A feature ships, every dashboard stays green because the code works, and people quietly stop using it. Nobody can say why, because nothing in the process was built to notice. That is the gap in the new playbooks the AI labs publish for running software development through agents. In a 30 October 2026 analysis, Nielsen Norman Group’s Tanner Kohler read Anthropic’s, Google’s and OpenAI’s versions and found that all three “sideline UX”. Teams are told to feed the AI their design rules. No step checks whether real people want, or can use, what comes out.
Design becomes a file the agent reads
Take Anthropic’s AI-native SDLC playbook , published 21 August 2026. Its design stage says requirements and design “collapse into one session”. The product owner mocks the design up in Claude Design, Anthropic’s beta design tool, and exports it to Claude Code. The owner “reviews that spec, but doesn’t write it”, and a tech lead decides on higher-risk items.
The prerequisite is telling: “brand, security, compliance, and UX policies written as skills”. UX arrives as a rulebook the agent consults. Designers appear only in the playbook’s description of the old process, where analysts write requirements and “designers then parse those back into a design”. Kohler reports “no mention of user research whatsoever”, and says discovery drops out because ideas begin as AI-generated intent files.
A flow can confuse everyone and still look healthy
The Maintain stage makes the gap concrete. A script watches production for control-band breaches: CI test failure rate, post-deploy 5xx rate, PR cycle time. Support tickets appear only as a trigger for work. Adoption and abandonment appear nowhere, and no designer or researcher is named as an approver at any stage. So a flow can confuse every user and still look healthy by every metric the playbook asks anyone to watch.
Kohler’s phrasing is blunter: “There are no usability-focused signals built into the monitoring system.” He also notes that the AI checks output against team artifacts, not against user research, so evaluation drops out alongside discovery.
These playbooks did not remove design. They turned it into inputs, and dropped the two jobs that cannot be written down ahead of time.
Consider who this changes. A UX researcher or product designer at a company adopting such a playbook stops being the person who can hold a release over a confusing flow. They write the UX skill file the agent reads. Their findings have no stage to go into and no gate to stop anything, unless they deliberately take one back.
The playbooks are not blind, but they are quiet
The fair objection is that these documents do mention design. Anthropic puts “design review” before any code is generated and lists “affected users” in its intent template. OpenAI’s guide says agents let teams prototype in high fidelity early, enabling “customer testing far sooner”, and leave designers “more time evaluating user flows”. Faster prototypes could mean more contact with users, not less.
Two things weaken that reading. Anthropic’s review happens on a document, “when changing course is still a matter of editing a document”, which reads as a check on the spec rather than on anyone’s experience of it. And OpenAI’s own delegate/review/own table gives the evaluation no step: the team reviews that components follow conventions and accessibility standards, and owns the “final direction of the user experience”. Testing with customers is promised in prose and absent from the table. NN/g, a UX consultancy, has a stake in UX keeping a formal role, though it also flags that AI labs benefit when companies adopt these models.
What the text shows, and what it doesn’t
Be precise about the reach here. These sources analyse what three vendor playbooks prescribe: stages, approvers, monitored metrics. They do not measure what adopting teams actually do, or whether products built this way show worse usability or adoption. NN/g’s failed-launch story is explicitly a hypothetical.
Read across from text to practice, though, the pattern suggests a quiet trade. A team that follows the document as written would have no scheduled moment to learn what is worth building, and no scheduled moment to learn whether it works. That echoes NN/g’s survey finding that saved time moves downstream : speed arrives, and the checking work goes unowned.
Kohler’s fixes are modest and specific: user-value acceptance criteria in the intent template, defined triggers for UX review or quick usability tests when navigation changes or a flow breaks convention, and “UX distress signals” such as task abandonment, conversion drops and support-ticket spikes in monitoring. None requires slowing the agent. Each needs someone with standing to read the signal.
Until a playbook names someone to ask the question, the green dashboard will keep answering a different one. Whether anyone can use the thing is, for now, the one check that has to be taken back by hand.
