Every design team I speak to has already crossed the line. Someone is pasting interview notes into a chat window. Someone is generating five layout directions before the kickoff call. Someone has an AI summarising a survey nobody had time to read. The tooling arrived faster than the judgement about how to use it, and that gap is where the risk sits.
This is not an argument for or against AI in UX. That argument is over. The useful question now is narrower and much harder: for any given design task, how much of it should a machine do, and what has to stay human?
What follows is the framework we use at Adi Design to settle that question task by task, with an interactive version you can run against your own work.
The short version
- Treat AI involvement as a five-level dial, not an on/off switch.
- Four questions place any task: consequence, evidence, verifiability, visibility.
- Work you would not be willing to disclose never belongs in a deliverable.
- If you cannot verify it and the stakes are high, no level of delegation is defensible.
- Log what was generated. The record is the control, not the intention.
What actually changes when AI enters a UX process
Three things change, and only one of them is the one people talk about.
Speed changes, obviously. A first draft of anything — copy, a flow, a research summary, a component spec — now costs minutes instead of hours. That is genuinely valuable, and pretending otherwise is posturing.
The cost of being wrong changes shape. Bad work used to look bad. It was slow, thin, visibly unfinished, and a reviewer caught it. AI-assisted work that is wrong looks finished. A fabricated user quote reads exactly like a real one. A persona built from nothing carries the same confident tone as a persona built from twelve interviews. The error no longer announces itself, so the review has to go looking for it.
The centre of gravity moves toward the average. These models are trained on what already exists, so left alone they return the most common solution to your problem. Often that is correct — most date pickers should look like a date picker. But if your brief needs a distinctive answer, the model is pulling in precisely the wrong direction, quietly, and you will not notice because the output is competent.
Bad work used to look bad. Now it looks finished.
The five failure modes worth naming
- Synthetic users. Personas, journeys and “user needs” generated without a single real participant, then treated as evidence in the next meeting. The most damaging one, because it corrupts every decision downstream.
- Research laundering. Real data goes in, a confident summary comes out, and the nuance — the contradictions, the one participant who broke the pattern — quietly disappears. Summaries are lossy, and the loss is not random.
- Accessibility theatre. A model will happily tell you a design is accessible. It can check contrast ratios and ARIA attributes. It cannot tell you whether a screen reader user can complete the task.
- Data leakage. Client NDAs, unreleased product plans, participant names, screeners, recordings — pasted into a tool with no agreement covering it.
- Silent authorship. Work shipped with no record of what was generated, so nobody can later answer “where did this recommendation come from?”
A simple test for any AI-assisted deliverable: if a client asked you to show your working, could you? If the honest answer is no, the task was delegated too far.
The framework: five levels of delegation
Rather than a blanket policy, treat AI involvement as a dial with five settings. Every task is placed on one of them, deliberately, before the work starts.
Level 0 — Out of the loop humans only
No AI touches this task or its inputs. Reserved for anything involving live participants, confidential material under a restrictive NDA, and final accountability for a recommendation. Moderating a usability session sits here. So does signing off an accessibility claim.
Level 1 — Sounding board think, don’t produce
AI is used to think, never to produce. Argue with it, stress-test a hypothesis, ask what you might have missed. Nothing it writes enters the deliverable. The right setting for sensitive strategy work where you want a second perspective but the output must be yours.
Level 2 — Divergence engine options, not answers
AI generates options; a human selects, discards and rewrites. The value is breadth, not quality — you want twenty naming directions so you can reject nineteen. Nothing survives untouched. Most early-stage ideation belongs here.
Level 3 — Drafts with sign-off verified by a named human
AI produces a complete first draft that a human edits, verifies against source material and takes responsibility for. Where most production work lands: synthesising notes you took yourself, drafting documentation, turning a defined flow into copy. The sign-off is not a formality — it is the whole control.
Level 4 — Supervised autopilot spot check and ship
AI drafts and the output ships after a spot check rather than a line-by-line review. Only appropriate when the task is low-consequence, reversible and trivially verifiable: alt-text first passes, component documentation, formatting and conversion, changelog entries.
