The signals I see today point to more adaptive interfaces, more delegated actions, and a larger responsibility for designers to define the limits

The most important change in UX may not be visible on a screen.

Software is beginning to interpret goals, assemble interfaces, choose tools, and act on a person’s behalf. Google Search can now generate interactive graphs, simulations, and mini apps in response to a query. Google’s A2UI project lets agents compose interfaces from trusted component catalogs. Gartner expects task-specific agents to appear in 40% of enterprise applications by the end of 2026.

None of those signals proves what UX will look like in 2027. Together, however, they point to a shift in the object we design. A fixed flow describes what a person can do. An adaptive system must also define what the product can infer, generate, and do for that person.

My own use of AI is much less dramatic than the demos. I use it to organize information, recover details from transcripts, prepare placeholder copy, name layers, connect variables, and explore product data. In a financial organization, confidentiality and internal policy sharply limit where experimentation is appropriate. That gap between public capability and responsible use is not incidental. It is part of the design problem.

This is not a prediction that interfaces will disappear. It is an evidence-based hypothesis: if software takes on more decisions and actions, UX will need to spend more of its craft defining behavior, limits, and ways to return control.

The interface is not disappearing, but it may stop telling the whole story

Traditional software waits for a sequence of instructions. Agentic software is designed to interpret a goal, plan steps, use tools, and return with a result.

That changes what an interface needs to do. I can make this concrete with two ERP products I worked on earlier in my career. AppHealth had to be modernized and standardized across an ecosystem that included financial and management operations. AzulControle also concentrated operational rules, records, and reports in a conventional ERP structure.

Consider one recurring management task: understanding why the financial result for a period changed. In a fixed ERP, the manager first chooses a company or unit, selects a date range, opens a financial report, applies categories and status filters, compares the result with another period, and often exports the data to assemble an explanation elsewhere. Each step may be valid, but the system asks the person to translate a business question into its information architecture.

An intent-based version could begin with: “Why did this unit’s result fall compared with last month?” The system could identify the relevant period and unit, compare revenue and expense categories, surface the largest variations, and compose a chart for that question.

The interface would not end there. Before the analysis could support a decision, it would need to show which records were included, which definition of “result” it used, whether any data was missing, and how the comparison was calculated. If the manager asks the system to act, such as changing a forecast or approving a payment, permissions, confirmation, and a reversible record become mandatory.

This example changes the flow from navigation by system structure to navigation by intent. It does not eliminate the ERP’s complexity. It relocates some of that complexity into interpretation, where mistakes can be harder to notice.

This is where I think some discussions about AI move too quickly. Reducing visible steps is not the same as reducing complexity. Sometimes the complexity has only moved from the screen into the system’s behavior.

From fixed flows to constrained adaptation

Generative UI is often described as an interface created or adapted according to a person’s goal and context. Nielsen Norman Group connects the idea to outcome-oriented design, in which the experience focuses more directly on the user’s desired result.

We already see small versions of this when an AI conversation produces buttons, checkboxes, tables, or other controls instead of asking the user to type everything. The important part is not that AI can generate a button. It is that the system is beginning to choose which interaction makes sense at that moment.

Google is already testing a more literal version of this idea. Its 2026 Search announcements describe generative interfaces that build custom graphs, trackers, dashboards, simulations, and mini apps for a query. Google Research reported that people strongly preferred its generative UI prototypes to baseline language-model responses, while interfaces built by experts still received the highest ratings. The signal is meaningful, but so is the limitation: generation improved the response without replacing design expertise.

By 2027, more interfaces may behave like temporary explanations of what the system understood, what it recommends, and what it still needs from the user.

I do not think this will make fixed interfaces obsolete. Stable structures support learning, spatial memory, predictability, and confidence. These qualities matter even more in healthcare, finance, government, and enterprise products, where a vague status or misunderstood permission can affect an entire operation.

The useful direction, in my view, is not unlimited personalization. It is constrained adaptation. The system can respond to context, but inside boundaries defined by accessibility, business rules, permissions, consistency, and risk.

