AI in Design Software
Mirel distinguishes easing the burdens around complex work from simplifying the professional judgement the work requires.1 This is a useful starting point for choosing what to automate.
The arrival of AI in creative software has been met with roughly equal parts enthusiasm and dread, and with good reason on both counts. Generative AI that produces images, music, or 3D assets from a text prompt is, depending on whom you ask, either a revolution or a threat. But this debate — fascinating as it is — obscures a far more useful question: what happens when you point AI at the friction around the creative work rather than at the creative work itself?
Support for automation depends on the task, the result and the control available to the user. The distinction that matters is between AI that performs the task and AI that supports the human's process. Automation research made this case long before generative AI existed: Lisanne Bainbridge's classic paper on the "ironies of automation" showed that handing a task entirely to a machine erodes the very skill and vigilance a person needs for the moments the machine gets it wrong.2 This maps precisely onto what we already know about complex software design: the goal is to reduce avoidable effort while preserving the understanding and control the user needs. The work is the point; the friction around it is not.
The Four Roles of AI in Creative Tools
There are four distinct roles AI can play in professional creative software without stepping on the creative process. Each targets a different kind of friction, and each has different design implications.
1. Doing the Otherwise Impossible
Some things are simply beyond practical human effort. Turning a 2D photograph into a spatial, navigable 3D scene. Reconstructing a complete environment from a handful of reference images. Generating plausible motion capture data from monocular video. These aren't tasks that artists used to do by hand and now don't have to — they're tasks that couldn't be done at all without AI, or required equipment so specialised that most studios couldn't justify the cost.
Neural Radiance Fields (NeRFs) are a good example. What once required expensive LIDAR rigs can now be approximated from commodity camera footage, opening up entirely new creative possibilities for independent artists and smaller studios.3 The AI here isn't replacing a workflow; it's creating one that didn't exist. This is the least contentious use of AI in creative tools, because there's nothing to feel replaced by.
The design implication is straightforward: when AI enables something genuinely new, the challenge is in making the output editable and composable within existing workflows. An AI-generated 3D reconstruction is only useful if it can be lit, textured, and rendered like any other scene element. The magic trick means nothing if it produces an opaque, untouchable result.
2. Eliminating Tedium
Every creative discipline has its version of housekeeping — the work that's necessary but creatively unrewarding. Rotoscoping in VFX. Retopology and UV unwrapping in 3D. Naming and organising layers. Tagging assets. Conforming exports to delivery specifications. These are tasks where the artist's expertise is needed to verify the result, but the process of getting there is mechanical drudgery.
This is where AI can be transformative without being threatening. Nuke's CopyCat uses supervised machine learning to accelerate rotoscoping — a task that previously consumed days of artist time per shot.4 Auto-retopology tools generate clean mesh topology and UV maps from high-poly sculpts. Autodesk's AI assistant for Maya and 3ds Max is explicitly positioned as being "less about automation and more about removing friction when you're stuck."5
The critical design principle here is that AI should do the grunt work and then hand the result back for human review. The artist must remain the authority. An auto-generated roto mask that can't be adjusted by hand is useless; one that gets you 90% of the way there and then lets you refine the remaining 10% is worth its weight in gold. This maps directly to Mirel's distinction between easing burdens and simplifying work — you're easing the burden of mechanical execution without simplifying the professional judgment required to know when the result is correct.1
The same automation research offers a useful caution here: offloading effort tends to offload engagement along with it, whether or not the task in question actually warranted less scrutiny.2 Decide the degree of automation by the consequences of error, the user's intent and the review available. Naming layers can be consequential in a shared pipeline; a colour-grading suggestion can be useful when it remains inspectable and editable.
Choosing Which Steps to Delegate
The appropriate degree of automation depends on the task and the user’s intent. Let users choose which steps to delegate, inspect the result, and continue manually where their judgement is required.
