In my last post, I argued that manufacturing faces two massive risks: fewer people who know where critical knowledge resides, and more systems that make that knowledge difficult to access. I called these the Human Limit and the Systems Limit.
This one asks the obvious follow-up question.
Manufacturers have spent billions on CAD, PLM, ERP, MES, QMS, and every other enterprise platform imaginable. If those systems were supposed to capture and connect engineering knowledge, why is it still so hard to access?
The clearest version of this problem shows up in sourcing. Every sourcing leader has experienced this. A new RFQ arrives. The organisation has likely bought something similar before, qualified a supplier before, negotiated a price before, solved manufacturability issues before. Yet the sourcing team behaves as if it is encountering the component for the first time. The result is duplicated qualification effort, longer RFQ cycles, fragmented supplier spends and missed opportunities in negotiations.
The answer, I think, lies in how manufacturing software evolved — and what each generation was built to do.
Each generation solved a different problem
The first generation of manufacturing software asked a simple question: can we create engineering definitions digitally?
CAD answered it. Engineers could design parts with precision, create drawings, generate models. That was transformational.
The second generation asked: can we manage those definitions across an organisation?
PLM answered it. Revisions, change management, lifecycle governance. Also transformational.
The third generation — and this is where it gets interesting — asked: does something similar already exist?
This produced classification systems, geometric search, and similarity clustering. For the first time, manufacturers could find parts that looked alike even when they had completely different part numbers, lived in different systems, and were created by teams who never knew about each other.
Each generation solved a real problem. Each delivered real value. But none of them were built to answer the question that matters most right now: what should the enterprise do with everything it already knows?
The part number was never designed to carry intelligence
Every enterprise system in manufacturing is built around the same unit. The part number.
ERP raises transactions against it. PLM files records under it. Suppliers quote against it. Quality logs findings to it.
The part number is excellent at administration. It was designed for administration. But it was never designed to carry relationships.
A bracket designed in Stuttgart in 2014 and an almost identical bracket designed by a different team in Detroit in 2024 have different numbers, live in different systems, belong to different programmes, and are known to different people. Administratively, they are separate. Practically, they may be the same thing.
I’ve seen this play out enough times to know it isn’t an edge case. Engineers redesign parts that already exist. Sourcing teams qualify the same supplier twice for geometrically identical components at different prices. Tooling gets commissioned for parts that are near-identical to tooling already sitting in a store somewhere.
The part number, for all its administrative value, is a ceiling on intelligence.
Geometric search was a genuine breakthrough
The industry did not ignore this problem. CAD and PLM software recognised that part numbers were hiding relationships and built tools to surface them. Instead of asking what a part was called, they asked what it looked like using geometric similarities. Tools such as OnePart and Geolus have been used by engineers for over a decade. That was the right insight, and it produced real value.
For the first time, manufacturers could see families of geometrically similar parts across product lines and programmes. A cluster of forty-seven similar brackets could become visible. Reuse opportunities emerged. Duplicate designs became easier to catch. Standardisation became possible.
More recently, these tools have evolved further by combining geometric similarity with metadata, purchasing information, classification workflows, and documentation to help organisations standardise catalogues and improve sourcing decisions.
This has created significant value for manufacturers. But it has also revealed something important about where the problem continues to live.
Clusters surface similarity, not enterprise understanding
Here is what I’ve observed in practice.
When a similarity search surfaces a cluster of forty-seven similar brackets, someone — usually an experienced sourcing head or engineer — still must manually work out what to do about it.
The question isn’t simply “Which parts are similar?” A sourcing question typically is: which suppliers have successfully manufactured this family of geometry before, at what cost, with what quality outcome, and for which programs?
Which of those parts can come from existing inventory rather than triggering a new purchase order? Which suppliers have manufactured something in this geometric family before — and how did it go? Does a quality issue on one of those brackets apply to the others? Which design should become the standard, given cost, supplier history, and programme requirements? What happens downstream if one of those designs changes?
None of those answers live in the cluster. They live across ERP, quality systems, supplier records, manufacturing routings, and inventory databases — all organised by part number, not by geometry. A similarity tool tells you what looks alike. It doesn’t tell you what to do because parts look alike.
That gap — between finding similarity and knowing what to do because of it — is the gap that has persisted for over a decade despite widespread availability of search and clustering tools. If geometric similarity alone had been enough, duplicate rates across OEM catalogues wouldn’t still be sitting at 20 to 40%. The fact that they are tells you something important about where the real problem lives.
Why the major vendors haven’t closed it
This is the question I get most often when I describe ShapeSense to people who know the industry. If this gap is real and the value is obvious, why haven’t CAD and PLM giants already solved it?
The honest answer is structural — and it’s not a criticism.
These companies built extraordinary engineering authoring tools and PLM systems. Classification tools and similarity engines that followed were built on top of those foundations, in service of engineering productivity and design governance.