This is also why UX cannot become a recipe executed by an agent. What works for one audience, product, or operational context may fail in another. The difficult part is noticing those differences.

When software starts acting, trust becomes part of the interaction

There is enough movement around agents to take them seriously. Gartner projected that up to 40% of enterprise applications would include task-specific AI agents by the end of 2026, compared with less than 5% in 2025.

The counter-signal is equally important. Gartner also forecast that more than 40% of agentic AI projects may be canceled by the end of 2027, citing cost, unclear business value, and inadequate risk controls.

Together, these numbers suggest experimentation at scale, followed by a much harder question: which forms of autonomy create enough value to justify their cost and risk?

Brazil is not a side note. It is a stress test

The tradeoffs around agentic products become easier to see in markets where capability and maturity advance at different speeds.

Cetic.br reported that AI use among Brazilian companies grew from 13% in 2024 to 17% in 2025. The difference by company size was substantial: 50% of large firms used AI, compared with 15% of small firms. This is part of a broader pattern. Across OECD countries with available data, 52% of large firms used AI in 2025, compared with 17.4% of small firms.

Trust, then, may become a set of interaction primitives: permission, provenance, confirmation, receipt, reversibility, and escalationThe gap matters for UX because an interface can be identical while the conditions behind it are not. One company may have clean data, specialized teams, cloud infrastructure, and mature governance. Another may depend on fragmented records, manual processes, outsourced technology, and little capacity to monitor an agent after launch.

In that second environment, token consumption is not an abstract technical detail. It affects which interactions are economically sustainable. A generated analysis may look effortless while requiring repeated model calls, retrieval, validation, and infrastructure that the organization cannot support at scale. A system that depends on perfect data may also fail precisely where automation appears most attractive.

Brazil is therefore useful to a global discussion, not as an exception but as an intensified version of a common problem. The OECD describes an emerging AI divide in which larger firms, knowledge-intensive sectors, and more innovative regions move faster. Designing for 2027 means designing across that unevenness, including lower-cost fallbacks, explicit uncertainty, human escalation, and experiences that still work when the agent cannot.

For designers, trust can no longer mean only explaining a recommendation. If a system can act, people need to understand what it will do, which information it can access, when it will ask for confirmation, and how an action can be interrupted or reversed.

McKinsey’s analysis of the automation curve in agentic commerce makes a useful distinction: delegation is likely to vary according to trust, regret risk, and the value of human involvement. Reordering a familiar product is different from choosing an investment, approving a financial operation, or making a healthcare decision.

Regulation is moving in the same direction. Since August 2026, the European Union’s AI Act transparency rules have required people to be informed in certain situations when they are interacting with AI. The European Commission’s Article 50 guidance also addresses machine-readable marking of generated or manipulated content.

Trust, then, may become a set of interaction primitives: permission, provenance, confirmation, receipt, reversibility, and escalation.

Products may need to work for people and their agents

Another signal changes the meaning of the word user. Nielsen Norman Group describes AI agents as users, interacting with products alongside people. Google’s Universal Commerce Protocol offers a concrete example of infrastructure designed to let AI surfaces communicate with merchants and perform actions such as checkout.

The customer remains human, but a machine may search, compare, interpret, and operate the product before the person sees the final result.

This makes structured content, APIs, permissions, error handling, and machine-readable policies part of experience design. A person may never reach a product if their agent cannot understand its information, verify a rule, or complete a task safely.

It also creates a new risk. Optimizing only for agents could flatten brands and experiences into whatever is easiest for a machine to compare. Designing only for humans could make the product invisible or unusable in delegated journeys. UX may need to mediate both without forgetting who ultimately lives with the consequence.

Research can become faster without becoming automatically better

My own experiments with website data show both sides of AI-assisted analysis. Connecting analytics and Search Console data can help me identify patterns, profile the real audience of the site, understand how people arrive, and find opportunities for new content. It can bring signals together much faster than reading isolated reports.

