Product Leadership · 8 min read
Five Ds, Two Gates, One Loop
A feature used to be done at launch. Now AI keeps changing what the product does, so the design work has to keep going too.
Three years ago, a feature was “done” at launch. The team researched, defined, designed, built and shipped it, then moved on to the backlog. Now an AI feature can behave differently next month than it did on launch day. An agent may even act inside the product on someone’s behalf. The old line doesn’t describe the work anymore. We need a process with feedback, oversight and room to change - without letting the product change in ways people can’t see or control. Launch is a checkpoint, not the finish. When a product keeps adapting, the team still owns the next round of observation, evaluation and correction.
What changed in three years
The shift isn’t just that teams can make prototypes faster. It’s that products can respond, generate and act after launch. That changes what we design, what we test and who is accountable for what the product does.
Nielsen Norman Group’s “State of UX 2026” (January 16, 2026) calls 2025 “post-hype AI” and 2026 “the year of AI fatigue.” It says organizations will “ask more of each role,” that “anyone will be able to make a decent-looking UI,” that “trust will be a major design problem,” and that “human direction, curation, and verification will continue to be essential.” That last point matters. A polished interface is not proof that the product is useful, safe or right.
At Config in May 2025, Figma launched Figma Make, which it called “a new prompt-to-app capability” for exploring and refining prototypes. Maze launched an AI moderator for user interviews in 2025 and expanded it on April 2, 2026. Maze’s own 2026 report says 69% of researchers use AI in some of their work and 82% say interpreting nuance and emotion still requires human judgment. These tools can accelerate work. They don’t remove the need to decide what matters.
A UX product designer’s plate now includes tasks that barely existed three years ago:
- Designing model behavior and system instructions as design material.
- Writing evaluation cases: test prompts with expected answers.
- Designing confidence, uncertainty and error states.
- Setting agent permissions and scopes.
- Designing receipts and undo for agent actions.
- Designing feedback capture into the product.
- Deciding how consent and data use work for learning.
- Designing for trust.
This is not a side quest for the design team. It is product design.
The 5Ds, rebuilt with gates
The five-stage version many teams use is Discover, Define, Design, Develop and Deliver. The stages still help organize work. The change is that each one needs two gates: an AI gate for what the machine does at speed, and a human gate for what a person must judge, approve or be able to override.
A gate is a decision point before work moves forward. It makes responsibility visible. AI can propose, sort, draft and check. People still decide whether the result fits the problem, the people and the risks.
| Stage | What changed | AI gate | Human gate | |---|---|---|---| | Discover | More signals can be gathered and synthesized continuously. | AI synthesizes voice-of-customer, support tickets, session data and AI-moderated interviews at scale. | A researcher decides which signals are real, sits in on sessions and protects consent. | | Define | Problems and possible outcomes can be drafted quickly, but choosing the right problem remains human work. | AI clusters problems and drafts hypotheses and success measures. | The product and design lead chooses the problem, writes the outcome, and writes the boundary of what the AI may and may not do. | | Design | Prototypes and variants can be generated rapidly, including working ones. | AI generates variants and working prototypes. | Design review checks the design system, accessibility and brand, plus judgment and taste. | | Develop | Parts of the build and checking process can be automated. | Agentic coding, automated evaluation runs and automated accessibility checks happen in the build. | A person signs off on model behavior and failure modes; the build blocks what a machine can check. | | Deliver | The product is monitored and may keep changing after release. | Staged rollout and live monitoring happen while the system learns. | Real oversight. |
The AI gate isn’t a rubber stamp. It asks whether the machine’s output is ready for the next stage and makes failures visible. The human gate isn’t a vague request to “keep a human in the loop.” It names the judgment a person must make and the control they need.
For high-risk AI systems, the EU AI Act’s Article 14 says they must be designed so they “can be effectively overseen by natural persons” while in use, including being able to override or reverse an output and to stop the system. That’s a concrete bar for oversight, not a decorative button labelled “human review.”
And Deliver feeds Discover. New usage, feedback, agent actions and task outcomes become inputs to the next round of work. The line becomes a loop.
Cybernetic UIs
In 1948, Norbert Wiener published “Cybernetics, or Control and Communication in the Animal and the Machine,” which MIT Libraries describes as “a new theory of control, feedback and communication.” The name comes from the Greek word for steersman.
A cybernetic UI is an interface built around a loop: it senses, compares, adjusts and shows. It senses behavior, feedback and outcomes. It compares them against the person’s goal. It adjusts ordering, suggestions, density or defaults. Then it shows what changed and lets the person steer or reset.
That last part is easy to miss. If a product changes itself but doesn’t show what changed, the person has no useful way to judge it. A cybernetic interface isn’t just an interface that adapts. It makes adaptation legible and leaves room for control.
