Tool comparisons date badly. A ranked list of AI products written this year will be wrong within months, because the products change faster than the reviews. What does not change is the structure of a lighting design workflow, and that structure decides where a generative tool can be trusted and where it cannot.

The useful question is therefore not which tool is best. It is which stage of the work you are at, and whether that stage tolerates an output that is plausible rather than derived. Concept work tolerates it. Calculation does not. Here is how the stack divides.

Stage 1 — Concept and visual argument

This is where generative imagery has changed the job. General-purpose image models produce lighting concepts fast enough to explore several directions in an afternoon, which used to be a week of rendering. For a client conversation that is a genuine gain: the discussion moves from adjectives to something the client can point at.

Two practical constraints apply. First, structure must be constrained by a real image rather than described in words — feeding the model a photograph or a plan of the actual space keeps the geometry honest, and generating from a text description alone produces a room that does not exist. Second, nothing from this stage carries forward as a number. The image is an argument about intent.

Alongside the generative tools, the conventional stack still does the work it always did: SketchUp and hand sketches for massing and spatial studies, Photoshop for composition and adjustment.

Stage 2 — Render acceleration

Enscape, D5 and V-Ray in Revit have absorbed AI techniques into rendering itself, principally denoising and upscaling. This is a different category of AI from image generation, and it deserves the distinction. The physics is still solved by the render engine; the neural component removes noise from a solution that already exists, or increases resolution on an image that was correctly computed.

The result is faster iteration on renders that remain physically grounded. Where a scene previously needed a long sample count to resolve cleanly, a denoised render at lower samples arrives in a fraction of the time with the light transport intact. Nothing is invented — which is exactly why this class of tool sits comfortably inside a professional workflow.

Stage 3 — Photometric calculation and compliance

This is where the AI conversation ends, and for a structural reason rather than a conservative one.

AGI32 and DIALux EVO calculate illuminance from luminaire photometric files, room geometry, surface reflectance and maintenance assumptions. ElumTools does the same inside a Revit model, which keeps the calculation attached to the building rather than to a separate file that drifts out of date. These are deterministic solvers: the same inputs produce the same result, the result can be traced back to the inputs, and the output can be checked against related Australian Standards (e.g. AS/NZS 1680 for interior lighting) by anyone who repeats the calculation.

Every property that makes a generative model useful at the concept stage disqualifies it here. A design that must satisfy maintained illuminance targets, glare limits or lighting power density requirements under Section J7 of the NCC needs results that are calculated, traceable and verifiable.

On the electrical side, voltage-drop calculations and protection requirements under AS/NZS 3000 and AS/NZS 3008 require figures that are derived from verified inputs, not simply generated. A certifier reviewing the submission is entitled to ask how each figure was calculated.

Stage 4 — Documentation and coordination

Revit and the wider BIM environment carry the design into construction documentation, and this is where language models have quietly become useful in a limited way — drafting specification prose, cross-checking a schedule against a layout for omissions, summarising a long standard for a colleague, tidying a report.

The limit is the same one that applies everywhere else. A language model is a competent editor and an unreliable authority. It will produce a confident and wrong clause number, and standards are precisely the material where a plausible error is most expensive. Anything it drafts that touches a compliance statement gets verified against the standard itself before it leaves the office.

The stack

How to assess a tool you have not used before

Three questions settle it quickly, and they will still work when the current products have been replaced.

Can the output be traced to its inputs? If changing the ceiling height changes the result in a way you can follow, the tool is calculating. If it changes the result in a way nobody can explain, it is generating.

Would you put the number in a submission? This is the practical version of the first question. Anything destined for a certifier, an energy assessor or a construction drawing has to survive being asked where it came from.

What does it cost when it is wrong? A concept image that misses is discarded in the meeting. An illuminance figure that misses is discovered at handover, and by then the ceiling is closed.

What we do not let AI touch

We do not allow AI to independently produce or approve final lighting calculation reports, lighting and electrical layouts issued for pricing, coordination or construction, final assessment and compliance reports, documentation requiring professional accreditation or certification, or any material intended—or potentially intended—for legal or regulatory purposes. AI may assist with limited drafting and checking tasks, but every final document is prepared, reviewed and approved by a suitably qualified professional.

FAQ

SL Design Consultant is an independent lighting and electrical design consultancy in Sydney, delivering lighting and electrical design across commercial, hospitality, retail, heritage and public projects. Our product-neutral recommendations are guided by design intent, technical performance and project requirements.