But a generated persona is not automatically a real person, and a correlation is not an explanation. The output is useful when it helps me ask better questions. It becomes dangerous when a polished synthesis makes assumptions look like evidence.

Maze’s Future of User Research Report 2026 found that 69% of researchers were already using AI in their workflows. Respondents still considered interpreting nuance and emotion, making ethical decisions, and framing the right questions areas where human involvement was essential.

This matches what I see in practice. AI can help recover information from a call, group data, summarize patterns, and suggest hypotheses. The designer or researcher still needs to know what was lost in the summary, whether the sample supports the claim, and whether the finding makes sense in that specific context.

Research may therefore become faster at processing and more demanding at judgment. We will also need to research the AI itself, testing how it behaves across ambiguous requests, incomplete context, conflicting constraints, and failures. A successful happy path tells us very little about a probabilistic system.

Design systems are becoming context for machines

I am already seeing the beginning of this discussion in design systems. Teams are exploring Figma and FigJam automation, component generation, and ways to document a system so that an AI agent or MCP integration can understand components, variables, layers, naming, and usage rules.

This is close to what I have started testing with Figma’s AI. Layer naming or variable linking can save time, but only when the tool understands enough of the local structure. Without context, automation creates another cleanup task.

Figma’s 2026 AI Report, based on 8,403 survey responses and 639 qualitative interviews, found that the boundaries between design and development were already becoming more porous. The share of designers participating in development doubled to 41%, while developers doing design work increased from 44% to 60%.

If agents generate interfaces and code, design system documentation must explain more than what a component looks like. It needs to express when the component should be used, what it can contain, how states behave, which accessibility rules are mandatory, and which alternatives should be avoided.

Google’s A2UI project makes this direction tangible. Instead of allowing an agent to send arbitrary executable interface code, A2UI lets it describe an interface using components from a catalog controlled by the receiving product. The agent can compose an experience, but the client retains control over rendering, styling, security, and available components.

Components become vocabulary. Tokens become constraints. Documentation becomes context. Governance becomes a way to prevent plausible output from drifting away from the product.

Does the designer become an auditor?

There is an uncomfortable consequence to this shift. If machines generate more screens and states, designers may spend more time reviewing outputs, writing constraints, checking exceptions, and investigating failures. The craft can begin to resemble system auditing.

Some of that work is necessary. A probabilistic system cannot be evaluated only through a polished happy path. Someone needs to test ambiguous requests, conflicting permissions, incomplete data, accessibility failures, and actions that cannot easily be reversed. The object of critique expands from a screen to a range of possible behavior.

But turning designers into quality control for machine output would be a poor future for the profession. It would place UX at the end of the process again, approving what a model or engineering system has already produced. It would also confuse governance with design.

The more useful role begins earlier. Designers can help decide which intentions the system should support, which decisions it may make, what evidence it must expose, what remains stable, and when it must stop and ask a person. Writing agent-readable rules is part of that role, but so are research, framing, prototyping, and the creation of new interaction models.

The day-to-day craft may therefore split into two connected activities. One is generative: imagining the relationship between a person and a system that can interpret and act. The other is evaluative: testing whether that relationship remains understandable, inclusive, and accountable across many possible outputs.

This is more than prompt writing. A prompt describes a desired response. Product constraints define an operating space. They connect research, policy, data, components, permissions, business rules, and recovery. If designers only polish prompts, the role narrows. If they shape that operating space, the role becomes more consequential.

Accessibility cannot be a final check in a generative system

The WebAIM Million 2026 detected more than 56 million accessibility errors across one million home pages, an average of 56.1 per page. At the same time, the average number of elements on those pages grew 22.5% in one year, reaching 1,437.

That relationship between complexity and errors should make us cautious. If AI can generate new states, layouts, controls, and content, accessibility cannot depend on someone reviewing a few predefined screens at the end.

It must exist inside the components, constraints, generation rules, automated tests, human evaluation, and monitoring. Accessibility itself is not a trend. What may change is how impossible it becomes to treat it as a static checklist.

