AI for Engineering Knowledge Management
How to evaluate an AI tool for Autodesk Vault in 2026: the file and item split, category and lifecycle rules, property mapping gaps and duplicate parts.
·
⏱
7 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
Vault is not a generic document store, so a generic evaluation will not tell you much. Four things decide whether a tool works on your deployment. It has to read both the file structure and the item bill of materials, and say which one it answered from. It has to treat categories, lifecycle states and revision schemes as rules that constrain an answer rather than labels attached to it. It has to keep working when property values are missing, which means an index built on geometry and not only on metadata. And it has to surface the duplicates that Copy Design leaves behind, on the files you have today. Run a two week pilot on your own vault, point it at the messiest corner you have, and ask every answer where it came from.
Autodesk Vault does its job. Files are versioned, references hold, and the history of a release is there if you know where to look. What Vault does not do is answer a question that crosses a file, an item, a revision scheme and a change order from two years ago in a single pass. That is the work an AI layer is supposed to take on, and it is also where a tool that looks fine in a generic demo fails on a real Vault deployment. The reason is specific: Vault is not a file store with metadata bolted on. It has its own object model, and a tool that cannot read that model will give confident answers built on half your data. Here is what to test before you commit.
Start With What Vault Already Does Well
An honest evaluation starts with the baseline you already own. Vault manages the files needed to produce a product, tracks every version of them, and holds the relationship information between an assembly, its parts and its drawings. Where used and uses queries follow those relationships in both directions. Property search finds a file whenever somebody filled the property in correctly. For controlled release of CAD data that baseline is solid, and a tool that only duplicates it adds nothing.
The gap opens on questions that cross objects. An engineer rarely wants a filename. They want to know whether a bracket like this one has been built before, what it cost, which change released it, and whether the material callout survived the last revision. Answering that means reading the file structure, the item that represents the part on the shop floor, the lifecycle state of both, and the change history, then applying engineering judgment to what comes back. Vault holds all of that data. It does not reason across it.
Two things are worth reading before you shortlist anything. The first is why property based search leaves engineers stuck even in a well run vault, which we covered in why engineers cannot find what they need in Vault. The second is what Autodesk's own assistant capability covers and where it stops, which we went through in the Vault 2027 AI assistant review. Between them they set the floor. Everything below is about what has to be true of a tool that sits above it.
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
The File and Item Split Decides What a Tool Can Answer
This is the first Vault specific test, and it is the one tools fail quietly. Vault Professional separates the files that describe a design from the items that represent what a company manufactures, assembles and sells. An item is identified by its item or part number, and the Item Master is the complete list of items in the vault. When a file is assigned an item, Vault generates a bill of materials that tracks the relationships between associated items, using the file relationship information it already holds.
So the same product is described twice, in two structures that are related but not identical. A tool that indexes only files will answer questions about CAD and miss the released manufacturing view. A tool that reads only items will miss everything still in design. Neither behavior is wrong, and both sound equally confident in a demo.
Three questions settle it:
Does it read both the file structure and the item bill of materials, and does the answer state which one it came from?
Does it understand that a released item can be associated with a file that has since moved on, and does it flag that mismatch instead of averaging over it?
What does it do on a site that runs Vault without items at all, where the file structure is the only structure there is?
The third case is common and it is not a defect in how those teams work. It does mean the tool has to degrade gracefully rather than assume an Item Master is populated. The same split shapes what happens downstream, where a design structure becomes a manufacturing one and every unmapped attribute turns into somebody's phone call.
Categories, Lifecycle States and Revision Schemes Are Rules, Not Labels
Vault categories are not decoration. A category groups objects logically and assigns a defined set of behaviors and rules to them, which can include user defined properties, a lifecycle definition and a revision scheme. A lifecycle definition assigns security, behaviors and properties according to where an object sits in the design process, using states such as Work in Progress, For Review and Released. Revision schemes assign the sequence values and formats applied to files and items that have been revised, and they reach an object through its category.
The practical consequence is that the state of an object carries meaning an answer has to respect. Quoting a dimension from a Work in Progress version as though it were released is not a small error. It is the kind that reaches a supplier.
Three checks belong in the shortlist:
Does it filter on lifecycle state before it answers, and does it label the state in the response so the reader can see what they are looking at?
Does it interpret your revision scheme, or does it compare revision labels as plain text and decide that revision 10 comes before revision 9?
Does it inherit access rules instead of building a second permission model beside the vault, so an answer never surfaces something the user could not open directly?
The third point is where pilots stall in security review, so ask for the posture in writing at the start rather than after. Leo AI is SOC-2 certified and GDPR compliant, no customer data is used to train models, and your intellectual property stays yours.
Property Mapping Means Your Metadata Lives in Two Places
Vault properties can be pushed to and pulled from the CAD files themselves. A user defined property created in Vault can be set in the client and pushed into an Inventor iProperty, and a value can be pulled the other way when the file is the authority. It is a sensible design with a predictable consequence: the same fact lives in two places, and across a few years of migrations, bulk edits and rushed releases the two drift apart. Material is filled in on some parts and blank on others. Finish becomes a free text field with nine spellings of the same treatment.
A tool whose only index is properties inherits every one of those gaps and then reports them as absence. Ask it for stainless brackets and it returns the subset where somebody typed the material in. This is the main reason a search pilot looks convincing on a curated folder and disappoints on the production vault.
The way past it is an index that does not depend on metadata being complete, because geometry is already in the files. Leo AI reads shape, so a part can be found by what it looks like rather than by what somebody remembered to type, and each answer carries the internal source with it so an engineer can verify rather than trust. We went through how shape based retrieval behaves in geometric search for CAD. Leo offers integrations with leading PDM and PLM platforms, including SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, Arena PLM and others, and it works as an intelligence layer on top of the system you already run rather than a replacement for it.
Copy Design, Reuse and the Duplicates It Leaves Behind
Vault gives you a controlled way to start a new design from an old one. In the Copy Design dialog you choose, file by file, whether to copy, reuse, reuse a branch, replace, exclude or detach, and files marked for reuse keep their names and original properties instead of being renumbered. Used deliberately, that is how a team reuses library content and hardware across products.
Used in a hurry, it is also how a vault fills up. Copy where reuse would have done and you get a second part, geometrically identical to one that already exists, carrying a new number, its own drawing and eventually its own supplier record. Nobody raises it, because nothing is broken. The cost lands in procurement, in inventory, and in qualification work repeated on a part that was already qualified. We put numbers on that pattern in the real cost of duplicate parts.
Three questions for the shortlist:
Can it find near identical geometry across the whole vault, independent of item number, filename and category?
Before any consolidation, can it show where each candidate is used, so the decision is made with the impact visible?
Does it work on the files as they are today, with no remodelling or retagging project first?
That last question decides whether a rationalisation effort ever finishes. A tool that needs clean data before it produces value is asking you to solve the problem you were buying it to solve. For the wider framework that is not specific to Vault, the criteria in how to evaluate an AI tool for PDM still apply.
FAQ
Put your Vault data to work
See what an AI layer answers across your files, items and history
Leo connects to your PDM and PLM, reads the structures your team already maintains, and answers with the internal source attached so engineers can verify what they are given.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
