White humanoid robot extending an open hand toward the camera in a bright room with a grey floor
White humanoid robot extending an open hand toward the camera in a bright room with a grey floor

•

8 mins read time

Designing for AI Agents as Users: When Your Next Customer Is Software

Designing for AI agents means giving every key task a machine-readable front door. Human UI vs agent-ready product, the Legible, Callable, Accountable test, and the designer's fear answered.

Ayoub Kada

•

8 mins read time

Designing for AI Agents as Users: When Your Next Customer Is Software

Designing for AI agents means giving every key task a machine-readable front door. Human UI vs agent-ready product, the Legible, Callable, Accountable test, and the designer's fear answered.

Ayoub Kada

Agents will not end interface design. They make unclear flows fail loudly, and they reward products that are legible, callable and accountable.

Designing for AI agents: your next customer may not have eyes

Designing for AI agents means building products that software can understand and operate on a person's behalf, not just products a person can tap through. The short answer for product teams: keep designing for humans, but give every important task a second, machine-readable front door with clear names, explicit inputs, predictable results and a human checkpoint before anything consequential happens.

This is no longer theory. The Model Context Protocol describes itself as "an open-source standard for connecting AI applications to external systems", and assistants such as Claude and ChatGPT already use it to reach tools and data. On the web, the WebMCP draft from the W3C Web Machine Learning Community Group proposes a way for web pages to offer JavaScript tools that agents can call. It is a draft community report, not a W3C standard, but the editors come from Google and Microsoft, which tells you where browsers are looking.

The hype vs the reality

The hype says websites and apps are about to become invisible pipes, that agents will do all the shopping, booking and admin, and that interface design is a dying craft. That story skips over how often agents still fail on real interfaces: ambiguous buttons, multi-step modals, hidden state, captchas, pop-ups and forms that only make sense if you can see the layout.

What holds up is narrower and more practical. A growing share of tasks will start in an assistant rather than in your app. Some of those tasks will be completed by an agent calling your product directly. The products that agents handle well are the ones that were already clear, structured and honest about state. The products that agents handle badly are the ones that rely on visual tricks to steer people.

Verdict: the "no more UI" claim is overhyped. The "agents become a real user segment with their own needs" claim is underhyped, and it lands first in commerce, booking, support, internal tools and SaaS admin.

If you have read the DIGCY pieces on agentic UX design and apps without screens, this is the other side of the same coin. Agentic UX is about your product acting for the user. This piece is about other people's agents using your product.

Human-first UI vs agent-ready product

The table below is the quickest audit you can run on an existing product.

Design decision

Human-only UI

Agent-ready product

How a task is found

Navigation, menus, search

Named actions with plain descriptions

Inputs

Form fields inferred from labels and layout

Explicit parameters with types and allowed values

Feedback

Toasts, colors, animations

Structured result plus a readable summary

Errors

Red text near a field

Error codes with a reason and a suggested fix

Confirmation

A modal the user clicks through

A preview the agent must show the human before committing

Identity and permission

One login, full access

Scoped, revocable access for a delegated agent

Pricing and terms

Shown in the checkout flow

Stated before the action, in data, not only in pixels

Persuasion

Urgency banners, upsell steps, dark patterns

Removed or ignored, so honest value must carry it

Look at the last row carefully. An agent does not feel urgency from a countdown timer. Tricks that push humans into choices become noise or, worse, reasons for an agent to treat your product as unreliable. That is good news for honest products and bad news for growth teams that lean on friction.

The Legible, Callable, Accountable test

A simple framework for deciding whether a task in your product is ready for agents. Run your top ten tasks through three questions.

Legible. Can an agent understand what this thing is and what state it is in without seeing the screen? Product names, prices, availability, order status and policies should exist as text and structured data, not only as images or layout. If your pricing page only makes sense visually, an agent will get it wrong, and the human will blame you.

Callable. Can the task be started with a clear action name and a small set of explicit inputs? "Book a 30 minute demo for next Tuesday afternoon" should map to one action with a date, a duration and a contact. If the task needs seven screens and a hidden toggle, it is not callable yet. MCP tools and WebMCP tools both push teams toward this shape: a name, a description, a schema for inputs and a predictable output.

Accountable. When the agent acts, can the human see what happened, approve it before it matters and undo it after? Designers earn their keep here. Previews before payment, summaries a person can scan in seconds, activity logs that name the agent, and an easy way to revoke access. Without this layer, agent access is a liability you will end up switching off.

A task that passes all three can be offered to agents. A task that passes only Legible should be read-only for agents for now. A task that fails Accountable should never be exposed, however easy it is to call.

What product teams should do in the next 12 months

You do not need an "agent strategy" deck. You need a handful of changes that also make the product better for people.

