AI for Parts & BOM Management
How to build a part classification and standardization scheme engineers use: what a usable taxonomy contains, what AI changes, and how to prove it in one quarter.
·
⏱
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
Part classification is not a data project, it is a retrieval and decision project that happens to need data. The scheme you already have is probably close enough; what is missing is population and enforcement. AI changes the economics of population, because a model can read geometry, drawings, documents and usage context to propose property values across a legacy vault that no team would ever classify by hand. It does not change the economics of enforcement, which stays a human decision about which variant survives and what happens to the rest. Teams that treat the output as a proposal to confirm, prove the loop on one commodity family, and measure new part number creation afterwards get value in a quarter. Teams that start by redesigning the top level tree usually deliver a document.
Most engineering organisations already have a part classification scheme. It usually lives in a spreadsheet, it has more levels than anyone can hold in their head, and a large share of the parts released last quarter never touched it. The scheme is not wrong. It is unused, which in practice amounts to the same thing.
That gap is why part classification keeps coming back as an artificial intelligence project. If a model can read the geometry, the drawing and the supplier document, then the taxonomy stops depending on whether a busy engineer fills in a form correctly at release, on a Friday, for a benefit that lands on somebody else six months later.
The interesting question is not whether AI can classify parts. It can. The question is whether the classification it produces is one your engineers, buyers and manufacturing team will actually act on, and whether standardization follows from it or stalls in a report nobody owns. This guide covers what a usable scheme contains, what AI genuinely changes, and how to sequence the work so it pays for itself before it is finished.
Why part classification fails in practice
Classification schemes rarely fail because the taxonomy was badly drawn. They fail for three structural reasons, and all three are worth naming before any tooling decision.
One hierarchy is asked to serve three functions. Design wants form, fit and function. Manufacturing wants process family, setup and machine group. Procurement wants commodity and supplier category. A single tree forces one of those to win, and the other two quietly keep their own spreadsheets.
The cost and the benefit sit in different places. Classifying a part correctly takes an engineer several minutes at release. The payoff goes to whoever searches for it later, or to the buyer consolidating spend next year. Any process with that shape decays.
Nobody owns maintenance. A new commodity arrives, no node exists for it, and the engineer picks the nearest node or a general catch-all. Repeat that for two years and the catch-all is the largest class in the vault.
The visible symptom is duplicate part numbers. Two engineers specify functionally identical hardware, neither can find the other's release because it was never classified, and the organisation now carries two part numbers, two supplier relationships and two sets of inventory. The downstream cost of that is well understood and lands mostly outside engineering, in procurement and inventory carrying cost.
The less visible symptom is redesign. An engineer who cannot retrieve a similar part designs a new one, which is the single most expensive way to answer a question that a search should have answered. That failure mode is covered in more depth in our piece on why engineers spend so much of their week redesigning parts that already exist.
IN PRACTICE
The geometry search has been invaluable, helping me find standard parts instead of designing new ones, saving a huge amount of time and effort. The search system is smart and CAD-aware.
- eytan s., R&D Engineer
What a usable classification scheme actually contains
This problem is older than CAD. In 1970 Hermann Opitz published a classification system for workpieces that became the reference point for group technology: a nine digit code in which the first five digits describe the shape of the part and the remaining four describe dimension, material, raw form and accuracy, with an optional four character extension for the production sequence. The insight has held up for fifty years. Useful classification is not a folder path. It is a set of attributes with defined meanings.
The modern version of that idea is a property dictionary. The common dictionary schema shared by IEC 61360-2 and ISO 13584-42 formalises it: each class carries a list of property definitions, each property has a stated data type and unit, and each part instance carries values against those properties. Three layers, and the discrimination lives in the middle layer rather than in the depth of the tree.
In practical terms, a scheme your team will use has four properties.
A shallow hierarchy. Three or four levels to reach a class such as hex flange screw or radial lip seal. If you need eight levels, you are encoding attributes as folders.
Six to twelve properties per class, with units. Thread size, length, drive type, head style, material, coating, strength grade. These are what searches, comparisons and consolidation decisions actually run on.
Controlled values. A material property whose values are free text is a property you cannot group on.
Separation from the part number. Classification changes as the business changes. Part numbers must not, which is the whole argument for non-significant numbering covered in our comparison of intelligent and non-significant part numbering systems.
If you get this layer right, everything downstream becomes tractable. If you get it wrong, no amount of model quality will rescue it, because the model will be filling in fields that mean different things to different readers.
What AI actually changes about classification
The honest framing is that AI does not invent a better taxonomy. It removes the data entry bottleneck that has always been the reason good taxonomies go unpopulated, and it does so by reading sources a release form never had access to.
There are four such sources, and a system that uses only the first is doing less than it appears.
Geometry. Shape descriptors computed from the solid model let a system group parts by what they are rather than by what someone typed. This is what makes it possible to find a near identical bracket released under a meaningless file name.
