Practice 9 min read

Putting AI and agentic tools into a corporate UX practice

Most enterprise teams have bought AI tools and seen no change in how design actually happens. The gain isn't generating screens faster — it's compressing the slow, unglamorous parts of the process, and knowing which parts must stay human.

The licences get bought first and the thinking second. A team gets access to a research-synthesis tool, a prototyping model, an agent framework — and six months later the process looks identical, except a few people privately paste transcripts into a chat window and hope nobody in legal notices.

The failure isn't the tooling. It's that nobody redesigned the process the tooling was meant to change, and nobody decided where human judgement is non-negotiable. Below is how I'd approach it — stage by stage, with what the machine does and what the designer keeps.

Stage by stage: what AI takes, what stays human

Discover
Synthesis in hours

Interview transcripts, support tickets and survey free-text clustered into themes with quotes attached — the analyst reviews and corrects rather than starting from a blank affinity board.

Human keeps: deciding which theme matters
Define
Evidence-linked artefacts

Personas, journey maps and JTBD statements drafted from the research corpus, each claim traceable to a source — so nobody argues from a persona invented in a workshop.

Human keeps: framing the problem
Design
Prototypes, not mockups

Working, clickable flows generated in a day and tested with real users the same week — plus variants for the arguments a team would otherwise spend a month debating in the abstract.

Human keeps: the design judgement
Validate
Continuous audit agents

Agents that run accessibility and design-system conformance on every build, file issues with severity and a suggested fix, and flag drift from the pattern library before release — not after.

Human keeps: what ships and what waits
Operate
The system maintains itself

Component documentation, spec-to-code handoff and release notes generated from the design system itself, so the library stops rotting the moment the designer who built it moves on.

Human keeps: governance and taste

The pattern across all five: AI drafts, humans decide. Every stage where that inverts is a stage where quality quietly leaves the building.

Where corporate AI adoption actually fails

Tools without process

Licences get bought, nobody changes a ritual, and the tool becomes a faster way to produce the same artefacts nobody reads. The process has to be redesigned, not decorated.

No governance for confidential data

In regulated and IP-heavy organisations, the first serious question is what may leave the building. Answer it up front or the practice stalls in legal review — usually after a team has already built a workflow that depends on the answer being yes.

Skipping the user

AI can produce plausible research. Plausible is not true. Generated insight is a hypothesis to test, never a substitute for talking to someone — and in expert domains the model has no access to the tacit knowledge that makes the difference.

Designers left to figure it out alone

Adoption is a capability problem, not a licensing one. Without shared prompts, review norms and a legitimate way to say "this output is wrong", the team splits into power users and sceptics — and the sceptics are often the strongest designers.

What adoption needs to work

A process map before any tool

Find where the weeks actually go. Only then does it become clear which parts AI should touch — and which are already fine. Most teams discover the bottleneck is decision latency, not production speed, and no model fixes that.

A data boundary agreed with legal and security

What may leave the building, settled up front. In IP-heavy organisations this is the difference between a practice and a stalled pilot — and the answer is often "more than you assumed, under these conditions".

One real project, not a sandbox

Start where the payback is fastest — research synthesis and prototyping — on work that ships. The result is then evidence rather than a demo, and evidence is what unlocks the next round of adoption.

Shared norms, not individual habits

Agreed prompts, review expectations and an escalation path for bad output. Codify it, or the practice dies the moment its champion changes jobs.

Where this comes from

I've designed AI features for expert users who distrust automation, and built the trust patterns that make ML output usable at all — patent pending on coverage-pattern visualisation, and published work on VR/AR data analysis and Scrum with UX. The same scepticism I designed for is how I judge AI inside a design team.

Working through this in your organisation?

aazizhana@gmail.com · LinkedIn

Get in touch