  1. Write an action inventory. List the jobs users come to do, phrased as sentences a person would say to an assistant. This list becomes the backbone for tools, help content and onboarding at the same time.

  2. Separate meaning from layout. Every price, limit, status and policy should exist as text and data, not only as styling. This also helps answer engines, which is why it overlaps with the shift described in the future of landing pages.

  3. Design the human checkpoint. For any action that spends money, sends a message, shares data or deletes something, design the preview the human sees before the agent commits. Treat it as a component in your design system, not a one-off modal.

  4. Design delegated identity. Who is acting, on whose behalf, with which scope, until when? Emerging efforts such as the Personal Agent Protocol and agentic checkout flows like Google Universal Cart show the direction: delegation needs visible, revocable consent.

  5. Rewrite your errors. An agent cannot guess what "Something went wrong" means. Neither can a person. Give every failure a reason and a next step.

  6. Measure agent traffic separately. If agents start completing tasks, your funnels, A/B tests and conversion metrics change meaning. Agree early on how you will tell agent sessions from human ones.

Where this does not apply yet: products whose value is the experience itself, such as games, creative tools and social feeds, gain less from agent access. Expose read-only data there first and keep the crafted experience for humans.

Where design goes over the next one to three years

Expect designers to own two artefacts for every important flow: the screen a human uses and the contract an agent uses. The contract covers action names, input rules, result summaries, confirmation moments and failure messages. It sounds like engineering, and engineers will build it, but deciding what a task is called, what a human must approve and what the summary says is product design work.

Expect brand to move too. When an agent compares three services, the visual identity matters less in that moment. What matters is clear terms, reliable results and a reputation for not surprising people. Brand becomes behaviour you can verify, and the screen becomes the place people come back to when they want to check, change or enjoy something.

Teams building this layer usually pair interface work with AI product development so the agent contract and the screens are designed together rather than bolted on.

The fear: if agents use the product, who needs designers?

This is the worry behind most questions on this topic, and it deserves a direct answer. Agents remove some screens. They do not remove decisions. Someone still has to decide what actions exist, what they are called, what counts as safe, when the human must step in, and how the product explains itself when things go wrong. Those are design decisions, and agents make them more visible, because a confusing flow now fails loudly and at scale.

The designers who should worry are the ones whose work is mostly decoration on top of unclear flows. The ones who will be in demand can write a clear action, map a permission model, design a confirmation that people actually read, and work with engineers on schemas. For a wider take on the replacement worry, read will AI replace UX designers.

FAQ

What does designing for AI agents mean?

Designing for AI agents means making a product's tasks understandable and operable by software acting for a person. In practice that means named actions, explicit inputs, structured results, clear errors and a human approval step before anything consequential, alongside the normal interface people use.

Do I need an MCP server to make my product agent-ready?

Not on day one. Start with legible content, clear actions and good errors, which help every agent and every human. An MCP server, or WebMCP tools on a website, is the next step once you know which tasks you want agents to perform and how humans will approve them.

Will AI agents replace website and app interfaces?

No. Agents will complete some routine tasks without opening your interface, but people still come back to check, compare, change and enjoy things. The interface becomes the place for trust, oversight and experience, while agents handle repeatable jobs through a structured layer.

Will designing for AI agents replace UI designers?

No. It shifts part of the work from screens to contracts: action names, permission scopes, previews and summaries. That is still design. Designers who learn to define these pieces alongside screens become more valuable, not less.

Is designing for AI agents overhyped?

The idea that every product must become a headless API this year is overhyped. The practical version, making key tasks legible, callable and accountable, is real and already useful for search, support and assistant integrations. Start there.

Start with ten tasks

Pick the ten tasks people most often come to do and run them through the Legible, Callable, Accountable test. The gaps you find will improve the product for humans first, and they will tell you exactly what an agent integration needs. If you want a partner for that work, a UI/UX agency in Casablanca like DIGCY works with teams globally on web app design and AI products. DIGCY is a Dribbble Selected Agency, and its work on ZeroDrift, which raised $2M from a16z, is a good example of designing an AI product people need to trust.

Let’s keep in touch.

Discover more about high-performance web design. Follow us on Twitter and Instagram.

Building an AI product?

By submitting, you agree to our Terms and Privacy Policy.

Let's Talk

Tell us about your AI product. After you send this, you can pick a 20-minute slot on our calendar. Prefer email? hello@digcy.com We reply within one business day.

Building an AI product?

By submitting, you agree to our Terms and Privacy Policy.

Let's Talk

Tell us about your AI product. After you send this, you can pick a 20-minute slot on our calendar. Prefer email? hello@digcy.com We reply within one business day.

Building an AI product?

By submitting, you agree to our Terms and Privacy Policy.

Let's Talk

Tell us about your AI product. After you send this, you can pick a 20-minute slot on our calendar. Prefer email? hello@digcy.com We reply within one business day.