Onboarding for AI Products: Getting Users to Their First Good Answer
Onboard AI product users by stating what it can and can't do, offering example prompts and sample data, and reaching a useful answer in the first session.
8 min read

Umut Durgut
Product Designer

Contents
An AI product should onboard new users by getting them to one useful answer in their first session. Say plainly what the product can and can't do, replace the empty prompt box with example prompts and sample data, and suggest a first task that is low-risk and quick to check. Explain the benefit, not the model.
Picture a small consultancy that signs up for an AI tool promising to "summarise any client meeting." The first screen is an empty text box with a blinking cursor. One partner types "help me with my meetings" and gets a generic list of tips. Another pastes a two-hour transcript and gets a summary that misses the decision that mattered. Neither comes back. The tool might have been good. Nobody showed them what a good request looked like, or what the tool needed from them to give a good answer.
Why is onboarding harder for AI products than for regular software?
AI products are harder to onboard because users have to describe what they want instead of clicking a button, and the output varies from one try to the next. A new user can't see the product's limits from its menus, so they learn them by failing.
Jakob Nielsen called this out early in AI: First New UI Paradigm in 60 Years (June 2023): "The current chat-based interaction style also suffers from requiring users to write out their problems as prose text." Traditional software shows its options. A prompt box hides them.
There is a second problem: expectations. Google's People + AI Research team writes in its Mental Models chapter that "Many products set users up for disappointment by promising 'AI magic' will help them accomplish their tasks." A user who expects magic and gets an ordinary first answer decides the product doesn't work.
So onboarding an AI product has two jobs that regular onboarding doesn't: teach people how to ask, and set honest expectations about what comes back.
What does a "first good answer" mean?
A first good answer is the first output a new user would actually use: correct, relevant to their own work, and better or faster than what they would have done alone. It is the AI equivalent of an activation event, and onboarding should be designed backwards from it.
To define it for your product, answer three questions:
What task does a new user most likely arrive with? Not the most impressive task, the most common one.
What does the model need to do that task well? Context, a file, a connected account, a clear instruction.
How will the user know the answer is right? If they can't check it quickly, they won't trust it, even when it is correct.
The answers usually show that the first session should be narrower than the product. A tool that can analyse any document should start with one kind of document and one kind of question.
How do example prompts and sample data help new users?
Example prompts show users what a good request looks like, and sample data lets them see a real result before they've uploaded anything of their own. Together they replace the empty box with a first step that is likely to work.

Microsoft's Guideline 1 for Human-AI Interaction, Make clear what the system can do, lists patterns for exactly this, including "Demonstrate possible system inputs" and "Show a set of system outputs". The Nielsen Norman Group's guidance on designing empty states points the same way: "Use Empty States to Provide Learning Cues" and "Use Empty States to Provide Direct Pathways for Key Tasks".
What makes a good example prompt
It is specific to the user's job. "Summarise this sales call and list the next steps with owners" teaches more than "Summarise a meeting."
It shows the shape of a good request. Include the context, the task and the format you want back, so users copy the pattern, not just the words.
It works every time. Test every example against the current model. An example that fails on first click does more damage than no example.
It changes with the user. If you know the user's role or industry from sign-up, show examples for that role.
When to use sample data
Sample data is a ready-made file, workspace or record set that ships with the product so new users can try it without risk. Use it when the product's value depends on the user's own data, such as documents, a CRM or a spreadsheet, and connecting that data takes effort or approval. Label it clearly as sample data, and make the step from "try it on ours" to "try it on yours" a single, obvious action.
What should the first session include?
The first session should include a short statement of what the product does and doesn't do, one suggested task the user can finish in minutes, a result they can check, and a clear next step. Everything else can wait until the user needs it.
This checklist is a practical way to review a first session:
A one-line statement of what the product does well, written as a benefit, not a technology.
A plain note on what it can't do yet, shown before the first request, not after the first failure.
Two to four example prompts for the most common starting task, specific to the user's role where possible.
Sample data, if the product needs the user's data to show its value.
One suggested task that is low-risk and reversible.
An answer that shows its sources or working, so the user can check it.
A way to give feedback on the answer, such as a rating or a "this is wrong" button.
One clear next step, such as connecting their own data or saving the result.
Google's PAIR guidance supports most of this. It advises teams to "Be up-front about what your product can and can't do the first time the user interacts with it", to "Suggest a low-risk or reversible action they can try right away", and, on talking about the AI, to "focus on how it specifically makes part of the experience better or delivers new value."
How do you explain what an AI product can't do without scaring users away?
Explain limits as specific, practical facts at the moment they matter, not as a list of warnings on the first screen. Users trust a product more when its limits are clear and the product behaves within them.
The PAIR Mental Models chapter offers a simple message pattern for this: "Right now, it's not able to" followed by the main limitation. Pair each limit with what to do instead. "It can't read scanned images yet; upload a text PDF or paste the text" is a limit and a path forward in one sentence.
Three habits help:
Show limits in context. PAIR notes that "People learn better when short, explicit information appears right when they need it." Tell users a file type isn't supported when they try to upload it, not in a welcome tour they will skip.
Make uncertainty visible. If the answer is incomplete or the model is unsure, say so in the answer itself, and show where the information came from.
Keep the marketing and the product in line. If the landing page promises more than the first session delivers, onboarding starts with a broken promise. PAIR advises teams to work with marketing on "appropriate and consistent messaging."
How do you know if your AI onboarding is working?
Your onboarding is working when most new users reach the first good answer you defined, in their first session, and come back to do it again. Measure the answer, not the tour.

