AI has a lot of catching up to do. Where human craft still wins.
AI creative tools are genuinely useful. They are also genuinely limited. Here is where the gap sits, why it matters, and how Sant works inside it.
20 April 20268 min read
The problem with AI creative tools is not that they are bad. The problem is that they are good enough to produce outputs that look right to someone who does not know what right looks like. That is a different category of problem from a tool that is obviously poor. Obviously poor tools get caught. Tools that produce plausible outputs pass review, reach clients, reach customers, and only reveal their limitations when the audience responds in a way that was not planned for.
Nattu Adnan, a designer at LottieFiles, put the motion design version of this precisely: AI can generate animation code, but it cannot feel timing. That sentence describes the gap more usefully than most industry commentary on AI creative tools, and the gap it describes is not specific to motion design. It runs through every creative discipline that has a craft layer beneath the output layer.
Two layers in every creative discipline
Every creative discipline has two layers. The output layer is what the tool produces. A piece of animation code, a typographic layout, a brand colour palette, an editorial illustration, a paragraph of copy. The craft layer is the ability to evaluate whether the output is correct in a way that requires experience to detect.
AI is good at the output layer. It is not good at the craft layer. This distinction is the whole argument. Everything else follows from it.
The output layer is learnable from examples. Train a model on enough motion graphics files and it will produce code that runs. Train it on enough typographic layouts and it will produce arrangements that resemble good typography. The model is pattern matching against a corpus. The outputs are real. They work. They pass a non-expert review.
The craft layer is not learnable from examples alone. It is built from experience observing real outputs across real contexts over real time. A motion designer who has watched three hundred animations on actual screens, across actual products, with actual users responding to them, knows when an easing curve is technically correct but communicates the wrong weight. A typographer who has spent years in editorial production knows which element in a layout is doing nothing and should be removed. A brand designer who has worked with colour in actual print and screen contexts knows when two colours that test well in isolation fight each other at scale.
That knowledge is felt. It is the accumulated result of being wrong, noticing the wrongness, understanding why, and carrying that understanding forward. It is not the result of pattern matching a corpus of past examples. The model cannot be wrong and learn from it. The model can only produce outputs and generate more.
Where AI is genuinely ahead
Motion design, content production, brand asset generation, and editorial illustration are all areas where AI creates real value for clients right now. Pretending otherwise is defensive territory protection that does not help anyone.
Speed at the iteration stage is the clearest gain. Producing ten versions of a visual concept for review used to take a skilled designer half a day. AI can generate those ten versions in minutes. The designer's time moves to evaluation and selection rather than generation. That is a legitimate improvement in how creative work gets done.
Scale in production is the second gain. A campaign that runs across twelve platforms in three languages with five creative variants used to require weeks of reformatting, resizing, and adaptation. AI handles most of that mechanical work. The creative direction is set once. The variations follow.
First drafts are the third gain. A copywriter staring at a blank document at nine in the morning benefits from having a serviceable first draft to react to. The draft will be wrong in specific ways that the writer knows how to fix. Fixing a wrong draft is faster than generating a right draft from nothing.
All three gains are real, and all three are at the output layer. The efficiency is genuine. The creative quality of those outputs, whether they are right in the ways that matter, still requires the craft layer to verify.
Where AI is genuinely behind
Timing and weight in animation. The easing curve that runs without error is not the same as the easing curve that communicates the right intent. An element that enters too quickly feels aggressive. An element that enters too slowly loses attention before it arrives. The difference between those two outcomes is a few milliseconds and decades of trained perception. The model produces the code. It cannot tell you which curve feels right for this element, on this screen, in this context.
Hierarchy in typography. Which element leads, which recedes, and which does nothing is a judgment that requires understanding what the reader needs to know first, second, and not at all. A layout can be technically balanced and functionally wrong. The model produces arrangements. It cannot evaluate which arrangement serves the reader's actual path through the content.
Restraint in layout. What to remove is harder than what to add. A skilled art director looks at a composition and sees what does not need to be there. The model tends toward inclusion. Generating more elements is easier than identifying which elements are doing work and which are filling space.