Dynamic UX changes with the current context and content. Adaptive UX changes over time as the system learns from this person and from everyone. Those are different kinds of change, and they need different checks.
There are two learners here. The AI learns from data. The team learns from people. The first can find patterns across many interactions. The second can understand whether those patterns represent a real need, a frustrating workaround or a harmful mistake. Neither replaces the other.
The feedback system
Feedback isn’t a phase at the end of a project. It’s the product’s nervous system. The useful question is not just how much feedback arrives, but what each method can tell us and what it can’t.
| Method | What it tells you | AI’s role | Human’s role | |---|---|---|---| | In-app voice of the customer | Quick ratings and short reasons at the moment of use. | Group comments and surface patterns. | Decide what the signal means in context. | | Agent interactions | What people asked, corrected, undid or overrode - the strongest signal. | Identify repeated requests, corrections and overrides. | Review why they happened and what the product should change. | | Moderated testing | The why, with follow-up questions. | Help organize observations and notes. | Ask follow-ups and interpret the response. | | Unmoderated testing | Speed and scale. Nielsen Norman Group notes results can come back within hours and reach dozens or hundreds of users, but participants behave less realistically in tasks that depend on imagination, decisions or emotion. | Summarize results and identify patterns. | Choose tasks carefully and account for less realistic behavior. | | AI-moderated interviews | Scale. Maze says the AI moderator does not replace human moderators. | Run interviews at scale. | Judge nuance and emotion; decide what needs human follow-up. | | Guerrilla testing | Quick tests with passers-by on early concepts. | Help organize observations. | Watch what people do and ask brief questions. | | Blind testing | Unbranded comparisons that remove brand bias. | Compare responses across options. | Interpret what the comparison means for the design. | | Focus groups | Attitudes and language, not behavior. | Group and summarize responses. | Interpret what people say without treating it as observed behavior. | | Surveys | Trends at scale, not the why. | Summarize trends. | Decide what to investigate further. | | Shadowing and field observation | Human contextual observation of real work, plus AI-assisted session review with consent, judged by whether the real task got resolved. | Review sessions with consent and help surface patterns. | Observe the context and judge whether the task was resolved. |
The measure that matters most is real task resolution: did the person finish the job they came to do, and did they have to override the system to do it.
The case against constant change
People hate moving targets. Feedback prompts cause fatigue. Watching usage raises privacy concerns. Adaptive interfaces are harder to test, support and keep accessible. Those aren’t objections to wave away. They’re reasons to put limits on what changes and make those limits clear.
Adapt the content and the offer, not the map. Keep navigation and core layout stable. A person shouldn’t have to relearn where things are because the system thinks a different arrangement might suit them today.
Ask for feedback sparingly and at moments of completion. A short question after someone finishes a task is more useful than another prompt in the middle of one. Get consent and show what is learned. People should understand when their actions or feedback may inform changes.
The human gates exist exactly for this. They make the team responsible for what adapts, what remains fixed and what a person can stop, override or reverse. Without those decisions, “adaptive” can become a polite word for unpredictable.
Icon lens
An adaptive interface needs fixed meanings. Glyphs must not drift as the UI learns; what changes is order, emphasis and suggestion. A learning system also needs visible states such as “suggested,” “learned from you” and “changed by the agent.” In Digital Sanskrit, Agent (karta) is a root, so agent changes compose onto any object; the manifest is the fixed point an adaptive UI cannot rewrite.
Try this
**Goal:** Make the two gates real for one AI feature.
**Steps:** 1. Pick one AI feature. 2. For each of the five Ds, write its AI gate and its human gate. 3. Choose three feedback channels from the table and write when each fires. 4. Define one override or undo rate you will watch. 5. Write the one thing the system may never change on its own.
**Check:** For every stage, a named person can stop, override or reverse what the AI did.
**Reference:** Digital Sanskrit, Appendix A.7 (Building an icon system, step 12) and Appendix B.7 (Headless architecture).
Sources
- [Nielsen Norman Group, State of UX 2026 (Jan 16, 2026)](https://www.nngroup.com/articles/state-of-ux-2026/)
- [Figma, Config 2025 recap (Figma Make)](https://www.figma.com/blog/config-2025-recap/)
- [Maze, AI moderator, expanded (Apr 2, 2026)](https://maze.co/blog/ai-moderator-expanded/)
- [EU AI Act, Article 14: Human Oversight](https://artificialintelligenceact.eu/article/14/)
- [MIT Libraries, 150 Years in the Stacks: 1948, Wiener's Cybernetics](https://libraries.mit.edu/150books/2011/04/04/1948/)
- [Nielsen Norman Group, Unmoderated User Tests](https://www.nngroup.com/articles/unmoderated-usability-testing/)