is seductive. It rewards accepting a good-enough image and calling the thinking done. If you let it, it will homogenize your work and let feasibility problems slip through, so the discipline has to be that a person still owns every judgment. To a nervous client, I am direct. No machine designed your space. The tools help us explore and communicate faster, but the decisions, the proportions, the materials, the way a room should feel when you walk in, are made by our designers, and everything we hand you is developed and checked by people who are accountable for it. Clients tend to relax once they understand where the line sits. They do not actually care whether an early study was rendered by hand or by a model. They care that someone with judgment stands behind the result. The imagery is a faster way to have the conversation, not a substitute for designers who know what they are doing.

Give me a concrete moment where this mattered on a real project.

I will describe a representative case rather than name a client. We were pitching a boutique hotel restaurant, and the operator was torn between two moods: a warm, textured, timber-and-brass room, and something cooler and more architectural in stone and plaster. In the old process we would have rendered one, probably committed to it, and hoped. Instead, in a single working session we generated both directions from the same floor plan, held the layout constant, and produced perhaps a dozen credible variations across materials and lighting, all conditioned on our own sketch of the seating plan. Seeing them side by side, the operator realized what they actually wanted was neither pure option but a specific hybrid: the warm materials with the cooler, quieter lighting. That insight would normally have surfaced weeks and several billed revisions later. We reached it in an afternoon. The point is not that the model designed the room. It did not. It let us externalize enough real options quickly that the client could recognize their own preference, and then our team designed to it.

Looking ahead, what does this change about the profession and the client relationship?

It shifts where a designer's value sits. When producing an image is cheap, the scarce skill is no longer visualization; it is judgment, editing, and the ability to hold a coherent point of view across the hundred options the tools can now generate. I expect the studios that struggle will be the ones that let the tool set the aesthetic. The ones that do well will use it to explore more of their own thinking and spend the recovered time on the parts of the work that are irreducibly human: understanding how a client's business actually runs, how people will move through a space, what it should feel like at eight in the morning and again at midnight. For the client relationship, the change is pace and transparency. Clients are brought into the exploration earlier and see more of the reasoning, which builds trust. My caution to the profession is to resist letting speed become the product. What we sell is considered design. These tools should buy us more room to do that work, not tempt us to do less of it.