Design for AI

Why this matters

In 2025, more than half of people said they were nervous about AI. Even so, AI was already inside almost every product those same people used. That tension is the starting point.

Designing with AI isn’t designing one more system with integrations. The system learns. It anticipates. Sometimes it acts unprompted. What used to be a button becomes an intent. What used to be a screen becomes a dialogue, or nothing at all. What used to be deterministic becomes probabilistic.

When the work shifts like this, the old methods don’t quite stretch. This guide is a synthesis of the principles, patterns, and traps I see day-to-day with teams building with AI.

The fundamental shift

For 60 years we designed for determinism: if the user clicks button A, X happens. Always. Predictable, testable, mockable in Figma.

AI breaks this. The same prompt can return different answers in the same conversation. The same agent can pick one tool today and another tomorrow for the same task. Our weapon of choice, the static screen, stops describing what will happen.

This shift sits at the centre of almost every other topic in this guide. I dig into it in From deterministic to probabilistic design.

Anatomy of an agent

An agent is not a chatbot. A chatbot answers when you talk to it; an agent has already started working before you realised you needed it. The four components worth knowing:

• Tools: what the agent can do. Each tool is, at root, a bundle of prompts plus an API. Search hotels, send emails, read files, generate images.
• Memory: what it remembers between interactions. The context window is what’s alive right now; persistent memory is what survives into the next conversation.
• Learning: how it adjusts from experience. Could be fine-tuning, could be a simple update of rules from feedback.
• Plan: how it breaks an objective into steps. This is where the agent separates from the assistant.

I wrote a primer on this in The Agentic Era: what does this mean for Product Design?. For the five human-AI interaction types that still hold up, see 5 Human-AI Interaction Types.

Mental models for AI experiences

When a person interacts with AI, they’re applying one of three mental models. Picking the right one changes the tone, the expectation, and the level of control.

• Co-pilot: AI sits next to me, suggests, I decide. Code completion, email autocomplete, suggestions inside design tools.
• Assistant: I ask, I receive. Conversations with ChatGPT, Perplexity, AI search. The model is reactive.
• Learning Partner: AI learns with me, I learn with it, the relationship evolves. Adaptive tutors, tools that personalise over time.

Each model asks for a different interface. Mistaking an Assistant for a Learning Partner sets up personalisation expectations the product won’t meet. Covered in detail in Mental models for AI design.

Multi-agent and orchestration

For complex tasks, a single agent hallucinates. Too many tools, too much context, too many objectives competing. The fix is to split the work across specialist agents and orchestrate the collaboration.

Seven orchestration patterns show up in practice:

1. Sequential: agent A’s output feeds agent B. E.g. designer → accessibility → engineer.
2. Parallel: the agent duplicates itself to handle parts in parallel.
3. Hierarchical: a supervisor delegates to specialists.
4. Hand-off: the supervisor routes by request type.
5. Swarm: agents negotiate with each other in real time.
6. React: cycles of reasoning + acting with observation in between.
7. Magentic: a dynamic orchestrator that adjusts the plan on the fly.

The choice of pattern has direct UX implications. Hierarchical asks for attribution (which agent did what). Parallel asks for simultaneous progress. I go deeper in Multi-agent orchestration for designers.

UX for agentic experiences

Three problems show up in almost every AI product:

Capability discovery. People don’t know what the agent can do. In a traditional interface, the buttons advertised themselves. In a conversational or agentic interface, the user looks at a text box and freezes. Design has to surface capability without listing everything in a menu. More in Capability discovery in agentic UX.

Observability. When the agent works in the background, trust dies. Claude Code nailed this: it shows the plan before executing, lists the steps as it goes, lets you pause and cancel. It’s the visibility of system status heuristic applied to systems that think. Covered in Observability in agentic UX.

Cost communication. Every token costs money. In tools like Claude Code or Cline, the real-time cost counter is a UX decision, not only a pricing one. When users see the cost, they tighten the prompt. When they don’t, they get a surprise bill.

Prompt as design work

Two years ago, “prompt” was an engineering term. Today, in teams shipping AI, designers write prompts side by side with data science. It’s a new discipline, with its own aesthetic.

Five patterns worth learning:

• Few shot: examples before the task.
• Chain of thought: ask for step-by-step reasoning.
• React: cycles of reasoning and acting.
• Reflection: the agent looks at its own output and adjusts.
• Reverse engineering: design the ideal message, then build the prompt that produces it.

Reverse engineering is the most natural method for designers. Instead of guessing the prompt from scratch, mock the output you want and work backwards. Covered in Prompt engineering as design work.

Ethics, limits, and boundaries

A real example: in an AI-driven property search, the agent relaxes filters when it can’t find results (3 bedrooms becomes 2, 100m² becomes 80m²). But if the user asked for ramps or grab bars, that’s an accessibility filter. It can never be relaxed.

Boundaries like this aren’t optional. They’re where design has to impose deterministic rules inside a probabilistic system. The prompt has to say, in big letters: “never relax accessibility filters.”

More on the principles guiding these decisions in Ethical principles in AI design and on the ethics in AI page. On the biases we inherit when designing with AI, read Cognitive biases explained.

Where to start this week

If you’ve never seriously worked with an agent, pay for a subscription (Claude, GPT, Perplexity) for a month. You can’t learn to design for AI without using AI properly.

After that, three concrete exercises:

1. Pick a product you work on and identify the points where AI could act instead of just suggest.
2. Write a prompt for a repetitive task of yours, in OpenAI’s Playground or equivalent, with chain of thought and few shot.
3. Map the boundaries of your product. Where can the system never improvise?

FAQ

What is Design for AI?

The practice of designing products and experiences with AI systems. It covers trust, explainability, error correction, human control, and the mental models users apply when interacting with AI.

What’s the difference between deterministic and probabilistic design?

Traditional design draws exact mockups for systems where output is predictable. In AI, the same input can produce different responses, so design becomes examples, properties and boundaries instead of fixed mockups. More in From deterministic to probabilistic design.

How does prompt engineering relate to design?

Prompts are conversations where you specify tone, format, rules and context. It’s design work: defining what the agent answers and how, translating the user’s problem into structured instructions. More in Prompt engineering as design work.

What changes with agents instead of assistants?

Agents are autonomous: they decompose goals into steps and execute without constant supervision. Assistants wait for commands. Design changes: you need observability, control, and attribution. Covered in The Agentic Era and Mental models for AI design.

What is capability discovery?

Patterns to help users discover what they can do with an AI agent, especially in conversational interfaces where capabilities aren’t visible. Detail in Capability discovery in agentic UX.

To close

AI won’t replace designers. It will replace designers who don’t know how to design with it. This guide is a starting point, not a destination. If something here stuck, or if you disagree out loud, write.