Onboarding and setting expectations
Introduce AI features so users know what they can do, where they fall short and how to get good results, without long tutorials.
On this page
Why this matters
Users arrive at an AI feature with expectations shaped by headlines, other products and their imagination. Some expect magic and are disappointed by the first imperfect answer. Others expect nothing, never try it, or try it once on the wrong task and leave. Either way, the feature fails before its quality has a chance to matter.
Onboarding is how you set the expectation in between: what this feature is good at, what it isn’t, and how to get a good result on the first try. For AI features, the first few interactions often decide whether someone becomes a regular user, so they deserve as much design attention as the feature itself.
The expectation gap
Two failure modes are common:
- Over-expectation. A feature introduced as “your AI assistant” invites any request. The first time it can’t do something obvious to the user, trust drops, and it’s hard to recover.
- Under-expectation. A feature hidden behind a generic sparkle icon, with an empty text box, gives users no idea what to try. Many don’t try at all.
Both come from the same root: the user doesn’t know what the feature is for. The fix is to introduce it through the job it does, not through the technology. “Draft a reply from your help center” sets a clear, achievable expectation. “Ask AI anything” doesn’t.
Showing, not telling
Long explanations and tutorials are skipped. What works better is letting users see a good result quickly:
Show text descriptionHide text description
A sequence of 5 steps.
- Name the job : One line on what it does: "Drafts replies from your help center"
- Show an example : A real before-and-after, or suggested prompts for this context
- First result : On the user's own content, with one click
- Show the controls : Edit, regenerate, sources: point them out at the moment they matter
- Second use : Make the next time easy to find and start
Techniques that help:
- Suggested prompts in context. In an empty state, offer three to five suggestions that work well on the user’s actual content: “Summarize this thread”, “List open questions”, “Draft a reply”. They teach by example what the feature can do.
- Smart defaults. Pre-fill the most common request so the first result is one click away.
- Real examples. A short before-and-after on representative content shows quality better than any description.
- Contextual entry points. Put the feature where the job happens: a “Draft reply” button in the reply box beats a separate AI page.
- Progressive disclosure. Introduce advanced capabilities after the first success, not before.
Being honest about limits
Telling users what the feature can’t do sounds like bad marketing. In practice it protects trust, because the first failure then matches what they were told instead of contradicting it.
- Name limits concretely and briefly. “Works best on threads under 200 messages” or “Uses your help center only; it won’t know about unreleased features.” Not a wall of disclaimers.
- Put limits where they matter. Near the input, or in the result, not in a terms page nobody reads.
- Say that output should be checked when it should. “Review before sending” next to a drafted email is honest and useful.
- Show sources so users can see what the feature knows. See citations.
A feature that clearly says “I only answer from your documents” and then does exactly that earns more trust than one that promises everything.
Measuring onboarding
Onboarding works if people get a good first result and come back. Measure:
| Metric | Definition | What it tells you |
|---|---|---|
| Discovery | Users who see the entry point ÷ active users | Whether people find the feature |
| Activation | Users who complete a first successful use ÷ users who tried | Whether the first try works |
| Second use | Users who use it again within a week ÷ activated users | Whether the first result was worth repeating |
| Suggested prompt use | First uses that started from a suggestion ÷ first uses | Whether your examples guide people |
Second use is often the most telling. A user who comes back has learned what the feature is good for. Product analytics events covers how to record these.
Common mistakes
- “Ask me anything.” An open text box with no guidance invites requests the feature can’t handle. Introduce it through the job it does.
- Long tutorials. Users skip them. Get them to one good result on their own content instead.
- Hiding limits. Unstated limits turn the first failure into a broken promise. State them briefly, where they matter.
- A separate AI page. Features placed where the job happens get used; features on their own page get forgotten.
- Measuring only first use. Curiosity drives first use. Activation and second use show whether onboarding worked.
Try it
Exercise
About 40 min
Write the first-run experience for a feature
- Write the one sentence a first-time user should believe about your feature.
- Write the entry point label, three to five suggested prompts and one honest limit.
- Sketch the empty state and the first result, including where edit and regenerate controls appear.
- Define how you’ll measure activation and second use.
Show a hint
Test your suggested prompts on real content before shipping them. A suggested prompt that produces a weak result is worse than none, because it’s the first impression you chose.
Show a worked solution
Example for “summarize this thread” in a team chat product:
- Belief: “It catches me up on long threads in a few bullet points, with links to the key messages.”
- Entry point: a “Catch me up” button at the top of any thread with more than 20 unread messages.
- Suggested prompts: “Catch me up”, “What was decided?”, “What’s still open?”, “What do I need to do?”
- Limit: “Summarizes this thread only; it can’t see other channels.”
- Measurement: activation = summaries viewed without immediately opening the full thread; second use = another summary within 7 days.
Check your understanding
Key takeaways
- Introduce AI features through the job they do, not the technology.
- Get users to one good result on their own content, using suggested prompts, defaults and contextual entry points.
- State limits briefly, concretely and where they matter.
- Measure activation and second use, not just first clicks.
Progress is saved in this browser and lights up the roadmap.
Related articles
- Communicating confidence and uncertainty Communicate uncertainty in AI features honestly: when to hedge, when to abstain, and why raw confidence scores usually confuse users more than they help. Level: Intermediate
- Jobs-to-be-done for AI products Use jobs-to-be-done to frame AI features around the progress users want to make, write job statements, and spot where AI changes the job. Level: Beginner
- Product analytics events for AI features Instrument AI features with product analytics, with an event schema for requests, outcomes, edits and costs, and the questions it lets you answer. Level: Advanced