Marcus Andrey Study
Ler em português Library

AI for UX · Step 5 · 6 min read

When not to use AI

Before designing an AI feature, decide whether it should use AI at all. Where AI fits and where it probably doesn't (Google PAIR), real cases where it backfired, the simpler alternatives that often win, the cost nobody sees, and a five-question check you can run with your team.

"Add AI" is not a user need. Many AI features ship because a competitor has one, or because a roadmap said "AI" in 2026. People notice: NN/g's State of UX 2026 calls it "the year of AI fatigue" (NN/g, Jan 2026). The most useful design decision about AI is often the first one: should this use AI at all? Saying no, with reasons, is a design skill.

What you'll be able to do after reading:

  • Tell when AI is likely to help a task and when it probably won't
  • Recognise the patterns behind real AI features that backfired
  • Propose simpler alternatives: rules, search, good defaults, better content
  • Run a five-question "should this be AI?" check before any design work

1. Start with the problem, not with AI

NN/g's advice is short: "Don't start with AI, start with the problem" (NN/g video, Jan 2026). Write the user's problem first, in one sentence, with no technology in it. Then ask which solution solves it best. Sometimes that's AI. Often it isn't.

Two hands cracking a walnut with a small glass nutcracker, next to a pile of walnuts on brown paper.
A nutcracker does one job, every time, with no surprises. Many product problems need a tool like this, not a model. Photo: Agaath, Wikimedia Commons, CC BY-SA 4.0.

2. Where AI fits, and where it probably doesn't

Google's People + AI Guidebook lists situations where AI is probably better and where it's probably not (PAIR, User Needs + Defining Success):

Probably betterPersonalising

Recommending what fits each person.

Suggestions, rankings
Probably betterRecognising and predicting

Patterns in images, text or behaviour.

Classifying, spotting anomalies
Probably betterNatural language

Understanding and drafting open-ended text.

Summaries, first drafts
Probably notPredictability matters

People need the same result every time.

Navigation, settings, policies
Probably notErrors are costly

A wrong answer harms or costs a lot.

Prices, legal terms, health
Probably notPeople value the task

They enjoy it or feel responsible for it.

Creative choices, care
Where AI tends to fit, and where it probably doesn't. Adapted from Google PAIR's People + AI Guidebook. Diagram: Marcus Andrey, CC BY 4.0.

The guidebook's "probably not" list, in its own words: "Maintaining predictability", "Providing static or limited information", "Minimizing costly errors", "Complete transparency", "Optimizing for high speed and low cost", and "Automating high-value tasks" (the ones people enjoy or feel responsible for).


3. When it backfired

Case What happened The pattern
Air Canada, 2024 A support chatbot told a customer he could claim a bereavement discount after the trip. He couldn't. A Canadian tribunal ordered the airline to pay: "It makes no difference whether the information comes from a static page or a chatbot" (CBC) Policy answers need to be exact: use the policy page
Apple, Jan 2025 AI summaries of news notifications produced false headlines; Apple paused them for news apps (TechCrunch) Short, trusted surfaces amplify errors
Cursor, Apr 2025 The company's AI support agent invented a login policy that didn't exist, and users cancelled subscriptions (Ars Technica) An AI speaking for the company needs limits and a person behind it
Klarna, 2024 to 2025 In 2024 it said its AI assistant did the work of 700 support agents; in 2025 its CEO said focusing on cost had led to "lower quality", and the company invested in human support again (Klarna; CX Dive) Cost savings aren't a user outcome
McDonald's, 2024 Ended its AI voice-ordering test at drive-thrus after about three years (CNBC) Noisy, real-world input is harder than a demo

None of these means "AI is bad". Each means the task needed exactness, accountability or reliability that the design didn't provide.


4. Simpler alternatives that often win

  • Better search and filters. NN/g found that many site chatbots leave users unsure what they're for, and warns against rebuilding search, filters or navigation as a slower chat (NN/g, Mar 2026). For critical facts, search beats generated answers (NN/g, Feb 2026).
  • Rules and good defaults. If you can write the logic down, write it down. It's predictable, testable and cheap.
  • Better content. A clear help page, a well-written error message or a sensible form often removes the need for an assistant.
  • A person. For high-stakes, emotional or rare situations, a fast route to a human is the feature.

5. The cost nobody sees

Every AI answer costs money and energy, every time. The International Energy Agency estimates data centres used about 415 TWh of electricity in 2024 (about 1.5% of the world's), heading for about 945 TWh by 2030, with AI as the main driver (IEA, Energy and AI, Apr 2025). A feature that runs a model where a rule would do is paying that cost for nothing.

There are human costs too: people have a right to know when they're talking to an AI (the EU AI Act requires it, in an accessible way, in Article 50: text), and AI-only flows can lock out people who need another way. See Accessibility with and for AI.


6. The five-question check

  1. 1The problem?

    One sentence, no technology in it.

  2. 2A simpler way?

    Search, a rule, a default, better content.

  3. 3Wrong answers OK?

    Who notices, and what does it cost?

  4. 4Another path?

    People can always skip the AI.

  5. 5Worth the cost?

    Money, energy, maintenance.

Five questions before designing an AI feature. A "no" anywhere means: look again at the simpler way. Diagram: Marcus Andrey, CC BY 4.0.

7. Common mistakes

  • Starting from "where can we add AI?" instead of a user problem.
  • Replacing a working search or form with a chat.
  • Using AI where a wrong answer is costly and nobody checks.
  • Counting usage of the AI feature as success. See Measuring AI features.
  • Treating "no AI" as a failure of ambition rather than a design decision.

8. Vocabulary

Term Meaning
Deterministic Gives the same output for the same input, every time (rules, search, forms)
Probabilistic Can give different outputs for the same input (AI models)
AI fatigue People tiring of AI features that add little
Inference cost The money and energy spent every time a model answers
Fallback The non-AI path people can always take

9. Try it (no code needed)

  1. Pick one AI feature in your product, planned or shipped.
  2. Write the user problem it solves in one sentence, without the word "AI".
  3. Run the five questions in section 6.
  4. Sketch the best non-AI alternative. Compare the two honestly with a colleague.

Reflection questions: Which AI feature you've seen recently would have been better as search, a rule or a good page? What would make you confident enough to say "no AI here" in a roadmap meeting?


10. If you lead a team

  • Make "why AI?" a required line in every AI feature brief, answered from the user's problem.
  • Keep a list of AI features you decided not to build, and why. It's as valuable as the roadmap.
  • Measure AI features against their non-AI alternative, not against nothing.

Checklist

Learn more

Books, courses and videos to go deeper

Book links go to Amazon. As an Amazon Associate I earn from qualifying purchases, at no extra cost to you.

Agentic design · Designing for uncertainty · Measuring AI features · Ethics, bias and inclusion in AI design · Accessibility with and for AI · Hallucination

Saved only on this device.

↑ ↓ to move · Enter to open · Esc to close