What to track
Reached first good answer: the share of new users who complete the defined first task and keep, copy or save the result.
Time to first good answer: how long it takes from sign-up, and where people stall.
Example and sample use: whether users start from an example prompt or sample data, and whether those users do better than people starting from a blank box.
Negative feedback on first answers: which first requests produce answers users mark as wrong.
Second session: whether users return and try the product on their own data.
What to watch in user tests
Run a handful of sessions with people who match your target user, give them no instructions beyond the product itself, and watch the first ten minutes. Note what they type first, where they hesitate, and the moment they decide whether the product is useful. That moment is what onboarding exists to improve.
How we approach it at Vizio AI
We design onboarding as part of the product, starting from the first good answer rather than from screens. With the founding team, we pick the one task a new user most likely arrives with, define what a useful answer to it looks like, and work out what the model needs from the user to get there.
Then we build the first session around that task: example prompts written for the target role, sample data where the user's own data is slow to connect, and limits stated plainly before the first request. We test it with real target users early, often while the rest of the product is still a prototype, and change the examples and copy based on what people actually type. The same loop thinking we use for agents, covered in designing for AI agents with approvals and undo, applies here: show the work and make the next step obvious.
This is part of how our Venture Studio builds AI products. If you haven't confirmed the core task yet, start with validating a startup idea before you build. If the product is close to launch on mobile, shipping a mobile MVP that passes store review covers the next step.
Sources
Google PAIR, People + AI Guidebook, Mental Models: https://pair.withgoogle.com/chapter/mental-models/
Microsoft HAX Toolkit, Guideline 1: Make clear what the system can do: https://www.microsoft.com/en-us/haxtoolkit/guideline/make-clear-what-the-system-can-do/
Nielsen Norman Group, Designing Empty States in Complex Applications: 3 Guidelines (Kate Kaplan, 2021): https://www.nngroup.com/articles/empty-state-interface-design/
Nielsen Norman Group, AI: First New UI Paradigm in 60 Years (Jakob Nielsen, 2023): https://www.nngroup.com/articles/ai-paradigm/
FAQs.
Read More

Attribution That Survives Your Tool Stack: Keeping UTMs From Ad to Booking to CRM

Clean Your CRM Data Before You Add AI: Dedupe, Grain and Ownership

Semantic Layer First: Getting Your BI Ready for AI Agents

How F&B CMOs Use Generative AI In Business For Location Decisions

How Real Estate Firms Use Business Analytics For City Parks To Drive Value

Data Storytelling and Visualization

When Does a Business Actually Need Digital Transformation?

Predictive Analysis for Data Visualization

Data Analyst vs. Data Quality Analyst

Analytics for Product Recommendations

