Most design organizations that describe themselves as AI-enabled have done the same thing: adopted useful tools at individual stages.
Research summarizes interviews, design generates concepts, content drafts variations, and delivery drafts acceptance criteria. Every one of those is a real improvement, and collectively they have produced very little change in how long projects take.
The reason is the subject of this article, and it is not a tooling gap.
At every transition in the design lifecycle, context gets compressed into an artifact and then decompressed by someone else. Research becomes a deck. The deck becomes requirements written from memory of the deck. Requirements become a journey map. The journey map becomes screens. The screens become a handoff document. Each compression discards the reasoning and keeps the conclusion.
Accelerating the work inside each stage does nothing about the losses between them. It just gets you to the next lossy transition sooner.
What "Connected" Means
The useful definition is narrow: information generated at one stage remains available, in a usable form, to the stages that follow, and the connection is traceable in both directions.
Concretely, that means a designer can ask what research supports this journey stage and get an answer. A developer can ask why this step exists and get the requirement rather than an opinion. A researcher can ask which design decisions cited this finding. A measurement lead can ask which customer need this feature was meant to address and compare it to what happened.
None of that requires new capability. It requires that the reasoning survive the transitions, which historically it has not, because the reasoning lived in the heads of people who were in the room.
The best available evidence that context changes output is empirical rather than conceptual. Guriță and Vatavu, at the 2025 Web for All Conference, compared UI code generated from prompts that said nothing about accessibility against the same tasks with accessibility explicitly specified. Expert evaluation found a 58 percent violation rate in the first condition and 19 percent in the second. Keyboard navigation went from failing 80 percent of the time to failing none of the time.
That is a large, measured effect from supplying one category of context. It is also an incomplete one, since 19 percent is not zero, and accessibility is a single dimension among the many that determine whether generated work is usable.
Extrapolate carefully from there. Nielsen Norman Group's parallel framing, which they call UX-context design, is explicitly presented as a hypothesis rather than a result: their language is that experiments "suggest" curated context improves generated UI while "important questions remain," including which artifacts help most and how to measure context quality. Treat the W4A numbers as evidence and the broader framing as a reasonable extension of it, and be skeptical of anyone presenting the whole model as settled.
The Stages, and What Carries Between Them
Discovery into synthesis. What should carry forward is the evidence and its provenance, not the summary alone: which participant said what, in what context, how strongly, and what contradicted it. Teams that carry only the summary lose the ability to revisit a conclusion when new evidence arrives, which is why organizations relitigate the same customer questions every eighteen months.
Synthesis into design. What should carry is the prioritized need with its supporting evidence attached, plus an explicit record of what remains assumption. That last part is the one everyone skips, and it is the difference between a team that knows which parts of its plan are load-bearing and one that discovers it during a launch.
Design into prototype. What should carry is constraint: approved components and their states, accessibility requirements, technical limits, and business rules. Post 5 covers why generation without these produces polished output that cannot be built, and the W4A violation rates above are the cleanest measure of the cost.
Prototype into delivery. What should carry is behavior and rationale as well as appearance. Post 6 covers this in depth, because it is where the largest losses occur and where DORA's throughput-plus-instability finding actually originates.
Delivery into measurement. What should carry is the original claim: this design addresses this customer need, and here is how we will know. Without it, post-launch analytics tell you what happened and never whether it was what you intended.
Measurement back into discovery. This is the loop almost nobody closes: which assumptions were proven, which were wrong, and what the next team should not have to rediscover.
Where AI Fits in This Picture
Here, AI functions as a retrieval and translation layer across every stage, rather than a generator at each one individually.
That reframing matters because it changes what you would build. If AI is a generator, you buy tools per stage and measure output volume. If AI is the connective layer, you invest in making organizational knowledge structured, permissioned, and retrievable, and you measure whether context survives transitions.
The second investment is less exciting and considerably more durable. Tools will keep changing. A well-structured research repository, a design system that encodes behavior rather than appearance, and a traceable line from finding to requirement to interface will still be valuable in five years regardless of which vendor won.
Nielsen Norman Group's review of AI use across the research process points the same direction: AI performs strongest in planning and analysis, the stages that are about organizing and retrieving information, and weakest during live sessions, the stages that require being present with a human being. Retrieval is the capability. Presence is not.
Why This Is the Seventh Article and Not the First
Because almost nobody should start here.
Rebuilding how an organization carries context is a multi-quarter effort involving research operations, design system investment, delivery process change, and governance. It is the right destination. It is a terrible opening move, for three reasons.
It cannot be justified before anything has been proven. It requires cross-functional agreement that no design leader can secure on the strength of a thesis. And it delays every benefit until the whole structure exists, which means the first eighteen months look like cost.
The sequence that works runs the other way. Fix one bounded, measurable thing. Post 2 describes a good candidate: an accessibility and heuristic sweep that produces a prioritized remediation list with a defensible return. Then fix a second one. Somewhere around the third, the pattern becomes obvious to people outside the design organization, and the structural conversation happens with evidence and allies instead of a slide and a hope.
The organizations that arrive at a genuinely connected workflow mostly did not set out to build one. They fixed a sequence of specific, annoying, expensive problems and noticed that the fixes shared a shape.
The Test for Whether You Have One
Ask someone in delivery why a particular interface element exists. Time how long it takes to get an answer that traces to a customer need rather than a preference.
If the answer arrives in a minute, the context is connected. If it takes three days and two meetings, or if it never arrives and the team just decides again, that is the gap, and no tool at any individual stage will close it.
Read: Building the AI-Enabled Experience Design Workflow for In-House Teams
Sources
- Alexandra-Elena Guriță and Radu-Daniel Vatavu, "When LLM-Generated Code Perpetuates User Interface Accessibility Barriers, How Can We Break the Cycle?" Proceedings of the 22nd International Web for All Conference (W4A '25), April 2025.
- Tony Alicea, "UX-Context Design: Using UX Knowledge to Inform AI-Generated Design," Nielsen Norman Group, July 24, 2026.
- Kate Moran and Maria Rosala, "Accelerating Research with AI," Nielsen Norman Group, September 27, 2024, reviewed January 21, 2026.
- DORA, "Balancing AI tensions," 2025.