For most of the history of digital product development, one of the biggest constraints on innovation has been our ability to make things.
Ideas needed designers. Designs needed developers. Products needed months of planning, prototyping, building, testing and refinement. Even relatively simple changes had a meaningful cost attached to them.
AI is starting to dismantle that constraint.
The economics are changing remarkably quickly. Stanford's AI Index found that the cost of running an AI model performing at roughly GPT-3.5 level fell from $20 per million tokens in November 2022 to $0.07 by October 2024; a more than 280-fold reduction in around 18 months.[1]
At the same time, generative AI is finding its way into almost every part of the product development process.
We can generate concepts in seconds. Prototype experiences in hours. Produce functioning code increasingly quickly (to an extent). Create hundreds of variations of content. Analyse enormous datasets. And build products capable of adapting their behaviour to individual users.
The cost of exploring possibilities is collapsing.
And that creates an interesting problem.
When we can make almost anything, deciding what is worth making becomes much more important.
Which brings us back to human-centred design.
The problem with human-centred design
Human-centred design has become one of those phrases that is difficult to disagree with.
Of course we should understand users.
Of course we should solve real problems.
Of course products should be useful, accessible and intuitive.
But somewhere along the way, the principle became confused with the process.
Human-centred design became synonymous with workshops, personas, stakeholder interviews, journey maps, research reports and walls covered in Post-it notes. Entire projects could spend weeks documenting the world before anybody actually changed it.
Some of that was necessary. Some of it probably wasn't. And AI should force the design industry to confront that.
There is also an uncomfortable possibility that some of what we've historically called human-centred design wasn't particularly human-centred at all. It was organisation-centred design dressed up in research: long discoveries shaped around procurement processes, artefacts created because a methodology expected them, and validation exercises designed to give organisations confidence in decisions they had largely already made.
AI won't fix that. But it might make the inefficiency harder to justify.
If a research team can use AI to interrogate thousands of support conversations, reviews, analytics events and research transcripts, should people still spend weeks manually categorising every observation?
If a designer can create ten functioning prototypes in the time it previously took to create one, should we spend two weeks debating which idea deserves to be prototyped?
If we can put functioning software in front of users almost immediately, should a static representation of that software always come first?
Probably not.
The future of human-centred design shouldn't be about protecting the processes we've become comfortable with. It should be about protecting what those processes were supposed to achieve.
Human-centred doesn't have to mean human-powered
There is a strange contradiction in some of the conversation around AI and design.
We talk about keeping humans "in the loop", while continuing to ask those humans to perform enormous amounts of work that machines are becoming very good at.
That isn't necessarily human-centred. It's just inefficient.
There is already evidence that AI can materially improve productivity in some types of work.
A large study of more than 5,000 customer-support agents found that access to a generative AI assistant increased productivity by around 14% on average. Interestingly, the effect wasn't evenly distributed. Less experienced and lower-skilled workers improved by around 34%, while the impact on highly experienced workers was minimal.[2]
An earlier controlled experiment involving software developers found participants using GitHub Copilot completed a defined programming task 55.8% faster.[3]
But the story isn't as simple as AI makes everyone faster.
A later randomised study by METR looked at experienced open-source developers working on real repositories they already knew. Developers expected AI to make them faster.
Instead, using early-2025 AI tools made them 19% slower.[4]
That contradiction is important. AI doesn't automatically create productivity. The task matters. The expertise of the person using it matters. The quality of the system matters.
And, perhaps most importantly, how we design the relationship between the human and the machine matters.
That itself is a human-centred design problem.
Execution is becoming abundant. Judgement isn't.
Imagine asking an AI system to redesign a customer onboarding journey.
Increasingly, it could analyse the existing experience, identify friction, generate alternative flows, write the interface copy, create the UI, produce the code, implement the analytics and potentially begin optimising the experience based on what happens next.
That's extraordinary. But there's a more difficult question:
What should it optimise for?
The shortest journey?
The highest conversion rate?
The greatest amount of customer data captured?
The lowest operational cost?
The highest customer lifetime value?
The fewest calls into the contact centre?
Because those aren't simply design problems. They're decisions. And decisions require context.
A business has commercial objectives. Customers have needs, expectations and anxieties. Technology has constraints. Organisations have operational realities. Society has expectations around privacy, accessibility, fairness and trust. The "best" experience depends on how those things are balanced.
AI can help us understand that landscape. It can't absolve us of responsibility for deciding what to do about it.
We're moving from designing interfaces to designing behaviour
There is another, potentially bigger change happening. Most digital products we've designed over the last 30 years have been relatively deterministic.
A user clicks something. The system responds.
A user fills in a form. The system processes it.
Designers could map the journey because, broadly speaking, we knew what the system was going to do. AI changes that relationship. Increasingly, digital products won't simply present interfaces - they will interpret intent.
Recommend.
Generate.
Predict.
Decide.
And, increasingly, act.
A banking experience might not ask a customer to navigate through six screens to move their money. An intelligent system might understand what they're trying to achieve and orchestrate the process for them.
A commerce experience might not require somebody to browse hundreds of products. It might understand the problem they're trying to solve and assemble an appropriate solution.
A business application might not require employees to learn a complicated interface at all. They might simply describe the outcome they need.
In that world, designing the screen is only a small part of designing the experience.
We need to design how the system behaves.
When should it make a decision?
When should it ask?
When should it explain itself?
When should it remember something?
When should it forget?
When should it act autonomously?
When should it hand control back to a human?
And what happens when it's wrong?
These aren't hypothetical UX concerns. Microsoft's human-AI interaction research identified many of exactly these problems years before the current generative AI boom: setting appropriate expectations, making clear what a system can do, allowing users to correct it, explaining why it behaved in a particular way and giving users meaningful control over its behaviour.[5]
The technology has changed dramatically whilst the human problems haven't.