The drawing. Material and finish callouts, tolerance classes, thread specifications and surface finish symbols carry most of the properties a fastener or seal class needs, and they are already recorded.
Released documents. Supplier datasheets, specifications and qualification records supply the properties that geometry cannot, such as strength grade, temperature range or approval status.
Usage context. Which assemblies a part appears in, and which bills of material, tells you how it is used, which is often a better guide to its functional class than its shape.
The workflow change matters more than the model. Classification stops being a form an engineer fills in and becomes a proposal an engineer confirms or corrects. Review is faster than entry by a wide margin, and correction produces training signal that entry does not.
The larger prize is the back catalogue. No team is going to hand classify forty thousand legacy parts, which is exactly why the legacy vault stays unsearchable and why new duplicates keep being created against it. Running classification retrospectively over everything already released is the one job in this area that is impossible manually and routine for a model, and it is usually where the return shows up first.
Standardization is a decision problem, not a detection problem
Here is where most part classification programmes stop short. Producing a list of near duplicates feels like the deliverable. It is not. A list of eight similar shoulder screws is a question, and somebody still has to answer it.
Answering it means survivor selection, and the criteria are rarely engineering criteria alone.
Supply position. Which variants are currently sourced, dual sourced, or single sourced from a supplier under review.
Qualification status. In regulated work, a variant already qualified on a released product is worth more than a technically superior one that is not.
Volume and price break. Consolidating onto the variant with the higher annual volume is usually where the saving is, not onto the cheapest unit price.
Tooling and inventory. Existing tooling, fixtures and stock on hand all argue for a particular survivor.
Then there is retirement, which is the part that actually changes behaviour. A survivor decision with no enforcement produces a cleaner spreadsheet and an unchanged vault. Enforcement usually means three states rather than two: preferred for new design, permitted on existing bills of material, and blocked. Phasing the blocked variants at the next natural revision avoids forcing a change order that nobody budgeted for.
If you are evaluating tools for this specific step, the criteria differ from search and are worth treating separately. We covered them in our guide to judging what actually cuts part count.
How to run this in a quarter rather than a year
The failure pattern for classification programmes is scope. A team sets out to classify the whole vault, spends two quarters arguing about the top level tree, and delivers nothing an engineer can use. The alternative is to prove the loop on one commodity family and then repeat it.
Pick one family with obvious pain. Fasteners, seals, bearings or sheet metal brackets. Pick the one your buyers complain about.
Define six to twelve properties for that family only, with units and controlled values. Timebox this to a week and accept that it will be revised.
Classify the full back catalogue for that family, not just new releases. The legacy data is where the duplicates are.
Review a random sample of one hundred parts by hand and record the agreement rate. This number is your evidence, and it is the only honest way to decide whether to widen scope.
Publish the survivor list, mark the rest blocked for new design, and measure new part number creation in that family for the following quarter.
This is where an intelligence layer over your existing systems earns its place. Leo is an AI assistant for mechanical engineers, trained on more than a million pages of standards, engineering books and technical articles, and connected to an organisation's own knowledge base: its PDM and PLM systems, local and network directories, and ERP. Leo offers integrations with leading PDM and PLM platforms (SolidWorks PDM, Autodesk Vault, PTC Windchill, Siemens Teamcenter, Arena PLM, and others). It sits on top of those systems rather than replacing them, which is what makes a classification effort survivable: the taxonomy is applied against the vault you already have, and the answers cite the source document so a reviewer can verify a property value rather than trust it.
The security position matters for this work, because classification touches every part record you own. Leo is SOC-2 certified and GDPR compliant, no AI is trained on customer data, and your intellectual property stays protected. If you are still choosing an approach at the platform level, our framework for evaluating AI tools for PDM covers the same ground one layer up.
FAQ
Opitz, H. A Classification System to Describe Workpieces. Pergamon Press, Oxford, 1970. The originating group technology classification and coding system, comprising a five digit form code, a four digit supplementary code covering dimension, material, raw material form and accuracy, and an optional four character secondary code for the production sequence.
IEC 61360-1:2017, Standard data element types with associated classification scheme, Part 1: Definitions, principles and methods. International Electrotechnical Commission. Defines the principles for property definitions, data types and associated classification used in product property dictionaries.
ISO 13584-42 and IEC 61360-2, common dictionary schema for parts libraries. International Organization for Standardization and International Electrotechnical Commission. Specifies the shared data model under which a class carries property definitions and part instances carry values against them.
Find the part before you draw it
See how Leo classifies and retrieves parts across your existing vault
Leo reads your geometry, drawings and documents to propose classification across parts you already own, so engineers retrieve a qualified part instead of designing a new one.
Schedule a Demo →
#1 New AI Software Globally - G2 2026
Enterprise-grade security
Trusted by world-class engineering teams
