A Full Project with AI

The project: a client wants a compact three-bedroom house on a tight urban plot, budget-conscious, climate-appropriate, with a small courtyard they've always wanted. That is the brief in a sentence, and the first AI touchpoint is right here - but it is not an image tool.

At the brief stage, a language model earns its place as a thinking partner, not an author. You can hand it the client's rambling wish-list and ask it to organise the requirements into a structured brief, surface questions the client hasn't answered (orientation? setback rules? who uses the courtyard?), and draft a first area programme you will then correct. The value is speed and completeness on the boring scaffolding; the judgement of what the house should be stays entirely yours. Crucially, everything it produces is a draft to edit - you verify every figure, every code assumption, every area, because the model will state a wrong setback with the same confidence as a right one.

At the concept stage the image tools enter. With the programme roughed out, you reach for the explore slot - Midjourney or an aesthetic tool - and generate atmosphere fast: twenty courtyard-house moods in an hour, testing whether the house wants to be inward and shaded or open and light, whether the courtyard reads as calm or green, what material world it lives in. This is where AI is genuinely transformative: it collapses days of reference-hunting into an afternoon of directed exploration. But note what the tool is doing and not doing. It is multiplying options; it is not choosing. Twenty moods are worthless until a designer looks at them against the site, the budget and the client and says 'this one, because'. The generation is AI; the selection - the actual design act - is human. That division will hold all the way through.

A chosen concept is not a building. Now the work turns from exploring to resolving, and the AI's role changes with it - from free invention to disciplined obedience.

You draw. The plan, the section, the massing - these come from your hand and your training, not from a prompt, because this is where geometry, structure, circulation and code live, and none of that is something an image model can be trusted to invent. A floor-plan generator might spit out a plausible-looking layout, but it does not know your setback, your client's grandmother's ground-floor bedroom, or the load path - so it stays a sketch-provocation at most, never a drawing you develop. This is the first hard boundary of the development stage: the resolved geometry is yours.

Where AI re-enters is visualisation of what you've resolved. Once you have an approved massing and plan, the obey slot - Stable Diffusion with ControlNet, or a wrapper like DesignAI that packages that control - lets you render your actual geometry into atmosphere: feed the massing or a line elevation as a condition, and get back a photographic study that honours your proportions instead of inventing new ones. You test the courtyard in warm evening light and flat monsoon light; you try the facade in laterite versus render; you check how the shaded verandah reads. Each of these is a fast, controlled experiment on a design you own, and the control is what makes it useful - an uncontrolled pretty render of a different building would be worse than no render at all, because it would quietly lie to you and the client about what you're building.

This is also where the roundtrip discipline from Module 7 pays off. The AI render is a study, not a source of truth; the drawing remains authoritative. If the render reveals a proportion that's wrong, you fix it in the drawing and re-condition, never the other way around. AI accelerates the seeing; it does not get a vote on the geometry.

The house is resolved and drawn. Now it must be presented - to the client, and perhaps published - and the AI touchpoint here carries a different priority: not beauty or control, but safety and honesty.

The hero renders that go on the board or into a proposal are client-facing deliverables, so the sell slot applies: reach for a commercial-safe tool like Firefly, or licensed sources, for anything you present or publish, because a rights problem in an image you've effectively sold is your exposure, not the tool's. You condition these final renders on your resolved geometry exactly as in development, but you produce them in the safe tool and keep your source images clean.

And you disclose. The professional standard - the one Module 9 argued for at length - is to tell the client that AI assisted the visuals, to be clear about what is a photograph of a real material and what is a suggestion, and never to present an AI render as a promise of an exact built outcome. A render is an atmosphere study, not a contract; the drawings are the contract. Said plainly, this disclosure builds trust rather than eroding it - clients increasingly know AI exists, and a designer who is straight about using it well reads as more credible, not less.

Stand back and look at the finished board. The concept moods that opened the client's imagination, the controlled development renders that tested the real scheme, the safe hero images that sell it - AI touched all of them, and the project moved faster and communicated better for it. But the house - the plan that works on that specific plot, the section that shades that courtyard, the material choices that suit that budget and climate, the drawings you'll stamp - is yours, start to finish. That is the shape of an AI-augmented practice: the pictures got faster and richer; the architecture stayed human.

It's worth naming, explicitly, the places in this project where AI had no role at all - because that map is as much a professional skill as knowing where it helps.

AI did not decide whether the concept was any good. It generated options; the judgement of which one deserved development was a human act of taste, experience and fit-to-brief that no model performs. AI did not resolve the structure, check the building against code, size the services, or work out the fire strategy - the entire technical spine of the building that makes it safe and legal to occupy. AI did not manage the client relationship, read the room in a meeting, or absorb the responsibility that comes with a professional stamp. And AI did not draw the construction documents that a contractor will build from and that you are legally answerable for.

Notice that these aren't arbitrary exclusions - they cluster. AI stayed out of exactly the places where consequence and judgement live: where being wrong hurts someone, where a decision expresses a value rather than an average, where a human must be accountable. It entered freely where the job was to generate and accelerate, and it stayed out where the job was to decide and be responsible. That line - generate versus decide, accelerate versus be accountable - is the single most useful thing to carry out of this whole course. Tools will get better at generating; the demo reels will get more astonishing. But the part of the work that is yours - the deciding, the judging, the being answerable - is not something you are waiting for AI to take over. It is the job. AI just cleared the busywork so you can spend more of your time on it.