Colour in context. Palette selection is learnable. How colours behave against specific backgrounds, at specific sizes, in specific ambient conditions, next to specific content is context-dependent in ways the model cannot fully account for. The palette that tests well in a Figma file on a calibrated monitor may behave differently on a cheap phone screen in direct sunlight.
Editorial judgment in illustration. What to show, what to imply, and what to leave out entirely is the difference between illustration that communicates and illustration that depicts. Depicting what was asked for is the output layer. Communicating what the reader needs to feel is the craft layer.
The common thread across all five is the same as Nattu's timing observation: the gap is not in what the tool produces. The gap is in detecting when what was produced is wrong.
What this means for clients commissioning creative work in 2026
The risk is not that AI produces obviously bad work. Obvious failure is caught. The risk is that AI produces work that passes a non-expert review and reaches customers before the craft layer has assessed it.
A brand video that is technically well assembled but has animation timing that feels slightly off will not generate complaints. It will generate less engagement than it should, with no clear cause visible in the data. A typographic layout that is visually balanced but hierarchy-wrong will be read out of sequence by users, with no clear cause visible in the analytics. A colour palette that tested well in isolation will look wrong in the context of the product, and nobody will be able to articulate exactly why.
These are not catastrophic failures. They are quiet underperformance, which is harder to attribute and harder to fix than a visible problem.
The practical implication for any client commissioning creative work is that AI at the production stage is an efficiency gain. AI at the evaluation stage is a risk. The evaluation layer requires craft, and craft requires experience that the current generation of tools does not have.
How Sant uses AI in creative production
AI is used in Sant's creative practice where it creates measurable value and not where it substitutes for craft judgment.
First drafts, variation at scale, reformatting and resizing, initial concept generation for review. All of these are production-layer tasks where AI improves speed without compromising quality, because the outputs go through evaluation before they reach a client.
The evaluation layer stays human. A designer or writer with craft experience assesses every output before it moves forward. Not as a compliance step but as the actual creative work. The generation is the starting point. The judgment is the job.
This is the framework Sant Grow's creative services operate from. AI tools are embedded in the production pipeline where they produce genuine efficiency. The craft layer that evaluates, selects, and refines is not automated. It is the thing clients are paying for when they engage creative services, whether they name it that way or not.
Frequently asked questions
Does using AI in creative production lower the quality of the work. It can, if AI is used at the evaluation layer rather than the production layer. Used at the production stage, with human craft judgment at evaluation, AI improves speed without lowering quality. The risk is treating AI-generated outputs as finished rather than as first drafts that require craft assessment.
Which creative disciplines are most affected by AI right now. Motion design, brand asset production, copywriting, and editorial illustration are all areas where AI tools have changed how production work gets done. The craft layer in each of those disciplines has not changed. Timing judgment in animation, hierarchy judgment in typography, editorial restraint in illustration, and voice precision in copywriting all still require experience that current AI tools do not have.
Should clients ask their agencies whether AI is being used in creative work. Yes. The useful question is not whether AI is being used but where it sits in the process. AI at the production stage with human evaluation is a workflow improvement. AI at the evaluation stage with no craft check is a risk. An honest agency should be able to answer that question directly.
Is this going to change as AI improves. The output layer will keep improving. The craft layer gap may narrow, but it narrows slowly because it requires the kind of contextual, consequential learning that comes from deploying work into real contexts and observing real outcomes. The timeline for AI closing the craft gap in any creative discipline is not short.
Closing
The tools are useful. They are also limited in specific, non-obvious ways, and the limitation matters most in exactly the situations where it is hardest to detect. An output that passes review but underperforms in the market does not announce its cause. It requires someone with craft experience to look at it, identify what is wrong, and say so.
That capacity is what takes years to build and what AI has not replicated. It is also, precisely because it is slow to build and hard to explain, the thing most likely to be undervalued in a procurement conversation. The model produces work that looks right. The experienced practitioner produces work that is right. The difference between those two things is the gap that still matters.