What I brought back from AU 2026

Back from Las Vegas. Autodesk University 2026 is behind me and, as tends to happen with good conferences, I brought back more than I carried in. On Thursday the 17th at 4:30 PM I presented MFG3731, Vault on Autopilot: a 60 minute technical deep dive on AI agents operating inside Vault Professional, in the Design and Manufacturing track.
What I went to say
The session revolved around three ideas I have been defending for months, and the AU stage forced me to distill them. Autonomous does not mean no human: the agent runs the full workflow, but a person authorizes the write. Every write is a claim: an API success code proves nothing, read-back does. And build the gate before you build the agent: control is not a security add-on, it is the architecture. All three come from real work on VaultPilot, which we built together with Avant Leap, and from a development log that does not forgive optimism.
I did not bring promises to the stage: I brought a log. Two months of development against a Vault Professional 2026 instance, with 45 distinct operations invoked from natural language, 249 calls that reached Vault, 150 verified by read-back and 80 writes stopped at a gate before touching data. It is a development log, not a production success rate, and I said so in the room: honest numbers persuade better than round ones.
What I brought back
The best of Autodesk University is never only in the rooms: it is in the hallway. Conversations with people who administer real Vaults, with real pains: teams losing hours every week to repetitive PDM operations, pre-release audits done by hand, and the question that came up most: how to bring AI into this without putting the data at risk. None of those conversations asked for magic. All of them asked for judgment. To me, that is the clearest sign the market has matured: nobody asks anymore whether AI can click; they ask who answers when the click goes wrong.
Thank you
Thank you to Autodesk University for the stage and the trust. Thank you to Avant Leap, partners in VaultPilot and in the thesis that AI inside the PDM is built with gates or it is not built at all. And thank you to every person who came to the session with questions: the real deep dive was that exchange, not the slides.
What comes next
What follows a week like this is not more stage time: it is work. Applying these ideas to the real Vaults of real teams, with their permissions, their lifecycles and their exceptions. If anything I presented describes a problem you have today, the conversation did not end in Las Vegas: it continues on whichever channel you prefer.
Did you miss the session?
The conversation did not end in Las Vegas: tell me your PDM case and let's talk about what to automate and with which controls.