The problem
“Add AI” was not a useful strategy.
Maya is powerful and difficult to learn. The question was not whether natural language sounded exciting. It was where language assistance fit the workflow and what users did when the model was wrong.
Program design
I used multiple methods because attitudes and behavior are not the same evidence.
Discover the learning breakdowns
Interviews across experience levels helped map where users struggled and what experts and novices needed from assistance.
Test the real tool over time
Diary research captured what happened when participants used the Codex-powered plugin outside a single moderated session.
Turn findings into direction
Opportunity and co-design work translated the research into product concepts, priorities, and recommendations for both companies.
Key finding
Failure experience was a product requirement.
When the model misunderstood a user, the important behavior was what happened next: rephrase, return to familiar menus, or stop asking. Opaque failure trained users to abandon the assistant.
Design principle: In professional AI tools, a legible failure can protect trust better than a confident but unclear response.
Expertise also changed the use case. Experts wanted speed. Novices needed scaffolding. One assistant posture could not serve both groups in the same way.
Impact
The research moved from user evidence to product strategy.
I synthesized opportunity areas and a recommended direction for leadership at Autodesk and Microsoft. The study also created a repeatable research-operations pattern for a two-company AI program.
What this demonstrates
AI use-case discovery, customer research, failure-mode thinking, cross-company program leadership, and the ability to translate technical uncertainty into product direction.