Experience debt is the bill AI passes to your customers

Technical debt slows your developers. Experience debt drives your customers away.

Most companies understand the first half of that sentence far better than the second. Technical debt has a language executives accept, dashboards that track it, and ratios that price it. It has earned its seat on the engineering agenda. Experience debt hasn’t. It’s the softer problem, the one you can defer. Design teams raise concerns during planning, and roadmaps quietly bump them to next quarter. It’s rarely shown as a number on a dashboard, so it rarely shows up as a priority.

What makes the distinction matter is who pays. Technical debt gets paid inside the building, by the engineers who work around the shortcut. Experience debt gets paid outside of it by the people trying to use what you shipped. And customers, unlike engineers, don’t file tickets against your backlog. They just leave quietly, and they rarely say why.

For most of the last decade, experience debt remained buried in forgotten backlogs. Then generative AI arrived, and the second half of the opening sentence gained a whole new meaning.

Generative tools changed what experience debt is, where it comes from, and who finds it first. The way we’ve caught it for thirty years no longer works. That’s what our Experience Design team explores in this blog.

What is experience debt?

To see what changed, it helps to remember where the word “debt” came from in the first place.

In 1992, Ward Cunningham reached for a money metaphor to explain something non-engineers found hard to fund: why a team would spend time reworking code that already ran. You can ship fast by borrowing against the future, he argued, but you pay interest on that loan in every release that follows ¹.

The metaphor was good enough that it moved beyond just engineering. By 2013, designers had borrowed it for “UX debt,” the widening gap between the experience a product delivers and the one it could have delivered if the research and care hadn’t been skipped ². By 2020, Deloitte had stretched it again into “experience debt,” calling it this generation’s version of technical debt ³. More recently, Thoughtworks introduced cognitive debt to describe the growing distance between what AI-assisted systems do and what the people maintaining them actually understand ⁴. 

Every version of the metaphor serves the same purpose: making a hidden, put-off cost clear to the people who hold the budget. And every version rests on the same assumption. The debt was chosen. Someone took the shortcut on purpose. Someone knew the corner was being cut, even if they never wrote it down. The interest was the price of a decision a person made.

That assumption is the thing AI breaks.

AI changed how experience debt accumulates

The old experience debt was a string of choices. Someone made a trade-off under a deadline, felt the right amount of guilt, and moved on. The work ran slowly enough that the choices stayed countable. You knew where the decisions were buried because you buried them, which meant you could, in theory, dig them up again.

The new experience debt works another way. It piles up from decisions no one consciously made, in screens no one fully wrote, shipped faster than anyone can check. There is no borrower. The line that caught the old world, “developers borrow and users pay”, assumed a developer knew a loan was being taken. Now the loan takes itself.

AI has crushed the cost of making an experience: a screen, a flow, a paragraph, a button. Checking that the experience actually works has gotten cheaper too, but checking never became automatic the way generating did. It still takes someone choosing to go and find out. Building and checking were always two jobs, and while people did the building, the checking more or less kept pace. That balance is gone. A model can generate a thousand screens before lunch. Learning whether any of them helps a real person still takes a deliberate act. 

No model sighs in frustration on a customer’s behalf. None of them hopelessly abandon a checkout for you to discover the problem in the numbers later.

Why AI-generated experiences look finished (at a glance)

Here is the trap inside the trap. AI is very good at making work that looks done.

Picture the review. A generated screen goes up on the shared display. The spacing is even, the copy reads clean, and the buttons sit where buttons are supposed to sit. Someone nods. Someone says ship it. Nobody in that room has watched a single real person try to use the thing, because watching takes a week and the screen took four seconds. 

The screen works. Whether it’s useful, whether it actually helps a real person do the thing they came to do, nobody has checked. 

Working and useful are not the same thing.

That gap is the whole problem in one meeting. A build like this clears the demo, then fails where demos never look: on the edge case nobody scripted, on a screen reader, on the hundredth step of a real workflow, in the messy domain reality the model never saw. 

And even when it runs clean, it can still leave the user no closer to what they came to do. The break surfaces later, and elsewhere: in a support queue, a drop-off curve, a churned account no one can explain. 

Cause and cost come unstuck from each other, and AI keeps prying them apart.

