The Impact of AI on Product Design

AI product design: where AI helps across stages, what generative models add, where they fail, and how teams validate output before production.

The Impact of AI on Product Design
Written by TechnoLynx Published on 24 Jun 2024

AI adoption in product design is most mature in the implementation stage, where optimisation, simulation support and geometry generation draw on available CAD and simulation data. It is exploratory in early concept work and thin in integration and validation. Generative models add fast, varied options, but their output is probabilistic: it can look plausible and still be physically impossible. Before any of it reaches production it has to be checked for topology, manufacturability and physics, and an engineer has to sign it off.

This article is about physical product design: objects that get moulded, machined or printed. Research on designers and user research is used where it applies.

How AI is used in product design today

A 2025 systematic review of data-driven methods in mechanical and mechatronic design found that machine learning and statistical methods dominate, while deep learning is less common but adopted more each year. The applications are practical ones: performance prediction, anomaly detection and design optimisation, used in both early and late development stages. In the system design stage, evaluating concepts is by far the most common use.

The generative side takes several forms. In generative design, the engineer encodes structure requirements into software and the AI evolves optimal designs. NASA Goddard’s Evolved Structures work frames this as the engineer defining requirements and objectives while AI generates compliant, optimised designs. Deep generative models (feedforward networks, GANs, VAEs and some deep reinforcement learning frameworks) have shown promising results in structural optimisation, materials design and shape synthesis. Further upstream, text-to-image tools such as Midjourney, DALL-E and Stable Diffusion produce concept visuals and moodboards, and designers use language models to get up to speed on a topic and generate ideas.

Which product design stages gain most from AI?

Implementation gains most, early design gains in an exploratory way, and validation remains the least addressed stage. The 2025 review sorted its 128 papers by stage (52 on system design, 30 on implementation, 26 on integration, 20 on validation), so publication volume and the variety of methods both fall from left to right across the V-model. Volume is not maturity, though. The review judges adoption most mature in implementation, because simulation and CAD data are available there. Early design attracts more papers, but that work is mostly exploratory and held back by a lack of structured design data and by the difficulty of linking model outputs to design decisions. Validation is conservative: it relies mainly on classical statistics because they are interpretable and easy to fit into risk-sensitive work.

Stage What AI does well there What limits it Maturity
Research and requirements Synthesising existing data, preparing studies, forming hypotheses Synthetic users cannot produce behavioural data and tend to please Complement to real research
Concept and ideation Large volumes of ideas and visual options; concept evaluation Unstructured data, weak link to decisions, and selection becomes the bottleneck Exploratory
Implementation (CAD, simulation, optimisation) Geometry generation, surrogate models, optimisation with CAD and simulation data Single-part focus and limited scope of analysis Most mature
Integration Predicting system behaviour, failure likelihood and remaining useful life Fewer studies, mixed methods Emerging
Validation Surrogate modelling, interpretable classification, test reduction Data scarcity, scalability, interpretability, domain specificity, compute Least addressed

The pattern makes sense in engineering terms. A model is only as useful as the data it learns from, and the implementation stage is where CAD and simulation data are already available. Concept work lacks available data, and the review found unsupervised methods (clustering, dimensionality reduction) used most in early design.

What generative AI adds to product design, beyond chat models

Generative AI is broader than large language models. The term originally meant models that generate signals such as images, audio and 3D assets, and treating it as a synonym for LLMs is a recent, market-driven confusion. VAEs, GANs, diffusion models and reinforcement learning frameworks have all been applied to design problems. Reinforcement learning needs no dataset, but it depends on reliable reward signals, which often means a high-fidelity simulation. GANs and VAEs, by contrast, can only use differentiable objectives.

A 2024 Scientific Reports study combined a VAE with reinforcement learning for cultural and creative product design and reported 94.5% accuracy, against 92.3% for a VAE alone and 87.0% for a GAN. In the same study, GPT and Llama-3 were slightly weaker at tasks that combine visual and textual data, and slower to train. In ideation, Wharton researchers found that one person working with GPT-4 produced about 200 ideas in 15 minutes and was roughly 40 times more productive at generating consumer-product ideas than a person working alone. They also found that 35 of the top 40 ideas, judged by purchase intent, came from the model rather than the students, in a study of consumer-product ideas.