Trust becomes part of the interface
Traditional usability was often about making technology easier to operate.
AI introduces another requirement: making technology easier to trust appropriately. Not blindly trust, but appropriately trust. That distinction matters.
If an AI system recommends a restaurant and gets it wrong, the consequences are relatively small. If it recommends a financial product, rejects an insurance claim, prioritises a job applicant or influences a medical decision, the relationship changes dramatically.
The US National Institute of Standards and Technology's AI Risk Management Framework identifies characteristics including reliability, safety, transparency, explainability, privacy and fairness as fundamental considerations in trustworthy AI.[6]
Those aren't simply technical requirements. They're experience-design requirements too.
A technically explainable system isn't necessarily understandable to the person affected by its decision. A system can expose a confidence score without somebody knowing what that confidence score actually means, and a user can technically have control while having no realistic idea when they should exercise it.
Design therefore has a new responsibility. It isn't simply about making AI feel easy to use.
It's about helping people build an appropriate mental model of what the system knows, what it doesn't know, and how much authority they should give it.
In some cases, the most human-centred thing an AI can do might be to admit that it doesn't know.
Personalisation changes what a "user experience" means
There's another assumption AI starts to challenge: that everybody needs to experience the same product.
We've traditionally designed experiences for groups: segments, personas and audiences.
We might personalise some content or change a recommendation, but the underlying product remains broadly the same. Generative systems make something much more radical possible.
The interface itself can become adaptive. Information can be presented differently depending on someone's expertise and complexity can expand or disappear depending on what somebody is trying to achieve. Journeys can be generated rather than predefined.
An interface might increasingly become something that is assembled in response to an individual rather than something designed once and served to everyone. At first glance, that sounds like the ultimate expression of human-centred design.
But it creates a paradox.
How do you design an experience that you can no longer completely predict?
The answer probably isn't more wireframes.
It's about defining how the system should behave, the constraints it should operate within and what happens when it gets things wrong.
The designer moves from specifying every possible interaction to defining the conditions within which good interactions can happen. That's a significant change in the discipline.
The interface might not be the product
There's an even more fundamental possibility.
For decades, improving a digital product usually meant improving its interface. Better navigation, better information architecture, fewer steps, clearer calls to action, better forms.
But increasingly, the best interface might be less interface.
If an AI system genuinely understands intent, why should someone navigate your organisation's internal structure? Why should a customer understand which department owns their problem? Why should they search through 500 pages of a website? Why should they complete a twelve-field form if the organisation already possesses most of the information?
Why should an employee learn a complex enterprise application simply to ask it for an outcome?
For years we've tried to make complicated systems easier for humans to operate. AI gives us the opportunity to ask a different question:
Why are we asking the human to operate the system at all?
That is potentially a much bigger shift than adding a chatbot to a website and it requires human-centred design to operate at the level of the service and organisation, rather than simply the interface.
Perhaps human-centred design needs less design
This may be the uncomfortable bit.
AI probably will automate a meaningful amount of the work currently performed by designers, researchers, developers, strategists and content teams, and that's not necessarily a bad thing.
There is little inherent value in manually producing 50 wireframes if a machine can generate them, or spending days transcribing interviews.
And there is certainly little value in writing a 100-page research report nobody will read.
The value was never really in producing those things. The value was in the understanding and decisions they enabled.
AI gives us an opportunity to strip away some of the ceremony that has accumulated around digital product development.
Research can become more continuous.
Prototyping can become almost immediate.
Testing can happen earlier.
Production and experimentation can move closer together.
The distance between idea → experience → evidence → improvement can shrink dramatically.
That's exciting.
Because it allows human-centred design to become less about documenting what people might need and more about continuously discovering what actually works for them.
The scarce resource is changing
For years, execution was expensive.
You needed significant time, money and expertise to turn an idea into functioning technology.
That's changing.
The rapidly falling cost of AI inference is one indication of just how quickly sophisticated computational capability is becoming commoditised.[1]
Ideas, variations and content are becoming cheap to produce. Prototypes are heading the same way, and increasingly so is code.
But judgement isn't.
Understanding why somebody behaves in a particular way, recognising when you've been asked to solve the wrong problem, navigating the operational reality of a business or spotting when optimisation is creating unintended consequences - none of those things suddenly become easy because the software is easier to make.
In fact, those capabilities become more valuable as the cost of execution falls.
Human-centred design at AI speed
So, is human-centred design still relevant?
Absolutely.
Arguably more than ever.
But that doesn't mean every human-centred design method we've used for the last 20 years needs to survive.
Some of them shouldn't.
We should automate research tasks that don't require human judgement. We should generate and discard ideas much faster. We should prototype before we've finished arguing about the prototype. We should replace static deliverables with working experiments wherever possible.
We should allow AI to do more of the production work.
And we should spend more of our time understanding people, businesses and the consequences of the systems we're creating.
Because the future isn't human-centred design versus artificial intelligence, and it isn't really about bolting AI onto human-centred design either.
It's about what happens when human-centred design can operate at AI speed.
When the cost of making something approaches zero, making something is no longer the achievement.
Knowing what to make is.
AI compresses the execution layer. It doesn't remove the need for judgement.
It makes judgement the scarce resource.
References
- Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report 2025: Research and Development.
- Brynjolfsson, Li and Raymond, Generative AI at Work.
- Peng et al., The Impact of AI on Developer Productivity: Evidence from GitHub Copilot.
- Becker et al., Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.
- Amershi et al., Guidelines for Human-AI Interaction.
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0).