AI in Web and App Design in 2026: Threat or Tool?

AI changes the cost of producing design drafts; people remain accountable for accepting them. Models can generate screens, critique a brief, propose a smaller scope and help design failure states. Those suggestions still need evidence and review. We use Claude in daily work; our ILTY and TripLeads.AI projects had 744 and 788 automated tests respectively. Those counts describe the suites, not proof of quality or a measured AI productivity gain. The useful question is which work AI assists and who checks the result.

Key Takeaways

• AI changes the cost of producing drafts, not the responsibility for accepting them.
• Inspect product behavior, failure states and evidence.
• Test counts describe a suite, not proof of safety or quality.
• Pricing examples are dated observations, not a current buying guide.

What AI actually took over

The useful question is which steps AI can assist, including the cost of prompting, reviewing and correcting its output.

From what we see in our own work and in the tools we pay for, AI took over the production steps:

• First drafts of screens. A prompt can produce a plausible layout quickly. Check it against the actual content, constraints and user task before treating it as a design direction.
• Variations. Ten versions of a hero, a card or an empty state, where a designer used to make two.
• Boilerplate code. Component scaffolding, form wiring, test stubs and migrations.
• Placeholder content. Draft microcopy, alt text, sample data for a dashboard.

The clearest sign that this is now routine is that the tools meter it. On 6 October 2026 we read three pricing pages. Figma gives every seat a monthly allowance of AI credits (500 on Starter and on Collab, Dev and View seats, 3,000 on a Professional full seat, 3,500 on Organization, 4,250 on Enterprise), shared across "all Figma AI features and products." Notion's $20 seat plan includes Notion Agent, while its $0 and $10 plans get a trial, and Custom Agents are "$10 per 1,000 monthly Notion credits." Linear lists a Linear Agent and an agent platform on its plans, with some features marked "Requires AI credits." When a capability becomes a quota on a pricing page, it has stopped being a differentiator.

For a tool-by-tool view of what each one produces and costs, we compared ten of them in the best AI tools for UI/UX design. This post is about what that shift does to the work.

What we keep a person accountable for

AI can assist with all three activities below. In our process, a named person approves the decisions and verifies the evidence.

Deciding what to build. A model can challenge a requested screen or suggest that the problem sits upstream in onboarding. Treat that critique as a hypothesis. Compare it with user observations, product constraints and the cost of alternatives before committing the scope.

Deciding what good looks like. Fluent output can be easy to accept without checking. We write down acceptance criteria and use evals to test model output, alongside application tests and review. AI can help draft those criteria, but someone must check their coverage. TripLeads.AI, an AI lead-capture product we designed and engineered end to end, has 788 tests across the platform, from the parser that turns "around 200" into a guest count to the billing guardrails.

Designing for when the model is wrong. Consider the slow answer, the wrong answer, the uncertain answer and the request the product should refuse. In TripLeads, we separated conversational generation from deterministic checks: server-side guardrails, schema-shaped extraction and lead caps enforced in the database. Out-of-scope requests receive a redirect in the visitor's own language. A model can propose or help implement these controls; design and engineering still have to choose, integrate and verify them. The TripLeads case study shows how.

Where the threat is real

It would be dishonest to say AI threatens nobody. It threatens work whose value was mostly the production step:

• Template websites assembled from a brief with no strategy behind them.
• Handoff work: redrawing a designer's file as code, pixel for pixel.
• Stock-style imagery and filler illustration.
• Large volumes of near-identical marketing pages written to a pattern.

The World Economic Forum's 2025 jobs outlook reports employers' expectations for 2025–2030: UI and UX designers appear among the expected fastest-growing roles, while graphic designers appear just outside the ten expected fastest-declining roles. These are forecasts, not measured job losses or proof of an individual designer's prospects. We discuss that distinction in are UX designers being replaced by AI.

For a client asking for faster production, compare the full workflow, including review and rework. For a client asking what to build, AI can generate options; choosing among them still requires evidence about the business and its users.

Another risk is accepting a familiar-looking draft without checking whether it expresses the product's actual differences. Review the wording, imagery and interactions against the brief rather than assuming either AI or human authorship guarantees distinctiveness.

What changed in how we work

The biggest change is not speed. It is who can own a whole feature.

We gave one senior responsibility across design, prompts and production code so that interface decisions could be checked against model behavior. That was our operating choice, not a comparison proving that one person outperforms a team. We used it for ILTY, an AI mental-health companion: a native iOS app with five distinct AI companions and streaming chat on the Claude API, a marketing site of more than 226 pages and an admin dashboard, seven supported languages, delivered in eight weeks with 744 tests. The ILTY case study has the detail.

We cannot tell you how much of that schedule AI saved, because we did not build a control version without it. Scoping, evals and failure states remained explicit parts of the work, with human review even when AI helped produce them.

Designing with AI is not designing for AI

These are different jobs and they get confused.

Designing with AI means using it as a tool in your own process: drafts, variations, code. Ask a studio how it reviews those outputs and what it has delivered; tool use alone says little about the result.

Designing for AI means building a product whose core feature is a model: the input, the streaming, the edit and retry, the citations, the confidence cues, the undo, the human in the loop. That is a newer discipline with its own failure modes, and we wrote about the principles in designing for AI.

For a first AI feature, define one useful task, its failure states and an evaluation set before estimating delivery. The model's ability to produce a demo does not establish the complete product scope.

How to use AI without losing quality

A short set of rules we hold ourselves to:

1. Generate wide, decide narrow. Use AI to see ten options, then have a person pick one and say why.
2. Never ship a draft unreviewed. Fluent output hides errors. Read every line of generated copy and code as if a junior wrote it.
3. Write the definition of good first. For any AI feature, decide what a correct answer is before you prompt for one.
4. Design the failure path before the happy path. Empty, slow, wrong and refused are the states users remember.
5. Review the distinctive parts explicitly. Check brand, voice and signature interactions against the brief, whether the first draft came from a person or a model.
6. Budget it. Credits, latency and per-call cost are now design constraints, not engineering afterthoughts.

The short answer

Use AI to draft, critique and explore alternatives, then verify the output against explicit criteria. Keep a named person accountable for scope, acceptance and failure behavior. If the product itself depends on a model, design and test the interface around uncertainty as carefully as the generated answer.

Pricing observations above were recorded on October 6, 2026; check current terms at Figma, Notion and Linear. Those pages were reopened October 20, but the historical numbers here should not be treated as fresh checkout quotes. Project facts are linked to the ILTY and TripLeads case studies.