Toyota Research Institute has said that generative AI tools serve designers as inspiration but cannot handle the engineering and safety considerations of actual car design. Its technique feeds constraints such as drag and chassis dimensions into text-to-image generation, which cuts the iterations needed to reconcile styling with engineering. In the Sketch2Prototype work, putting text between the sketch and the 3D model produced more diverse and more manufacturable models than going straight from sketch to 3D. The text step also gives the designer a place to add feedback. A review of deep generative models in engineering design argues that copying the training data is often not enough for engineering, and can even work against it. The authors propose performance, novelty and constraint adherence as training objectives instead.

Where generative AI stops being useful

Generative AI tools face decisive limitations at the point where a design has to converge. In the same VAE plus reinforcement learning study, the model’s work was often taken for human work in modern art and digital illustration, but it was weaker on traditional craft and product design. The industrial and UX designers interviewed for a 2024 DIS study (Uusitalo and colleagues) said much the same: these tools work in divergent phases, but they lack the precision and control needed to iterate on converging designs, such as product architecture.

Three costs come with the capability:

  • Data. Data sparsity is one of the biggest problems in data-driven design, and many design domains lack public datasets of any useful size.
  • Compute. The VAE plus reinforcement learning model’s heavy computational demands may rule it out for resource-constrained teams.
  • Novelty. Most deep generative models learn to imitate the parts of the design space that are already explored, which works against creative designs. The Wharton study rated human ideas as more novel than GPT-4’s.

How much does AI cut physical prototyping and iteration time?

This section draws on two examples: structural generative design at NASA and one production deployment of AI in retail. Each has a clear scope.

In structures, NASA compared expert-human and AI designs of the EXCITE tip/tilt bracket. Two expert humans (a designer and an analyst) took two days to develop theirs. The AI designs took about an hour each and were stiffer, stronger and easier to manufacture. NASA reports that combining generative design with digital manufacturing cuts structure development time from months to days, with one example showing more than a tenfold reduction in development time and cost. Its 2025 guide says simple parts with known requirements can be designed and FEA-validated in as little as a day by an experienced user, and that new designs are much faster to regenerate when requirements change. The scope is spaceflight structures and simple parts with known requirements, not every kind of product.

In retail, Alibaba deployed a text-to-image system that lets merchants design fashion items and generate photorealistic product images before anything is made. Production starts only after an item has received enough orders, which greatly reduces the need for physical prototypes. In a six-month online A/B test across 17 product categories, the AI-generated items had a 7.9% lower return rate than human-designed items in the same stores. Production cycles ran from 3 to 20 days, with a mean of 7. Expert review still found room for improvement in craft design. The result covers fashion items sold online, where the generated images served as blueprints for manufacturing. The wider set of retail techniques is covered in the AI innovations behind smart retail.

Will AI replace product designers, or change what they spend time on?

The sources describe AI changing designers’ work rather than replacing them. A CoLab Software guide for mechanical engineers puts the shift plainly: generative and text-to-CAD tools move engineers from authoring geometry to auditing it. The Wharton authors suggest that the critical task in innovation may shift from generating ideas to evaluating and selecting them, and they note that language models are not yet well suited to that task. The designers in the DIS study expected efficiency gains and the automation of tedious work, with human time redirected to sensemaking. One industrial designer said that innovative product work still needs human understanding of emotional connection and how bodies behave.

In Figma’s 2025 survey of designers and developers, 78% agreed that AI significantly improves the efficiency of their work, but only 32% said they can rely on its output. Autodesk’s 2025 State of Design & Make report found that 69% of professionals were confident in their company’s AI decisions, nine points lower than the previous year. Autodesk also found that a lack of skilled talent was the barrier named by 58% of respondents, up 15 points.

The useful question is not “can AI out-design us?” but “which repeatable task can AI take off our hands?”. Expecting a model to be smarter than the team is a different proposition from handing it a repeatable task. Handing over a defined, menial task such as generating bracket variants against encoded loads is far more likely to show measurable return. A field study of consultants using GPT-4 shows why the capability boundary matters. On tasks inside AI’s capability, consultants completed 12.2% more tasks and worked 25.1% faster, but on a task chosen to sit outside it, they were 19 percentage points less likely to reach a correct answer. The same study found that AI-assisted ideas were higher in quality but less varied. For a design team, that is a reason to keep people in charge of divergence.

