Product Design / AI
AI Product Design
AI product design is the design of software whose core output comes from a model: the input people give, the wait while the model works, the result they judge and the correction they make when it is wrong. Gestures designs AI products against real model outputs, sets how much the product may do on its own and designs every failure before the success screen.
AI product design with Gestures costs $2,250 per week, $4,000 for a two-week sprint or $7,500 per month for ongoing work across model and feature releases. Every plan covers strategy, design and review on a scope agreed before work starts.
Method
Design with real model outputs, not ideal sample text
An AI interface designed around ideal sample text breaks on the first real answer, so Gestures designs with a set of real outputs from your model. Your team exports the set, and every screen is laid out and reviewed against it.
- A typical answer and the longest answer the model produces.
- An almost-right answer that a person has to correct.
- A wrong answer delivered with confidence.
- A refusal and an empty result.
- A slow response and a timeout.
Each design review runs through the same set, so a layout that only works for the ideal answer is caught before it reaches engineering. When the model is not ready yet, outputs from a comparable model stand in and are replaced as soon as your own model produces them.
Autonomy
Decide how much the product may do on its own
The autonomy level decides most of an AI product's interface, because a product that suggests needs a different design from one that acts. We set the level for each task with your team before any screen is drawn.
| Autonomy level | The product | The person | Interface it needs |
|---|---|---|---|
| Suggest | The productOffers options or completions | The personChooses and applies one | Interface it needsInline suggestions with accept and dismiss |
| Draft | The productProduces a complete first version | The personEdits and approves it | Interface it needsAn editable result, comparison with the source and partial regenerate |
| Act with approval | The productPrepares an action in another system | The personApproves each action | Interface it needsA preview of the change, approve and reject, and an activity log |
| Act within limits | The productActs without asking inside agreed limits | The personReviews afterward | Interface it needsLimit settings, activity history, undo or rollback and alerts |
One product often uses several levels: a drafting assistant can draft text but only act with approval when it sends an email. Your team decides which actions the system is allowed to take; we design how each level looks and how a person moves between them.
Principles
The human-AI interaction guidelines we design against
Microsoft Research's Guidelines for Human-AI Interaction, 18 guidelines published at CHI 2019, are the checklist Gestures reviews AI interfaces against, alongside Google's People + AI Guidebook. Four of the guidelines shape most screen decisions.
| Guideline | What it means on screen |
|---|---|
| Make clear what the system can do | What it means on screenFirst-run examples and stated boundaries before the first input |
| Make clear how well the system can do what it can do | What it means on screenSource coverage or confidence signals, and a note on when not to rely on a result |
| Support efficient correction | What it means on screenEdit in place, regenerate one part and revert, instead of starting over |
| Scope services when in doubt | What it means on screenA clarifying question or a narrower task when the input is ambiguous |
Waiting
Design the wait by how long the model takes
Model latency decides the waiting design, so we ask your engineers for real response times before choosing a pattern. Jakob Nielsen's response time limits set the thresholds: 0.1 second feels instant, 1 second keeps the flow of thought and 10 seconds is about the limit for keeping attention, after which people need to know when the result will be ready.
| Response time | Waiting pattern |
|---|---|
| Under 1 second | Waiting patternShow the result directly, with no loading state |
| 1 to 10 seconds | Waiting patternStreaming output or a progress indicator tied to real steps, with a stop control |
| More than 10 seconds | Waiting patternA background run with a notification, the steps completed so far and a way to leave and come back |
| Minutes, for agents and batch jobs | Waiting patternA job list with status, an expected finish time, partial results and a summary when the run ends |
Failure inventory
A failure inventory before the success screen
Every AI product fails in predictable ways, so we list the failures with your engineers before the main screen is designed and give each one a screen state, a message and a next action.
| Failure | What the person sees | Next action offered |
|---|---|---|
| A wrong answer stated with confidence | What the person seesThe sources or steps behind the answer, so it can be checked | Next action offeredCorrect it, report it or regenerate |
| A refusal or blocked request | What the person seesWhy it was refused, in plain words, and what is allowed | Next action offeredRephrase with guidance |
| A usage or rate limit | What the person seesThe remaining allowance and when it resets | Next action offeredWait, queue the request or upgrade |
| The model is unavailable or times out | What the person seesThe input, saved, and the status of the retry | Next action offeredRetry now or switch to another mode |
| Input too long or an unsupported file | What the person seesThe limit, stated before upload rather than after | Next action offeredTrim, split or convert the input |
Feedback loop
Feedback that reaches your evaluation set
AI product feedback becomes useful when the product collects it at the moment of correction and sends it to the team that maintains the evaluation set. We design the correction itself as the signal: an edited answer, a rejected suggestion or a chosen alternative, with an optional reason.
