Intention debt is the new UX debt, there I said it. When AI repeatedly misreads what people mean, the product gets worse in ways your design system will never catch. For years, designers have had language for the slow accumulation of product mess. UX debt. It is the confusing setting that never got renamed. The onboarding step everyone knows is too long. The modal added to fix a launch problem that somehow became permanent. The flow that made sense three strategies ago. The inconsistency users trip over because every team solved the same problem slightly differently. UX debt is not always dramatic. That is why it survives. It hides inside reasonable compromises. We will clean it up later. We just need the release out. This edge case is rare. The user can figure it out. The tooltip explains it. Support can handle it. We know it is not ideal, but it works. Then, eventually, the product feels heavier than it should. Every improvement takes longer. Every new feature has to route around old decisions. Users feel the drag even when they cannot name it. AI products have their own version of this.

[I call it intention deficit]

Intent debt is what accumulates when a system repeatedly misreads, over-assumes, flattens, or redirects what users actually mean. It is not just a bad answer. It is a pattern of misunderstanding that becomes part of the product. And because AI products can sound confident while being wrong, intent debt can be harder to see than UX debt. The product heard the words, but missed the meaning Most product teams are trained to look for visible friction.

Where did users drop off?

Where did they click?

What did they not understand?

Which step took too long?

Those questions still matter. But AI introduces a different failure mode. The user can type a perfectly reasonable request. The system can respond fluently. The interaction can look successful from the outside. But something is off. The system answered the wrong version of the question. It solved for a generic user instead of this user. It gave a confident answer where a clarifying question would have been better. It optimized for completion when the user needed exploration. It collapsed emotion, context, constraints, or risk into a neat response. It moved too fast. It made the user feel handled rather than helped. That is intention deficit disorder. The words went in. The meaning did not survive. A simple example. Imagine a user asks an AI travel product:

"Can you help me plan a weekend away with my dad? He gets tired easily but does not like feeling old."

A bad system might produce a perfectly organized itinerary. Restaurants. Walking routes. Museums. Times. Links. A cheerful summary. On paper, it did the task. But the real intent was not "make an itinerary." The real intent included dignity, pacing, care, family dynamics, and emotional tact. The system needed to understand that the user's father may need rest without being treated as fragile. It needed options that preserve autonomy. It needed language the user could share without embarrassment. It needed to ask what kind of weekend would feel meaningful to both of them. If the product misses that, the failure may not show up as an error. It may show up as a user quietly not trusting it again. Intent debt often looks like successful task completion until you inspect the human residue it leaves behind.

[How intention debt accumulates]

Intent debt usually starts with reasonable product pressure. A team wants the AI to be fast, helpful, and proactive. They want fewer dead ends. They want the system to show value quickly. They want demos that feel magical. So the product begins to over-answer. It fills in missing context instead of asking. It treats ambiguity as inconvenience. It reduces messy human goals into clean task categories. It pushes toward output when the user is still forming intent. It optimizes for "done" because done is easy to measure. At first, this feels efficient. Then the product starts building a habit of being wrong in the same direction. That is the debt. Not one mistake. A repeated bias in how the system interprets people.

[Signs you have intention deficit disorder]

You may have intent debt if:

• Users keep correcting the system in similar ways.

• The AI answers quickly but users still ask follow-up questions that reveal the first answer missed the point.

• The product performs well in demos but feels brittle in real use.

• The system prefers producing an artifact over understanding the situation.

• Users describe outputs as "not wrong, but not quite right."

• The same prompt works for simple cases and collapses under personal, ambiguous, or high-context cases.

• The AI asks too few clarifying questions.

• The product has no clear policy for when to slow down, refuse, defer, or escalate.

• Teams evaluate outputs for accuracy but not fit.

• Support teams can name common misunderstandings that are not represented in design reviews.

The most important signal is repeated correction. When users keep saying some version of "No, I meant..." the product is teaching you where intent is leaking.

[Why this is different from hallucination]

Hallucination gets a lot of attention because it is easy to understand. The system made something up. It claimed a fact that was not true. It cited something that did not exist. That matters. But intent debt is broader. An AI system can be factually correct and still fail the user. It can summarize accurately but omit what matters. It can complete the requested task but choose the wrong level of detail. It can follow policy but sound cold. It can personalize output using the wrong signal. It can ask a question that is technically relevant but socially clumsy. It can be safe in a compliance sense and unsafe in a trust sense. This is why design needs to be involved. Intent debt lives in the gap between correctness and usefulness.

