Stop Calling Everything an 'AI Feature'
2026-02-11 · 7 min read · Janaina Maia
I was reviewing a product roadmap recently where everything had the word "AI" in it. AI-powered search. AI recommendations. AI chatbot. AI analytics. AI-driven notifications. The roadmap looked like someone had played buzzword bingo and won every round.
Here's my problem: these are fundamentally different things that need fundamentally different design approaches, risk assessments, and trust-building strategies. Lumping them all under "AI features" is like calling a bicycle and a Boeing 747 both "vehicles." Technically correct. Completely useless for making design decisions.
A More Useful Taxonomy
When I work with product teams, I push them to categorize their AI features into one of these buckets. Each has different UX implications:
1. Smart Defaults
AI that pre-fills, sorts, or organizes things based on patterns. Autocomplete. Smart categorization. Intelligent search ranking.
UX implication: These should be invisible. Users shouldn't have to think about them. The moment a "smart default" requires a tutorial, you've over-engineered it. If it's wrong, it should be trivially easy to change — like deleting autocomplete text.
2. Recommendations
AI that suggests actions or content. "You might also need..." "Similar projects used..." "Based on your data, consider..."
UX implication: These need to be clearly optional and clearly sourced. The user should understand why the recommendation was made and be able to ignore it without friction. Never auto-apply a recommendation.
3. Analysis
AI that processes data and produces insights. Pattern detection. Anomaly flagging. Trend identification.
UX implication: These need confidence indicators and methodology transparency. The user needs to know what data was analyzed, how, and how reliable the conclusion is. This is where my Confidence Spectrum framework lives.
4. Generation
AI that creates content. Report drafts. Design suggestions. Code. Natural language responses.
UX implication: These need clear "AI-generated" labeling, easy editing, and hallucination safeguards. Users should review before anything goes external. Always provide a diff or track changes view.
5. Automation
AI that takes actions. Auto-classifying data. Routing workflows. Triggering notifications based on conditions.
UX implication: These need audit trails, undo capability, and clear guardrails. Users need to see what was done, why, and be able to reverse it. This is where the Copilot-to-Autopilot maturity model matters most.
6. Agents
AI that orchestrates multi-step workflows autonomously. Research tasks. Complex data processing pipelines. End-to-end processes.
UX implication: These need everything — plans, progress visibility, interruption capability, handoff protocols. This is the hardest UX to get right. (See my agentic UX patterns post.)
Why This Matters
When a PM says "let's add an AI feature," my first question is now: "Which kind?"
Because a Smart Default needs maybe 2 days of design work and minimal risk assessment. An Agent needs weeks of design, extensive error handling, trust-building patterns, and probably a phased rollout strategy.
Treating them the same way leads to either over-designing simple features (adding confidence indicators to autocomplete — why?) or under-designing complex ones (shipping an autonomous agent with just an "undo" button).
The Conversation I Want to Stop Having
Stakeholder: "Does our product have AI?"
PM: "Yes! We have 12 AI features!"
This tells me nothing. I want to hear:
"We have smart defaults in search and data entry, recommendation engines for methodology selection, AI-powered analysis for soil classification with confidence display, and we're piloting an agent for automated report generation at Level 2 autonomy with human review."
Now I know what we're actually building, what the risk profile is, and what design patterns are needed.
Stop saying "AI features." Start saying what the AI actually does. Your design team — and your users — will thank you.