3. Filling Functionality Gaps
Professional workflows are idiosyncratic. No matter how comprehensive a DCC application is, there will always be moments where a user needs a specific tool or behaviour that simply doesn't exist. Historically, this meant writing a script, a plugin, or a workaround — or, more often, just living with the limitation.
AI can serve as a bridge here, generating bespoke tools or procedural operations on demand. Need a script that distributes objects along a spline with specific randomisation parameters? Need a shader that does something the material editor doesn't natively support? Need a batch operation that renames, resizes, and re-exports 400 assets according to a naming convention your studio invented last Tuesday? These are cases where AI acts less like a creative tool and more like a very capable technical assistant — one that understands the application's scripting API and can translate an artist's intent into executable code.
Test generated operations on a small, reversible scope before running them on the project. Show affected objects and external writes, expose failures and preserve a way back. Testing and Debugging User-Created Systems covers these safeguards.
The design challenge is in making this capability discoverable and trustworthy. An AI that generates scripts needs to be transparent about what those scripts do, and the user needs a way to inspect, modify, and save the result. This is fundamentally about maintaining the user's sense of agency — they're still the one deciding what to build; the AI is just removing the barrier between the idea and the implementation.
4. Pre-Flight and Quality Assurance
Automated pre-flight checking predates generative AI. AI may extend it, but deterministic checks remain useful for requirements such as dimensions, missing dependencies and file naming. Designers aren't accessibility experts. Animators aren't performance engineers. Compositors aren't colour scientists (well, some are, but you take the point). There are entire categories of quality concern that fall outside a creative professional's core expertise but nonetheless affect the final output.
AI-powered pre-flight systems can check a design against accessibility standards (contrast ratios, colour blindness simulation, screen reader compatibility), audit a 3D scene for performance issues (excessive polygon counts, inefficient shader graphs, texture memory overruns), or verify that an animation meets broadcast delivery specifications — all without requiring the artist to become an expert in those domains.
This is AI as a safety net, not as a creative partner. Distinguish advice from a delivery requirement. A recommendation may be dismissible; a contractual or technical requirement may need to block a particular delivery route. Explain the requirement, affected output and route to correction or authorised override. Do not present an uncertain model judgement as a definitive standards check.
The Trouble with Prompts
Whichever of the four roles AI plays, it needs an interface, and the default one — a blank text box — is usually a poor fit for professional work. A 2023 study of a Stable Diffusion forum collected 78,911 posts, read the highest-value 1,068 closely, and catalogued 25 recurring user behaviours across 8 user types.7 The pattern that stands out: serious users kept building their own tooling around the prompt box rather than living with it as supplied — tag autocomplete, prompt search, EXIF viewers to recover the prompt behind someone else's image, tag recommendation, tag-weight conversion. None of this exists because people enjoy building infrastructure in their spare time. It exists because a prompt is a leaky abstraction over the model's internals, and professionals will always route around a leaky abstraction if the capability underneath is worth having.
A prompt, in other words, is not a natural unit of professional intent. It is the model's API, handed to the user directly because building a proper interface on top of it is more work than not doing so. The fix follows the rest of this section: keep the model's input format out of the user's hands. Ask about the project, the constraints, the goal — structured questions, structured answers — and translate that into the prompt behind the scenes. The user should not need to know the word "prompt," let alone write one.
The Underlying Principle
These four roles share a common thread, and it's one that aligns with everything else in this guide: AI should reduce avoidable effort while preserving meaningful control. It should make the mechanical parts of creative work faster, make the impossible parts possible, fill gaps in expertise and tooling, and catch mistakes the user isn't equipped to catch — all while leaving the creative judgment, the aesthetic decisions, and the professional expertise firmly in human hands.
Or, to put it in terms of Larry Tesler's law of conservation of complexity8: AI can absorb complexity that the user shouldn't have to deal with, while users retain the ability to inspect, revise and take responsibility for the result.
The appropriate boundary should be established with the people doing the work. Test result quality, control and recovery as well as speed. A preference for automation in one task does not establish acceptance in another.