Maya Codex
“Add AI” was not a useful product strategy.
Autodesk and Microsoft wanted to understand where natural-language AI could genuinely help people learn Maya, and where early language-model behavior would break trust.
I led the Autodesk-side program across interviews, diary studies, co-design, synthesis, and strategy. The work focused on real task attempts, not opinions about hypothetical AI.
Two research rounds, then an explicit strategy phase.
Round one aligned stakeholders and mapped where the learning journey broke across advanced users, novices, and an educator. Round two narrowed to novice behavior and added diary studies over multiple days.
The strategy phase synthesized the evidence, ran opportunity and co-design workshops, reduced more than 20 concepts into five opportunity areas, and recommended one direction.
Trust was shaped by what happened when the model was wrong.
The most valuable moments were not successful commands. They were the points where the AI misunderstood the user and the user had to decide whether to rephrase, retreat to familiar menus, or stop asking.
That led to a durable design principle: failure should be legible. An assistant that fails opaquely trains users to stop using it; one that explains limits and offers a reliable next step can keep the user in the learning loop.
Distinguish between guidance and execution.
The research compared assistance postures: recommending trusted learning material, guiding a user step by step, or executing commands directly. Those modes earn trust differently depending on expertise and task risk.
The final work defined personas, opportunity areas, and a recommended direction that leadership at Autodesk and Microsoft could use beyond the study.
Research became product direction.
The program produced five opportunity areas from more than 20 concepts and one recommended direction. The findings informed Autodesk AI strategy and Microsoft’s understanding of LLM behavior with non-programmer users.
Autodesk publicly announced Maya Assist with Microsoft the following year. This case does not claim a direct one-to-one feature lineage; the evidence is that the program defined product questions, users, failure modes, and opportunity space before generative AI assistants became mainstream.
Working on a hard product problem?
I’m exploring GTM Strategy, AI Product, Product Strategy, and Forward Deployed roles.