UX validation won’t hold when generated experiences don’t hold still

We are not defenseless against this. For thirty years, we have run one loop to catch it: audit, fix, certify, ship. 

Run the usability study. Run the scan. Catch the regressions against a baseline. Sign it off. This loop works because of one assumption so basic we never say it out loud: the thing you are studying will hold still long enough to study it.

Generated experiences don’t hold still. 

It is like sending a building inspector to a house that redraws its own floor plan between visits. They can sign off on the staircase they see. They cannot sign off on the staircase the next visitor will climb. A real house, at least, stays a house. A generated interface makes no such promise. It breaks every part of the validation loop at once:

• You can’t usability-test a screen that’s different for every visitor. A finding describes one arrangement of one interface. Build that arrangement fresh each time, and the finding describes something that’s already gone.
• You can’t run a visual regression with no baseline. Regression testing asks whether what shipped matches what you meant to ship. A system built to draw something new on every render has no “meant to” to check against.
• You can’t certify a layout that never repeats. A certificate is a claim about a specific thing. Certify a generated interface today, and you’ve blessed one instance out of millions you will never see.

So the loop doesn’t slow down. It stops working. We built it to catch a target that stays put, and AI handed us one that moves. You can’t audit a moving target.

Read more tech insights on this topic: AI agents are assembling adaptive UI. Here’s how validation needs to evolve. 

Experience debt cascades into trust debt

Experience debt doesn’t stay at the surface. It comes from three places, and they stack.

1. It starts with research debt: the understanding that never happened. No real validation; calls made on a hunch; personas built once and never revisited, because the output already looked validated. Everything else gets built on top of that. 
2. Design debt comes next: broken patterns, drifting components, accessibility gaps piling up faster than any design system can govern them. 
3. Then comes seam debt: the experience stitched together from parts that each work on their own while the whole quietly stops making sense across screens, channels, and the teams that own them.

Those three are sources. They all roll into one cost: trust debt. The customer’s slow conclusion that your product can’t quite be counted on. It is by far the most expensive because customers pay it back the only way they can: they leave.

So this isn’t four problems, and it isn’t really four layers. It’s three things you can skimp on and the one bill that comes due when you do skimp on them. And all three begin in the same place: the gap between how effortlessly you can now generate and how easily you can skip the checking. 

Something used to guard each step. Research caught the bad assumptions. Design QA caught the broken patterns. End-to-end testing caught the broken seams. AI overran all three at once.

Each layer is worth a piece of its own, and we’ll give them that space in the blogs that follow. For now, hold on to the shape.

Customers, regulators, and plaintiffs find experience debt first 

Notice where the cascade ends: with the customer. This is where discovery has moved, too. Companies used to find their own experience debt first. A usability study surfaced it. An audit exposed it. A support trend pointed to it. Teams could decide what to do about it, and when. Now someone else usually finds it first.

Sometimes that someone is a customer, slipping away to a competitor without a word. Increasingly, it is a regulator or a plaintiff. Digital accessibility has crossed from good practice into enforceable law in major markets. The European Accessibility Act is in force ⁵, and the EU AI Act’s transparency rules are landing on a fixed clock ⁶. 

Read more tech insights on this topic: Are your UI application development processes compliant with the EU AI Act? 

And the same cheap production that fills your product with unchecked decisions has made finding those decisions cheap too. A growing share of accessibility complaints now come from people aiming automated tools at websites, surfacing violations in bulk, and acting on them ⁷.

Sit with that symmetry for a second. The same kind of technology that makes the debt is now the technology hunting it. The debt finds itself. And the finder isn’t always on your side.

This is a business problem, not a quality problem

It’s tempting to read all this as a familiar argument: good design matters, so care more about it. The argument is true. It is also already won. Few people dispute the business value of mature design; McKinsey spent years measuring it and found that the most design-mature companies outrun their peers, and no one is arguing the other side ⁸.

That is exactly why “care more” misses the point. The value of the destination was never in question. The road changed. AI has changed the economics of creating experiences. Teams can generate more screens, flows, and interactions than ever before. What they have not done is rebuild the systems that check whether those experiences work.

And checking isn’t even the slow, expensive thing it used to be. 

