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
Where corporate AI adoption actually fails
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.
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.
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.
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
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.
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".
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.
Agreed prompts, review expectations and an escalation path for bad output. Codify it, or the practice dies the moment its champion changes jobs.
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.
aazizhana@gmail.com · LinkedIn