Where AI-assisted product design fails

AI-assisted product design fails in recognisable ways, many of them because the model knows shape but not physics or process.

  • Hallucinated geometry. Probabilistic models produce shapes that are mathematically plausible but physically impossible or very weak. Models trained on visual datasets tend to favour how a surface looks over whether the volume is sound.
  • Lossy image-to-3D conversion. The Sketch2Prototype authors found image-to-3D to be the weakest stage of their pipeline. Details were lost, and meshes were often rough, fragmented or impossible to manufacture, needing postprocessing that still left holes and uneven surfaces.
  • Text-to-CAD failure cascade. The 2026 MUSE benchmark found that models drop off in stages: from code that runs, to valid geometry, to an engineering-ready design. Even the strongest models had limited success on fine-grained engineering criteria.
  • Out-of-distribution inputs. A data-driven design model may fail to produce decent solutions when the new task falls outside its training data.
  • Literal optimisation. Optimising for a vertical load can remove all the material that resists torsion. Lattice-to-skin transitions are often left without proper radii and become stress risers.
  • Missing process and material context. The model knows shape, not whether a part will be machined, cast or printed. That shows up as zero-draft walls and pockets that tools cannot reach. Material is often treated as an abstract variable, and generated designs may default to a factor of safety of 1.0.
  • Misleading confidence. A model’s confidence score cannot be relied on as a measure of its output’s quality.
  • Synthetic user research. Nielsen Norman Group found that synthetic users tend to please, care about everything equally and cannot produce behavioural data, so using them to test concepts is risky.

Human failure modes matter too. In the Harvard Business School study, the professionals who performed worse with AI tended to adopt its output without question.

Checking AI-generated designs before production: a validation sequence

This sequence draws on the CoLab Software framework and NASA Goddard’s Evolved Structures guide. Adjust the thresholds to your own process and standards.

  1. Audit the topology. Check for non-manifold artefacts such as T-junctions (edges shared by three or more faces) and bowtie singularities, which confuse slicers and crash CAM software.
  2. Review manufacturability for the intended process. Flag faces with less than 0.5° of draft, and pockets that would need a tool length-to-diameter ratio above 5:1. Generative parts often need edits to reduce overhangs and clean up artefacts such as voxel print-through.
  3. Validate the physics with detailed FEA. The coarse voxel model used during optimisation misses localised stress peaks near interfaces. Rerun FEA after every CAD edit, and converge the mesh on the load case with the lowest margin. Apply off-axis loads, check stress concentrations, and run a modal analysis aiming for at least 2x separation from operating frequencies.
  4. Apply material and compliance factors. Calculate margins of safety from material allowables. NASA-STD-5001 generally uses 1.4 for ultimate and 1.25 for yield. For additive parts, apply a surface-roughness knock-down to fatigue calculations.
  5. Check design-specific metrics, not just similarity. Score outputs on constraint satisfaction, functional performance and novelty rather than statistical likeness to the training set.
  6. Formalise human review and sign-off. Keep a system of record for every AI-generated model considered for production, not email chains. Review tolerances, assembly constraints and serviceability.
  7. Plan the inspection. Consult the inspectors when drawing the part. Spot-check organic shapes with a CMM. For additive parts, add internal flaw inspection such as CT where needed.

For consumer-facing concepts, the check involves real people. Purchase intent from about 15 people in the target market is a reliable measure of idea quality, and Nielsen Norman Group advises treating synthetic-user findings as hypotheses to test with real users.

Where we would point AI first in a design programme

We would start where the evidence is strongest and the risk is easiest to manage: parts with clear, encodable requirements, such as NASA’s bracket that went from two days to about an hour. Concept ideation comes second. It is cheap to try, but its value depends on whether the team can evaluate options faster than the model generates them. Validation is the least addressed stage and still relies mainly on interpretable statistical models. Before investing, every team should settle one question: does your bottleneck sit inside the current capability frontier, or are you hoping the model will supply judgement your process does not yet have?

Back See Blogs
arrow icon