THE KEY ANSWER
Good AI UX helps the user understand the system's capabilities, verify the result, and correct errors. Choose the interface for the task: sometimes a conversation, and sometimes a form, a comparison, or editing a specific field. Always show the status and the next step.
Does the user really want to write a prompt?
A person preparing a quote may know the product, quantity, and deadline, but not know how to describe it to the model. A form with appropriate fields can shorten the work and reduce the number of missing details. Conversation is useful where the problem requires clarification, but it does not have to be the only way to handle the entire application.
Start with the task, and then choose the interaction. For comparing document versions, a diff view will be useful. For data analysis, a set of results with the ability to check the source. If you turn everything into a long stream of messages, the user may spend more time finding information than before.
Context and references: Google PAIR: People + AI Guidebook
Set expectations before the first attempt
Show what the tool is suitable for and what data is needed. A short, concrete example of a task helps more than a general promise to "ask about anything". Explain whether the result is a draft, a suggestion, or an executed operation. The user should understand this difference without reading technical documentation.
Demonstration example: a "Prepare quote" button may suggest a ready-to-send document. If the system creates a working proposal requiring review, the name and the next view should indicate this. Clear communication does not weaken the product. It helps use it properly and evaluate the result according to its real capabilities.
Correction should be cheaper than starting over
Allow correcting a specific fragment while retaining approved data. If the user changes a deadline, they should not have to regenerate the entire document and check all fields. Show which parts may change after correction. For critical operations, keep a preview before execution.
Also design for the absence of a result: insufficient sources, unavailable integration, unsupported document. The message should say what the user can do now. "Try again" makes sense only if the next attempt can solve the problem. For missing data, a request for specific information is better.
Measure the effort needed for a useful result
Observe time to first value, number of corrections, drop-offs, and result utilization. Long time spent in the application may indicate engagement or difficulty. Check this through conversation and task observation. Do not optimize the interface solely for the number of sent messages.
In tests, include beginner users, experts, and people using different input methods. Provide the ability to interrupt a task and save the work. An AI product should increase the sense of control over the process. The surprise effect may encourage the first attempt, but subsequent use depends on predictable value.
WHERE TO START
Bring this into your project.
- Match the interface to the task, not to the technology.
- Explain the difference between a draft and an action.
- Enable correction without losing approved work.
- Measure effort, result, and reason for drop-off.
Choose one thing your process is missing today. It's a useful topic for your first conversation with the team.
QUESTIONS AND ANSWERS
Frequently asked questions.
Does every AI product need chat?
No. Many tasks are better handled by a form, a comparison panel, or a hint in an existing process. It is worth choosing chat when conversation truly helps clarify the goal.
How to build trust without pretending to be infallible?
Show sources, status, scope of action, and how to correct. The user needs a predictable process, not a declaration of 100% effectiveness. Limitations should be communicated where they affect the decision.
Sources and context
- Google PAIR: People + AI Guidebook ↗
This guide concerns the design of AI products with consideration for people. The examples of quotes, corrections, and effort measurement are original proposals by ALGOV.
Prepared by the ALGOV team. Current as of September 8, 2026. Examples describe possible scenarios, not results from client projects. How we create our guides.