[Preventative care]

Here is a practical way to start addressing it. Choose a recent AI interaction pattern in your product. It might be search, onboarding, recommendations, customer support, planning, summarization, writing assistance, research synthesis, or an agentic workflow. Then review real or realistic interactions against five questions.

1. What did the user literally ask?
Capture the surface request. Do not interpret yet. Write down the words, action, or selected option.

2. What might the user have meant?

List plausible underlying intents. Were they trying to decide, understand, compare, create, repair, explore, avoid risk, gain confidence, save time, feel reassured, or get unstuck? Most AI products fail here because they treat intent as singular.

3. What did the system assume?

Name the assumptions. Did it assume the user's goal was obvious? That speed mattered most? That the user wanted a complete answer? That the user had expertise? That the user wanted advice rather than options? That the emotional context was irrelevant? Assumptions are not bad. Invisible assumptions are.

4. What did the system optimize for?

Look at the behavior. Did it optimize for completion, brevity, accuracy, confidence, persuasion, engagement, safety, conversion, convenience, or user agency? Every AI product optimizes for something, even when the team has not named it.

5. What should have happened instead?

Define the better behavior. Maybe the system should have asked one clarifying question. Maybe it should have offered two paths. Maybe it should have named uncertainty. Maybe it should have slowed down. Maybe it should have refused. Maybe it should have escalated. Maybe it should have produced a smaller first step instead of a complete plan.

This is where intent debt becomes design work.

[A lightweight scoring model]

For each reviewed interaction, score four dimensions from 1 to 5.

[Dimension Question Recognition]

Did the system identify the likely user intent? Fit Did the response match the user's context, constraints, and stakes? Agency Did the system preserve meaningful user control? Recovery Did the product make it easy to correct, refine, or redirect?

The score is not the point. The pattern is. If recognition is consistently low, you may need better onboarding, research, memory, or clarifying behavior. If fit is low, you may need richer context, better examples, or more nuanced product principles. If agency is low, your product may be over-automating. If recovery is low, users may be trapped inside the system's first guess.

[Paying down intent debt]

Intent debt does not get fixed by telling the model to "understand the user better." It gets fixed through product decisions.

Create intent categories that reflect real user needs, not internal feature categories. Add clarifying questions where the cost of guessing is high. Define when the system should give options instead of answers. Design undo, correction, and refinement as first-class interactions. Collect "No, I meant..." examples and turn them into eval cases. Teach the system product-specific judgment through examples and anti-examples. Review high-context interactions with researchers, designers, support teams, and domain experts. Measure whether users accept, edit, reject, or repeatedly redirect AI outputs. Most importantly, decide what should be slow. That may be the least comfortable design decision in AI products. The pressure is always toward speed. Faster answers. Faster generation. Faster workflows. Faster task completion.

But some moments should not be collapsed.

Ambiguity deserves a pause. High stakes deserve a check. Emotional context deserves care. Irreversible actions deserve confirmation. User intent deserves enough time to become visible. Speed is not always respect. Sometimes respect is the system knowing not to rush.

[Intent is a design material]

Designers are used to working with layout, hierarchy, language, motion, flows, states, constraints, and systems.

AI adds another material: interpretation. The product is no longer just presenting choices. It is inferring what people mean, deciding what matters, and taking action on those interpretations. That makes intent a design material. If we do not shape it deliberately, the system will shape it anyway. It will inherit assumptions from training data, prompt fragments, team defaults, business incentives, and whatever got rewarded in the last round of testing. Intent debt is what happens when those assumptions compound without scrutiny.

[The new design review]

A useful AI design review should not stop at the screen. It should ask:

What did the system think the user meant?

Why did it think that?

What else could the user have meant?

When should it ask instead of answer?

What does it do when its first interpretation is wrong?

What kind of misunderstanding would damage trust?

Where are we rewarding completion over comprehension?

Those questions belong in product strategy, research, design critique, content design, QA, and launch readiness. They are not edge cases. They are the experience. The debt you cannot see in Figma. UX debt often leaves visual evidence. Screens pile up. Components fork. Flows sprawl. Copy gets inconsistent. The mess is visible if someone takes the time to look.

Intent debt is quieter. It lives in the relationship between what users meant and what the system did with that meaning. That is why it is so easy to ignore. A team can have a polished interface, a strong design system, a fast model, and a beautiful demo, while the product steadily trains users not to trust it with anything nuanced. The future of AI product quality will not belong only to teams that generate the best outputs. It will belong to teams that get better at understanding what those outputs were supposed to serve.

[That starts with treating intent as something worth designing]