AI for Engineering Knowledge Management
What an AI assistant can and cannot reach inside Autodesk Fusion Manage: items, workspaces, BOMs and change orders, and where the native geometry stops.
·
⏱
8 min read
Michelle Ben-David
Michelle Ben-David is a mechanical engineer and Technion graduate. She served in an IDF elite technology and intelligence unit, where she developed multidisciplinary systems integrating mechanics, electronics, and advanced algorithms. Her engineering background spans robotics, medical devices, and automotive systems.
BOTTOM LINE
Fusion Manage is a strong substrate for an AI assistant precisely because it is opinionated about structure. Items carry a shared attribute set, lifecycle state and revision are fields rather than filename conventions, and BOM, where-used and change relationships are explicit records. That makes answers checkable in a way a file share never allows. The limit is equally clear: what the Vault Connector moves across is items, BOMs, attributes and neutral-format files, so the native model and its design intent stay in the CAD seat and the vault. Pick a tool that knows which side of that line each question sits on, that reads the records and the geometry when a question needs both, and that cites what it read. That combination, rather than the size of the model behind it, is what makes the answers worth acting on.
Autodesk's cloud PLM has been running in production at hardware companies for years, and the question teams bring to it has shifted. It is no longer whether the system can hold a change order. It is whether anything can answer a question across everything the system already holds.
That question has a precise answer, and it depends less on the assistant than on the shape of the data underneath it. Fusion Manage keeps its records in a structure that suits retrieval unusually well. It also stops short of the geometry, in a way that is easy to miss until an assistant answers a question it had no business answering.
What follows is both halves: what an agent genuinely reaches inside Fusion Manage, and where the reach ends.
What Fusion Manage Actually Holds
Fusion Manage organizes everything into workspaces. Autodesk's administration documentation defines a workspace as a virtual space in which to store, view and work with items that share the same attributes, and that definition is worth taking literally. The shared attribute set is the thing that makes the data tractable later.
Items are the building blocks. An item is a unique data record inside a workspace, and Autodesk is deliberately broad about what one can represent: a specific part, an assembly, a person, a task, an organization, an asset, or whatever a given site has been configured to manage. A document control workspace holds items that are documents. A change order workspace holds items that are change orders. The abstraction is the same in both cases.
Workspaces themselves come in six configurations: a basic workspace, a basic workspace with workflow, a revision-controlled workspace, a revisioning workspace, supplier management, and supplier management with workflow. A typical site exposes workspaces along the lines of Items and BOMs, Change Requests, Project Management and Suppliers, and Autodesk ships Components and Drawings workspaces pre-configured for Fusion users.
In a revision-controlled workspace, the item header carries the descriptor, the lifecycle state and the revision. An item currently moving through a change is marked with a stamp that links directly to the change order driving it. Those three facts, state, revision and the link to the change, are stored as structured fields rather than left to a naming convention, and that distinction turns out to matter a great deal.
IN PRACTICE
We've started reusing parts we didn't even know we had, and that has real downstream impact on procurement and BOM costs.
- Verified User, Defense & Space
Why the Workspace Model Reads Well to an Assistant
Most retrieval problems in engineering data are not search problems. They are schema problems. An assistant pointed at a network directory has to infer, from a filename, whether Bracket-1042_RevC_final_v2.sldprt is current, superseded or abandoned, and it will sometimes infer wrongly. Fusion Manage removes that guesswork, because the same facts are fields.
Four properties of the workspace model do most of the work:
Uniformity. Every item in a workspace shares an attribute set by definition, so a query returns predictable columns rather than a ragged union of whatever each record happened to carry.
First-class state. Lifecycle phase and revision are stored fields on revision-controlled items, so "is this the released version" is a lookup rather than an inference.
Explicit relationships. BOM structure, where-used and the link from an item to the change order stamping it are records in their own right, not conventions a reader has to reconstruct.
Site-wide reach. Items can be searched within the current workspace or across the whole site, so a question that spans change requests and suppliers does not require three separate exports first.
The practical effect is that an assistant working over this data can be held to a standard. If it claims a part is at revision D and in a released state, that claim is checkable against a field, and a tool that cites the record it read can be audited by the engineer reading the answer. That is a meaningfully different posture from an assistant summarizing a folder of attachments, and it is the reason the PLM is a better starting point than the file share. For a wider view of how these systems divide, our guide to PLM and PDM for engineers covers the split in more detail.