- Feedback tied to the exact part of the output, not one rating for a whole conversation.
- A reason picker with a few categories your team can act on, such as wrong fact, wrong format or missing source.
- A plain statement of what is logged and how it is used, written with your team.
- A way for your team to see the corrections in a review queue.
Our own AI product
Product decisions in Nomicam, the AI app we design and build
Nomicam, the AI photo editor Gestures designs and builds for iPhone, works at the draft level: people choose a template such as Night Flash or New Hairstyle, add a photo and judge the result. The template replaces the prompt, so nobody has to describe a style in words, and free, local editing tools handle adjustments that need no model at all.
Ragip Diler, who leads Gestures, also designed security, analytics and MCP workflows for AI products as Senior Product Designer at Venice, an independent design studio for AI and venture-backed companies.

Scope and cost
AI product design scope and cost with Gestures
AI product design with Gestures costs $2,250 per week, $4,000 for a two-week sprint or $7,500 per month for ongoing work. The plan follows the size of the task loop, not the number of screens.
| Scope | What it covers | Plan |
|---|---|---|
| First AI feature in an existing product | What it coversOne task loop: input, wait, result and correction, with its failure states | PlanTwo-week sprint, $4,000 |
| Agent approvals and activity | What it coversAction previews, approval steps, activity history and undo | PlanWeekly sprint, $2,250 per week |
| Review of an AI feature in use | What it coversThe failure inventory and correction flows checked against real outputs | PlanWeekly sprint, $2,250 per week |
| An AI product from the first version on | What it coversCore loops, autonomy levels, failure inventory and a design system | PlanMonthly partnership, $7,500 per month |
For an overview of the AI screens we design and who owns what in an AI project, see our AI and LLM startups page. To build the product as well, our AI-native product development service takes a first version from prototype to a working product. The website that sells it is covered by AI website design.
Monthly partnership
$7,500$10,500USD / month
Studio $7,500 + Growth $3,000 / month
For an AI product that changes with every model and feature release. Set the current focus each month and review what needs attention next.
Strategy · Design · Review
Work starts within 1 day of payment.
Cancel anytime.
Start your projectWeekly sprint
$2,250$4,000USD / week / 2 weeks
+ $3,000 / month for Growth
For one AI task loop, such as a drafting feature, an agent approval step or a review of an AI feature already in use. Agree the output set and the engineering contact before starting.
Save $500 with two weeks: $4,000 instead of $4,500.
Strategy · Design · Review
Work starts within 1 day of payment.
Cancel anytime.
Start your projectAdd Growth to a studio plan
SEO, content strategy and content production, alongside your design team. Growth is billed monthly. Cancel anytime.
+$3,000 / month
Compare all plans or book a call to agree the first task loop.
Questions
Questions about AI product design
How much does AI product design cost?
AI product design with Gestures costs $2,250 per week, $4,000 for a two-week sprint or $7,500 per month. A first AI feature with one task loop fits a two-week sprint. See pricing for every plan.
What do you need from our team to start?
We need a set of real outputs from your model, the response times your engineers measure, the usage limits and the actions the system is allowed to take, plus one engineer who can answer questions during the work.
Can you design before our model is ready?
Yes. Outputs from a comparable model or carefully written stand-ins let the design start, labelled as stand-ins in every file. The screens are reviewed again as soon as your own model produces real outputs.
Do you prototype with live models?
Yes, when timing matters. A prototype connected to a model shows how streaming, waiting and errors feel at real speed; a clickable prototype with recorded outputs is enough for layout and content decisions.
What does engineering receive at handoff?
Engineering receives the screens for every state in the failure inventory, the output set used in review, the waiting pattern for each response time, the copy for each message and annotated prototypes.
Do you design AI features for mobile apps?
Yes. Mobile AI features follow the same method, with extra attention to small screens, background runs and notifications. Our mobile app design and development service covers the rest of the app.
Is AI product design the same as AI-native product development?
No. AI product design is how the product works for the people using it. AI-native product development is how Gestures builds software with AI tools, with or without AI features in the product.
A conversation
Send us ten real outputs from your model
Send the task your product performs and a handful of real outputs, including one that went wrong. We can agree the first task loop and whether Gestures is the right fit. For the wider service, see product design.