One agent, three surfaces: Vault, AutoCAD and Inventor

We built the agent inside Vault. And the first user reaction was not about the AI, or the tools, or the security. It was this: "I barely open Vault, I live in AutoCAD".

The comment that changed the design

That comment changed the course of VaultPilot, the AI agent we developed together with Avant Leap for Autodesk Vault Professional 2026. We had built a capable agent: 28 tools across 9 categories, a dry-run before every write, validation against the real Vault schema, a full audit trail. But we had put it in a single place: the Vault client.

The reality of an engineering office is that almost nobody lives there. The PDM administrator does. But the drafter spends the day in AutoCAD, and the mechanical designer spends the day in Inventor. For them, opening another application to talk to the agent was enough friction not to use it.

The three surfaces

Today the same agent lives in three places, each one designed for the person who actually works there.

A chat panel in Vault Client. A WPF panel embedded in the Vault client, for the PDM administrator and whoever manages lifecycles, properties and releases. It is the most complete surface: bulk operations, reports and audits run from there.

The VPSYNC command in AutoCAD 2026. For whoever spends the day on drawings, the agent is invoked from the command line they already know. No window switching, no extra application.

An Inventor 2026 add-in with iProperties sync. The mechanical designer works on their models and the agent keeps properties synchronized with the Vault, from inside Inventor.

One logic, not three products

The architectural point that matters is that these are not three agents: it is one. The three surfaces share the same logic, the same 28 tools and the same 9 safety layers. The dry-run with confirmation exists identically in all three. And on all three surfaces the agent runs with the user's own credentials: whatever you cannot do in the Vault, the agent cannot do for you either. No admin backdoors, no service accounts with superpowers.

That means the behavior is predictable no matter where you come from: the same question produces the same answer, with the same validations, in Vault Client, in AutoCAD or in Inventor.

What comes next

VaultPilot is on its way to the Autodesk App Store, and publishing a commercial add-in is a project in itself: packaging, review, documentation. But the three surfaces decision is already made and there is no going back, because it is what turned the agent into part of the real workflow instead of a demo you show once.

The adoption lesson

The lesson I take with me goes beyond VaultPilot: adoption of a tool does not depend only on what it does, but on where it lives. An excellent tool in a window nobody opens is a tool that does not exist.

So when I design a solution, the question is no longer just "what should it do", but "in which window does the person who will use it spend their day". Software has to go where people are. Not the other way around.