Their natural unit of value is the engineering record — the CAD model, the drawing, the revision, the part number. Their business model rewards more seats, more modules, more integrations within that ecosystem.
That is what they were optimised for. It is genuinely excellent at what it does.
But the problem of making accumulated engineering intelligence accessible — connecting geometry to live inventory positions, to supplier performance history, to quality findings across related parts, to manufacturing capability and cost — is not an engineering authoring problem. It requires starting from a different place entirely. Not from the record, but from the decision.
There’s also a structural incentive question. Geometric clustering and search tools operate within the CAD and PLM environment. They help engineers find things faster within that world. They weren’t designed to reach across into ERP, QMS, MES, and supplier systems. Doing that meaningfully would require deep integration across systems that have historically been treated as separate. And it would require a business model oriented around enterprise outcomes rather than engineering seat counts.
That’s a different company from the ones these vendors were built to be.
What accumulated engineering intelligence actually is
Before I describe what the next generation needs to do, I want to name what’s at stake.
What sourcing teams call supplier knowledge is often engineering knowledge in disguise. Every successful supplier qualification, tooling decision, cost reduction initiative, and quality correction is ultimately linked to a geometry, a process, and a manufacturing outcome. The challenge is that these learnings remain trapped inside individual part numbers and programs instead of becoming reusable enterprise intelligence.
The people making sourcing decisions increasingly depend on engineering intelligence, yet most engineering intelligence remains accessible only through engineering systems. A sourcing team trying to identify capable suppliers, evaluate manufacturability, assess tooling reuse, understand quality history, or find geometrically similar components is often forced to rely on engineers, specialist CAD tools, or PLM systems to retrieve the relevant information.
The issue isn’t that sourcing lacks expertise. The issue is that the underlying knowledge was created for engineering purposes and remains locked inside engineering environments. As organisations grow, this dependency turns accumulated knowledge into an organisational bottleneck, slowing decisions precisely when speed matters most.
Most manufacturers have been accumulating this intelligence for decades. It exists — in drawings, CAD files, quality records, supplier histories, and in the institutional memory of the people who built these organisations.
The problem isn’t that manufacturers lack intelligence. Most have more of it than they can access. The problem is that it gets harder to reach as organisations grow, systems proliferate, and experienced people move on.
The organisations that will pull ahead over the next decade won’t be the ones with the most data. They’ll be the ones that have figured out how to operationalise the intelligence they already have.
What the next generation requires
The foundation is geometry.
Geometry is the one signal that stays consistent across every system, every file format, every drawing standard, and every era of product history. A part’s three-dimensional form doesn’t change when the PLM system gets upgraded. It doesn’t vary between a drawing made in Stuttgart and one made in Pune. It doesn’t disappear when the engineer who designed it retires.
Starting from geometry means starting from the only thing that’s always true about a part.
Geometry reveals families. Families reveal relationships — between parts and inventory, parts and suppliers, parts and quality outcomes, parts and manufacturing capability. Relationships reveal intelligence. And once intelligence is connected across the enterprise rather than locked inside individual systems, it can work in two ways.
On demand — when a sourcing head asks whether existing inventory can satisfy active demand, or whether a particular supplier has proven experience with this type of geometry, the answer is available in seconds rather than weeks.
But on-demand intelligence has a limitation that’s easy to overlook.
It still requires someone to ask.
Most of the intelligence that matters most in manufacturing arrives at moments nobody anticipated. Inventory that quietly became reusable when a programme was cancelled. A quality finding on one part that should have flagged seventeen geometrically related parts in three other product lines — but didn’t, because nobody ran the search. A new part being designed right now that is geometrically identical to something already sitting in the approved supplier base.
These aren’t rare exceptions. They’re the normal operating condition of a complex manufacturing organisation.
Most systems wait. The engineer asks, the system responds. The engineer stops asking, the system goes quiet.
ShapeSense is built to work the other way. Always-on intelligence watches continuously — surfacing reuse opportunities as inventory moves, flagging quality patterns across geometric families as findings come in, catching duplicate designs before they’re commissioned rather than after.
One helps people make decisions faster when they’re already trying to make them. The other makes sure they find out about the decisions they didn’t know they needed to make.
That distinction — between a system that responds and a system that watches — is one of the sharpest differences between search and intelligence.
Where this leaves us
The first generation of manufacturing software helped us design parts. The second helped us manage them. The third helped us find similar ones.
The fourth — the generation the industry has been building toward without quite arriving — must help us understand what the enterprise already knows about those parts and what to do with that knowledge.
Clusters are part of that journey. They always were. But a cluster is a starting point, not a destination. Geometric similarity tells you where intelligence resides. Relationships tell you what that intelligence means. Decisions determine whether it creates value.
Shape reveals relationships. Sense turns relationships into decisions. And decisions are how engineering intelligence endures.