Skip to content
Article Nielsen Norman Group Oct 2026

NN/g: the role of UX in agentic software factories

What the article is about

Tanner Kohler of Nielsen Norman Group examines the AI-native software development lifecycle (SDLC) playbooks published by Anthropic, Google, and OpenAI — frameworks in which AI agents handle most of the building while humans approve work at defined gates. He argues that these playbooks reduce UX to background information that AI consults, rather than treating it as a research and decision-making process embedded throughout development.

Context: how the playbooks work

The playbooks describe a pipeline: the team states an intent, AI generates a plan, AI creates mockups, agents write code, AI verifies the build against the mockups, and AI monitors the deployed product. Humans approve outputs at each gate. The underlying assumption is that capturing UX knowledge in documents upfront is sufficient for AI to apply it correctly throughout the lifecycle.

Kohler identifies three distinct roles UX plays in any software project: deciding what to build, evaluating what gets built, and supplying context for those doing the work. The playbooks address only the third.

Where the playbooks fall short on UX

In the plan phase, no step exists to surface new user needs. Proposals can be made in isolation without the team input that typically catches blind spots.

In the design phase, AI applies visual design standards and produces mockups, but there is no validation step against user research. Visual polish is not the same as meeting user needs, and a well-structured mockup can still represent a bad product decision.

In the test phase, AI verifies the build against the team’s own mockups and specifications — not against users. Problems that were present in the original design pass through unchallenged.

In the maintain phase, monitoring focuses on technical health. A feature that functions correctly but consistently confuses users would go undetected until it surfaces as a support ticket or churn signal.

What UX teams should do

Kohler recommends several adjustments: build user-need questions into the intent template so researchers can document problems before solutions are proposed; work with product owners to define user-value acceptance criteria alongside the functional acceptance criteria; specify which changes require a human UX review, such as navigation changes or flows that break established patterns; add usability signals — task abandonment rates, support-ticket spikes — to monitoring dashboards; and take ownership of making design system knowledge and research insights machine-readable so AI agents can actually use them.

Who it is useful for

UX practitioners and design leaders in organizations that are adopting or evaluating AI-native development approaches. The article is a practical starting point for teams trying to understand what UX must actively claim in agentic workflows rather than waiting to be invited.