Four questions that place a task on the dial
- Consequence. If this output is wrong and nobody catches it, who pays — and how badly? A clumsy tooltip is a nuisance. A wrong assumption inside a persona costs a quarter.
- Evidence. Is this grounded in real user data that exists, or is the model being asked to invent the ground truth? Summarising ten interviews is a different act from imagining them.
- Verifiability. Can a human check this in reasonable time? If verification takes longer than doing the work yourself, delegation has bought nothing but risk.
- Visibility. Would you be comfortable telling the client and the end user exactly how this was made? Discomfort here is almost always information — it usually means the task already failed one of the first three questions.
Two hard overrides sit on top of the scoring. If you would not disclose the AI’s involvement, the task drops to Level 1 or below regardless of everything else. And if the consequence is high and you cannot practically verify the output, no level above 1 is defensible.
Run it against your own work
Pick a task below, or answer the four questions for something you are working on now. The tool places it on the dial and names the human checkpoint that has to hold.
The delegation check
Place one UX task on the dial. Start from an example, or answer the four questions yourself.
Common UX tasks
1. Consequence
If this is wrong and nobody catches it, who pays?
2. Evidence
What is the output standing on?
3. Verifiability
Can a human check it in reasonable time?
4. Visibility
Happy to tell the client and the user how this was made?
Answer the four questions
Each answer moves the task along the dial, from fully human to supervised autopilot.
Do this
Not this
0 of 4 answered
Where common UX work tends to land
The tool will disagree with this table sometimes, and it should — context moves things. As a starting position:
| UX activity | Usual level | What keeps it there |
|---|---|---|
| Moderating sessions with real users | Level 0 | A person is in the room; rapport and follow-up questions are the method. |
| Final recommendation to a client | Level 0–1 | Accountability cannot be delegated. |
| Research conclusions and framing | Level 2 | Summaries drop the contradictions that matter most. |
| Concept and layout exploration | Level 2 | Useful for breadth, dangerous as a finished answer. |
| Synthesising notes you took yourself | Level 3 | Real source material exists and can be checked against. |
| UI copy, error and empty states | Level 3 | Fast to draft, fast to verify, easy to revise later. |
| Accessibility checks | Level 2–3 | Catches mechanical failures, misses experiential ones. |
| Design system documentation | Level 4 | Low stakes, verifiable against the code. |
| Alt text first pass, formatting, conversions | Level 4 | Reversible and trivially checked. |
Four practices that make the framework hold
A dial on a slide changes nothing. These four habits turn it into practice.
Keep an AI log in every project. One line per use: what task, what tool, what level, who signed off. It takes seconds and it answers the only question that matters when something goes wrong six months later.
Label synthetic material loudly and permanently. If a persona, quote, journey or data point was generated rather than observed, mark it in the file itself — not in a footnote in a deck that gets reused without the footnote. Assume every artefact outlives the conversation that produced it.
Write the data rule down before you need it. Which tools are approved, what may never be pasted anywhere, how participant data is handled, what the contract actually permits. Most leakage is not malice; it is a designer at 11pm with no rule to follow.
Disclose by default, in plain language. “We used AI to produce first drafts of the copy and to summarise our own session notes; all findings come from the eight interviews listed in the appendix.” Clients respond well to this. What they object to is finding out later.
What does not change
Everything above is about delegation. It is worth being clear about what is not on the dial at all.
Deciding what problem is worth solving. Sitting with a user long enough to notice what they did not say. Choosing which of twenty generated options is right for this client, this market, this budget. Owning the recommendation when it turns out to be wrong. None of this is a drafting task, and no amount of model capability converts it into one.
The teams getting real value from AI in UX are not the ones using it most. They are the ones who decided, in advance, where it stops.
How we work at Adi Design. Every engagement starts with the delegation levels agreed in writing — which parts of the work are AI-assisted, at what level, and who signs off. If you would like that applied to a project you are planning, get in touch.