Unmoderated testing turns around in a day or two, and AI helps write the questions and make sense of the answers. The cost of learning whether something works is often a low single-digit percentage of the cost of building it.

Teams skip it anyway. 

They will spend millions on the tokens to generate the thing and hesitate to spend a few thousand validating them. Almost nobody has rebuilt around that reality.

And the bill is already arriving. 

In 2025, MIT researchers studying enterprise AI initiatives found that roughly 95% delivered no measurable return ⁹. The problem was not the models. Most of them worked.

The failure came from workflow design, operational context, and fit with real users: the exact work that validation used to cover before teams started skipping it.

Put human judgment back in the UX loop

Rebuilding sounds like a call to slow down, or to trust the models less. It is neither. 

AI plainly belongs in building digital experiences, and it isn’t going anywhere. The point is narrower and harder. The habit of validating work before it reached customers carried more weight than we realized. The systems we built around slower, human production relied on that habit. AI knocked it out.

The reflex is to automate the catching: more scanners, more tests, more gates in the pipeline. 

Some of that earns its keep. You can automate an accessibility scan. You can automate a regression check. What you cannot automate is the part that was always the point, the judgment call on whether a screen that works is actually useful, whether a flow holds together over a hundred real steps, whether the thing helps a real person or only looks like it does. 

A model cannot make that call about its own output because making that call is the exact work the model skipped.

So the fix isn’t to care more, and it isn’t to slow down. Nobody is going to do either at the scale this needs. 

The fix is to put judgment back into the loop and keep it there: a standing layer between the generative engine and the user, part automated check and part human call, running at the speed at which the work is now created instead of showing up once a quarter to take inventory. What that looks like across research, design, the seams between them, and the trust architecture that holds them is what we are covering in a series of posts here.

Go back to Cunningham’s metaphor one last time. His borrower always knew the loan was his, due on a date he picked. This new debt took that away. The interest still comes due, but on a customer’s schedule, or a court’s, and never on yours.

So the job isn’t to schedule the payments better. 

It’s to stop the loan from being taken in the dark. Experience debt used to be a choice. That is the part worth getting back. Put judgment back in the loop, and you turn the accident, slowly, back into a decision. One you make on purpose, in the open.

Get in touch with our experience design team if experience debt is a current concern.

References

1. The WyCash Portfolio Management System | Ward Cunningham
2. UX Debt | Adaptive Path
3. Experience Debt: How Digital Experiences Are Failing Customers | Deloitte Digital
4. Cognitive Debt | Thoughtworks
5. European Accessibility Act | European Commission
6. Artificial Intelligence Act (AI Act) | European Parliament
7. 2024 Report: ADA Digital Accessibility Lawsuits | UsableNet
8. The Business Value of Design | McKinsey & Company
9. MIT Finds 95% of GenAI Pilots Fail Because Companies Avoid Friction | Forbes

Tags

Artificial intelligence

Cross-industry

References

1. The WyCash Portfolio Management System | Ward Cunningham, https://c2.com/doc/oopsla92.html
2. UX Debt | Adaptive Path, https://adaptivepath.org/ideas/ux-debt/
3. Experience Debt: How Digital Experiences Are Failing Customers | Deloitte Digital, https://www.deloittedigital.com/us/en/insights/perspectives/experience-debt.html
4. Cognitive Debt | Thoughtworks, https://www.thoughtworks.com/insights/blog/generative-ai/cognitive-debt
5. European Accessibility Act | European Commission, https://ec.europa.eu/social/main.jsp?catId=1202
6. Artificial Intelligence Act (AI Act) | European Parliament, https://artificialintelligenceact.eu/
7. 2024 Report: ADA Digital Accessibility Lawsuits | UsableNet, https://blog.usablenet.com/2024-report-ada-digital-accessibility-lawsuits
8. The Business Value of Design | McKinsey & Company, https://www.mckinsey.com/capabilities/mckinsey-design/our-insights/the-business-value-of-design
9. MIT Finds 95% of GenAI Pilots Fail Because Companies Avoid Friction | Forbes, https://www.forbes.com/sites/jasonsnyder/2025/08/26/mit-finds-95-of-genai-pilots-fail-because-companies-avoid-friction/

Share

Follow

Subscribe

Follow