Where the Geometry Stops
Here is the boundary that decides what any tool can honestly promise. Fusion Manage is a record system. It is not where the native model lives.
Autodesk describes the Vault Connector as automating the synchronization of items and bills of materials, attaching neutral-format files, and matching attributes between Vault Professional and Fusion Manage. Read that carefully, because the word doing the heavy lifting is neutral. What crosses into the PLM is the record, the structure and a representation of the shape. What does not cross is the native parametric model: the feature tree, the sketch constraints, the mates, the design intent that explains why the geometry is the shape it is.
So an assistant reading Fusion Manage alone can tell you that an item sits at revision D, that it is used in four assemblies, that a change order opened against it last Tuesday, and which supplier record is attached to it. The same assistant cannot tell you the wall thickness, cannot tell you whether two brackets in different programs are the same shape with different part numbers, and cannot tell you whether a proposed change violates a draft angle the model has always carried.
That is not a criticism of the PLM. It is what a PLM is for. It does mean that evaluating an AI tool for Fusion Manage is mostly an exercise in working out whether the tool stays inside that boundary and says so, or quietly steps over it and starts guessing. A tool that also reads the vault and the drawings behind the records can answer the geometry questions honestly. A tool that only reads the PLM should decline them. Our write-up on AI across PDM and PLM and the companion piece on an AI tool for Autodesk Vault both come at this split from the file side.
The Questions Worth Asking Across Items, Changes and BOMs
Once the boundary is clear, the useful work is obvious. These are the questions that are expensive to answer by hand and cheap to answer from structured records:
Where is this item used, and which of those assemblies are themselves in a released state?
What changed between revision C and revision D, and which change order authorized it?
Which items are sitting in a given lifecycle phase longer than they should be?
Which change orders are currently open against anything in this product line?
Which supplier records are attached to the items on this BOM, and where does a single supplier carry more of the structure than anyone intended?
Has something close to this part already been through the system under a different number?
That last one is where the value usually concentrates, and it is also the one the PLM cannot settle alone, because deciding that two parts are the same thing is a geometry question wearing a records question's clothes.
This is the gap Leo is built for. Leo is an AI assistant for mechanical engineers, trained on more than a million pages of standards, books and articles, and it works as an intelligence layer on top of the systems a team already runs rather than as a replacement for any of them. Leo offers integrations with leading PDM and PLM platforms, including SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, Arena PLM and others, alongside local and network directories and ERP. Because it reads the records and the files behind them, a reuse question can be answered from the BOM and the geometry at once, with a citation back to whatever it read. Teams who run that search before releasing a new part number tend to find the savings in procurement rather than in engineering hours, which is covered further in our piece on PLM search across parts.
What to Check Before You Commit to a Tool
Most evaluations of an AI tool for a PLM go wrong by testing the demo question rather than the boundary. A short checklist keeps it honest:
Does it read the record model, or only the attachments? A tool that ingests files hanging off items will miss lifecycle state and BOM structure entirely, which are the two things the PLM is best at.
Does it respect revision and lifecycle state in its answers? Ask it something whose correct answer changed at the last release and see whether it notices.
Does it cross the geometry boundary deliberately? Either it also reads the vault and the drawings, or it should say it cannot answer shape questions. Both are acceptable. Silent guessing is not.
Does every answer carry a source you can open? An answer you cannot check is an answer you have to re-derive, which is the work you were trying to avoid.
Does the security posture survive procurement review? Leo is SOC-2 certified and GDPR compliant, no customer data is used to train the models, and customer IP stays protected.
Does it work across systems rather than inside one? The questions that matter rarely stay inside a single platform, as our overview of AI across CAD, PLM and ERP sets out.
Run those six against anything you are considering, including Leo. The tools that answer them cleanly are the ones still useful in month six, after the novelty of the first demo has worn off and the questions being asked have turned into the awkward ones.
FAQ
Put an assistant over your PLM
Ask one question and get an answer drawn from items, BOMs and changes.
Leo connects to the PDM and PLM systems your team already runs, reads the records and the files behind them, and answers with a citation you can open and check.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