What the evidence suggests for 2027

The signals do not support a world without interfaces. They support interfaces that are more adaptive, conversational, and connected to systems capable of acting.

They also show a gap between enthusiasm and maturity. Agent adoption is accelerating while projects are being canceled for cost, weak value, and inadequate controls. Generative interfaces are becoming more capable while expert-designed experiences still outperform them. AI speeds up design and research work while human judgment remains essential for nuance, ethics, and framing.

Taken together, these are not separate contradictions. They describe the work ahead.

Figma’s State of the Designer 2026 found that 89% of 906 surveyed designers said AI helped them work faster, 80% said it improved collaboration, and 91% said the tools improved their designs. The same report presents craft as a human differentiator.

Faster execution does not reduce the need for judgment. It increases the number of possible outputs that someone must evaluate.

If a system can create alternatives, UX still needs to decide which problem deserves attention. If it can adapt an interface, UX needs to define what cannot change. If it can act, UX needs to establish permissions, recovery, and accountability. If it can summarize research, UX needs to decide whether the evidence supports the conclusion.

By 2027, the role of UX may include drawing fewer predetermined paths and defining more of the boundaries within which a system can create a path. That is still a hypothesis, not a certainty. But it is one that current products, research, standards, and adoption data now make possible to test.

The interface will remain important. It just may no longer be the whole experience.

References and further reading

• Figma, State of the Designer 2026. Research conducted with 906 digital designers across multiple regions about AI adoption, speed, collaboration, craft, and the changing role of designers.
• Google Research, Generative UI: A rich, custom, visual interactive user experience for any prompt. An experimental comparison of generated interactive experiences, baseline language-model responses, and expert-designed sites.
• Google, Search I/O 2026 updates. Product examples of Search generating graphs, trackers, dashboards, simulations, and mini apps in response to queries.
• Google, Introducing A2UI. An open project that allows agents to compose interfaces from component catalogs controlled by the receiving application.
• Figma, 2026 AI Report. A three-year study based on 8,403 survey responses and 639 qualitative interviews with designers, developers, and product managers.
• Maze, The Future of User Research Report 2026. Research about the adoption of AI in user research and the areas in which human judgment remains essential.
• Nielsen Norman Group, Generative UI and Outcome-Oriented Design. An analysis of interfaces that can adapt or generate elements according to the user’s context and desired outcome.
• Nielsen Norman Group, AI Agents as Users. A discussion about products being accessed by AI agents alongside human users and what this changes for UX and accessibility.
• Gartner, 40% of enterprise applications will feature task-specific AI agents by 2026. A forecast about the rapid integration of task-specific agents into enterprise software.
• Gartner, Over 40% of agentic AI projects will be canceled by the end of 2027. A counterpoint to accelerated adoption, highlighting cost, unclear value, and inadequate risk controls.
• Cetic.br, AI use by Brazilian companies reached 17%. Results from the TIC Empresas 2025 survey of 4,174 companies, including adoption differences by company size.
• OECD, AI use by individuals and firms in 2025. Official cross-country data showing the adoption gap between large and small firms.
• OECD, Emerging divides in the transition to artificial intelligence. Analysis of uneven AI adoption across firm sizes, sectors, and regions.
• McKinsey, The automation curve in agentic commerce. An analysis of how trust, regret risk, and the value of human involvement can determine the appropriate level of delegation.
• European Commission, Guidelines on transparency obligations under Article 50 of the AI Act. Official guidance about informing people when they interact with AI and marking generated or manipulated content.
• Google, Universal Commerce Protocol. Technical documentation for an open standard that allows AI surfaces to interact with merchants and perform commerce actions.
• WebAIM, The WebAIM Million 2026. An accessibility analysis of one million home pages, including detectable errors and changes in page complexity.

I write about UX, product design, and emerging technology at CamaraUX, where I publish practical articles, research, and free resources for designers.

---

UX in 2027 may be less about interfaces and more about behavior was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.