Three ideas I defended at AU 2026

Autodesk University 2026 is over. On Thursday at 4:30 PM I presented MFG3731, Vault on Autopilot, a 60 minute deep dive on AI agents operating inside Vault Professional. Now that the stage is behind me, I want to put in writing the three ideas that structured the session, because they are the ones I plan to keep repeating until they stop being necessary.

1. Autonomous does not mean no human

A well built agent runs the full workflow: it understands the request, finds the files, prepares the operation and presents the result. But a person authorizes the write. That line, which looks like a design detail, is the difference between a tool your team adopts and an incident waiting for a date.

In VaultPilot, which we built together with Avant Leap, that translates into dry-run with confirmation: the agent shows exactly what it is about to change before changing it. Nobody finds out afterwards what the system decided on its own. And the whole conversation runs on the user's own credentials, with no admin backdoors: the agent cannot do anything the person using it could not do.

2. Every write is a claim

An API returning success proves nothing. In our development log, 23 operations received a success code that the subsequent read-back disproved. Without verification, those 23 false successes would today be corrupt data wearing the costume of correct data, which is the worst class of error: the one you do not know you have.

The rule I defended in Las Vegas is simple: every write is a claim, and read-back is the proof. An agent that does not verify its own operations is not autonomous; it is optimistic.

3. Build the gate before you build the agent

A script automates the rule; an agent handles the exception. And because the agent operates where the rules run out, it needs more control than a script, not less. That is why the gates (schema validation, hard limits, identity anchoring, a read-only switch, a full audit trail) are not a security add-on: they are the architecture. In our log, 80 writes died at a gate before touching data. Every one of them was an error that never got to exist.

Where to start on Monday

If these ideas ring true but you do not know where to begin, use the formula I shared in the session: frequency times manual minutes times people, divided by effort in days and risk, with risk scored 1 if the operation is reversible and 5 if it is not. Score your three most painful processes and automate the winner. In our examples, property assignment scored 360 and the pre-release audit 126; revision changes, with more risk than volume, barely 16.

And now, back to work

What follows a week like this is not more stage time: it is applying these three ideas to the real Vaults of real teams. If any of the three describes a problem you have today, let's talk.