Designing AI Products: The Vaaryan UX Playbook
By Vaaryan Team • • 4 min read
Most AI products fail for design reasons, not model reasons. The model can be accurate and still go unused because people don't understand what it can do, can't tell when it is wrong, or don't feel in control. Designing AI products means designing for a collaborator that is powerful, fast and sometimes wrong. This playbook distils the principles we use at Vaaryan across our own products — Tuton, Tasklo and Findlen — and client work.
Principle 1: Start from the job, not the model
Before any interface, answer:
- What job is the user trying to get done?
- Which part of that job is slow, tedious or error-prone today?
- What would "done" look like, and how would the user verify it?
AI should remove friction from a real workflow. If the answer to "why AI?" is "because we can", the feature will be ignored.
Principle 2: Choose the right interaction pattern
Chat is one pattern among many. Choose deliberately:
| Pattern | Best when | Example |
|---|---|---|
| Inline suggestion | Users are already doing the task | Autocomplete, suggested replies |
| Draft for review | Output is long but needs human judgement | Proposal or report drafts |
| One-click transform | A clear, repeatable operation | Summarise, translate, simplify |
| Proposed action | AI can act but stakes are real | "Refund ₹1,200 to this order?" |
| Conversational | Open-ended exploration or guidance | Tutoring, research, troubleshooting |
| Background automation | High confidence, low stakes | Tagging, routing, deduplication |
In Tuton, for example, learners do converse with an AI tutor — but much of the intelligence shows up as adaptive practice and personalised next steps, not as chat.
Principle 3: Set expectations early
Users arrive with wildly different mental models of AI: some expect magic, some expect nothing. Good onboarding:
- Shows concrete examples of what to ask or do
- States limits plainly ("Answers come from your company's HR policies, updated weekly")
- Demonstrates value in the first minute with a real result
Principle 4: Make the invisible visible
AI is invisible by nature; trust needs evidence.
- Sources and citations. For any factual answer, show where it came from and let users open the source. This is where good RAG meets good UX.
- Reasoning summaries. For multi-step agents, show what was checked and done, in plain language.
- Status and progress. Long-running tasks should show stages, not a spinner.
Principle 5: Design for being wrong
Assume the AI will make mistakes and design the recovery:
- Easy correction — edit in place, regenerate, or "this is wrong" with a reason.
- Calibrated confidence — avoid fake precision like "87% confident"; use clear states such as "verified from source" vs "needs your review".
- Graceful refusal — "I couldn't find this in your documents" beats a confident invention.
- Undo for every AI-initiated change.
Principle 6: Keep humans in control
Autonomy should be earned. Match control to stakes:
- Low stakes, high confidence — act automatically, allow undo.
- Medium stakes — propose, let the user approve with one click.
- High stakes — gather evidence, recommend, and require explicit review.
Approval screens deserve real design attention: show the proposed action, the evidence and the consequences, and make approve, edit and reject equally easy.
Principle 7: Design for India's users
Building for Indian users means designing for:
- Multiple languages and code-mixing — users write "mera order kab aayega?" as naturally as English.
- Voice as a primary input for many users.
- WhatsApp as the default channel for service interactions.
- Variable connectivity and mid-range devices — stream responses, keep payloads light, handle drops gracefully.
- Diverse digital literacy — plain language, clear affordances, no jargon.
Principle 8: Latency is a design material
A 10-second wait feels broken; a 10-second wait with streamed, useful partial output feels fast. Stream text, show intermediate results, do work in the background, and reserve the slowest, most capable models for steps where users will wait. Engineering and design must plan this together (LLMOps).
Principle 9: Measure what matters
Model accuracy is necessary but not sufficient. Track:
- Task success — did the user achieve the goal?
- Acceptance and edit rate — how often is AI output used as-is, edited or discarded?
- Time to outcome compared with the old workflow
- Override and escalation rate
- Retention of users who adopt the AI feature
- Qualitative trust from interviews and session reviews
Our AI product design process
- Workflow research — observe the job being done today.
- Opportunity mapping — where AI removes friction, and at what stakes.
- Rapid prototyping with real models — designers prototype with live AI from week one, because static mock-ups can't capture probabilistic behaviour.
- Evaluation-informed design — design states for every failure mode the evaluation suite reveals.
- Pilot and iterate — ship to a small group, watch sessions, refine.
At Vaaryan, design and AI engineering sit in the same team, because in AI products you cannot separate the behaviour from the interface. If you are planning an AI product, start with our generative AI development guide for the engineering side, and bring designers in from the first day — not after the model works.
Frequently asked questions
How is designing AI products different from designing traditional software?+
Traditional software behaves the same way for the same input. AI systems are probabilistic: they can be wrong, vary between runs and surprise users. Design must set expectations, make reasoning and sources visible, make errors easy to catch and correct, and keep users in control of consequential actions.
Is a chat interface always the right choice for AI?+
No. Chat is flexible but puts the burden on users to know what to ask. Many AI features work better embedded in existing workflows: suggestions in a form, a draft in an editor, a one-click summary, or an action proposed for approval.
How do you build user trust in AI features?+
Show sources, be honest about uncertainty, explain what the system did, let users correct it easily, and be consistent. Trust grows when users can verify the AI and recover from mistakes quickly.
What metrics should AI product teams track?+
Beyond model accuracy, track task success, acceptance and edit rates of AI output, time saved, retention of AI feature users, escalation or override rates, and qualitative trust signals from research.
Work with Vaaryan
Turning AI ideas into products that ship
Vaaryan is an AI consulting and product company based in India. We help teams find the right use cases, design the experience, and engineer AI systems that hold up in production.