Where AI Fits in the Design Process

Every mature practice runs on a staged process - call it concept, schematic design, design development and construction documents, or the RIBA and AIA equivalents. The stages are not bureaucracy; they exist because a building resolves from loose to fixed, and each stage carries a different tolerance for error. A concept can be wrong ten times an hour and lose nothing; a construction detail wrong once can hurt someone. Generative AI does not dissolve this structure - it slots into it, and it fits some slots far better than others.

The single most useful idea in this whole module is that AI's value is inversely related to how much precision a stage demands. Early, when you are searching for a direction and everything is provisional, AI's fluency and speed are pure upside - it hands you twenty starts before lunch and none of them has to be right. Late, when the work is about dimensions, coordination and code, that same fluency becomes a liability, because the model produces confident images with no underlying logic (the appearance-without-logic problem from Module 0.1). The picture looks resolved; nothing behind it is.

So the map is simple to state and hard to hold to under deadline pressure: use AI hardest where the work is still soft, and pull it back to a supporting role as the work hardens. The rest of this lesson makes that concrete stage by stage - and names the line AI must never cross.

This is where AI belongs, and where it genuinely changes how fast you think. In concept, the job is divergence - generating many possible directions so the good one has competition. AI is a divergence engine: feed it your intent and it returns a spread of moods, massings and material worlds in minutes, far more than you could sketch by hand. The point is not that any single image is the answer; it is that seeing twenty framings sharpens your sense of which one you actually want. AI is a conversation partner for the fuzzy front end.

In schematic design, the work shifts from 'what could this be' to 'let us test this direction' - and AI keeps helping, now more precisely. You can visualise a scheme under different light, try a facade in three materials, show a client two atmospheres for the same plan. Here AI is a fast rehearsal of decisions you will later make for real in your CAD or BIM model. Crucially, at this stage the image is still exploratory - nobody is building from it, so its lack of buildable logic costs you nothing.

The reason AI fits these stages so naturally is that their currency is ambiguity, and ambiguity is exactly what a probabilistic image model trades in. A concept sketch is meant to be loose; the model's cheerful invention is a feature, not a bug, because you are looking for provocation, not a spec. The skill at this stage is not restraint but volume and judgement: generate widely, then choose ruthlessly against your intent. Lesson 7.2 is about that judgement - keeping the seductive image from quietly rewriting the brief - but the placement is settled here: concept and schematic are where AI earns most of its keep.

Now the line. As a project moves into design development and construction documents, the work becomes dimensions, clearances, structure, services coordination, code compliance and the contract set that a professional stamps. This is the register of record - the drawings someone builds from and someone is liable for. Generative image AI has no business authoring any of it, and the reason is not caution for its own sake: the model does not know what it is drawing.

It has never held a load, never resisted a fire rating, never checked a door swing against a clearance. It produces a picture that looks like a resolved detail by imitating the surface statistics of ten thousand real ones, with none of the reasoning that made them correct. A rendered stair with treads that would fail code, a facade with mullions that carry no load, a plan that reads well and coordinates with nothing - these are not bugs to be prompted away; they are what the tool fundamentally is. Trusting it here is trusting confident fiction with life-safety consequences.

That does not mean AI vanishes at these stages - it means its role narrows sharply to visualisation of a design you have already resolved elsewhere. Once your BIM model says the building is right, AI can dress a view of it beautifully for a client or a competition board (Module 7.3 covers that roundtrip). The distinction is absolute and worth memorising: AI may illustrate the record; it may never be the record. The moment an AI image starts standing in for a drawing, a schedule or a coordination check, the process has been broken and someone is about to build from a hallucination.

The governing principle across every stage is a division of authority: AI proposes; the designer disposes. The model is allowed to suggest, provoke, visualise and accelerate. It is never allowed to decide. Every output passes through a human who owns the judgement - and, critically, owns the accountability, because no client, code official or court accepts 'the AI chose it' as an answer.

Making this real in practice takes two small disciplines. First, keep AI outputs clearly labelled and separated from your design record - a concept board is not a drawing set, and the two should never sit in the same folder as equals. When you carry a direction forward, you re-make it deliberately in your real tools, so the record is authored, not pasted. Second, keep a light decision log: when an AI image genuinely shifted a decision, note what you took from it and why. This keeps you honest about which choices are yours and defensible, and it is exactly the trail a reviewer, a partner or a client may later ask for.

There is also a professional-trust dimension here that Studio Matrx takes seriously. Clients are increasingly aware of AI, and the fastest way to lose their confidence is to hand them a photoreal 'design' that turns out to be an ungrounded render. The fastest way to build confidence is the opposite: use AI visibly for exploration, then show them the resolved, coordinated design it helped you reach. Used this way, AI is not a threat to your judgement - it is an amplifier for it, and the judgement stays unmistakably, accountably yours.

Put it together on one small project - a hillside weekend house. At concept, you write a one-line intent ('a quiet laterite house that steps down the slope and opens to the valley') and generate thirty images across three moods. None is 'the design'; together they tell you the stepped, open-to-valley direction is the one worth pursuing. Twenty minutes, and you have chosen a direction with evidence rather than a hunch.

At schematic, you sketch the stepped plan properly in your own hand or in Rhino, then use AI to test it: the same massing in laterite versus board-formed concrete, in morning versus evening light, as a client would experience arriving. You show the client two of these to confirm the material direction. The decisions are yours; AI just let you see them fast enough to choose well.

At design development, the work moves into your BIM model - real dimensions, the actual structure of those stepped floors, the retaining wall the slope demands, the code check on the stair. AI does none of this. When you want a hero image for the planning submission, you export a view from the resolved model and use AI to render it atmospherically (Lesson 7.3). At construction documents, AI is essentially absent from the record entirely; the set is drawn, coordinated and stamped. Trace that arc and the rule is unmistakable: AI was loudest when the house was an idea and quietest when it became a building. That shape - loud early, quiet late, always proposing and never deciding - is the whole of fitting AI into real practice.