Nothing cuts through this year’s AI hype about how the technology is shiny and new more effectively than practical demonstrations of how manufacturing has deployed the technology for over a decade, to the point of trusting crucial business entirely to unsupervised artificial intelligence.
Xometry is not a manufacturing company. It serves as a clearinghouse for a network of more than 10,000 manufacturing partners worldwide, facilitating parts-on-demand deals.
When a buyer submits part designs to Xometry, AI analyzes the CAD files and provides a quote. If the buyer accepts the quote, the AI reaches out to Xometry’s network to find a manufacturer willing to produce the part. Xometry profits on the difference between what it charges the buyer and offers the manufacturer.
AI generates the overwhelming majority of quotes with zero human supervision or interference. Trusting your entire profit model to a machine sounds crazy, but for Xometry, it works. In Q2 2021, gross margins sat at 23.5%. In Q2 2024, Xometry reported gross margins of 33.4%.
Company growth depends on feeding larger and more varied amounts of data to the AI without losing accuracy. When the company began, it dealt mostly in machining and 3D printing. In 2024, Xometry’s AI tackles quotes for die casting, injection molding, tube bending and cutting, extrusions and sheet metal cutting. Xometry also operates in Europe and Asia, which throws locality specific data into the mix.
Xometry can pull this off because it focused for over a decade on developing AI. Even so, the company has now partnered with Google Vertex AI because even Xometry needs help learning new, better techniques to keep up with the constant flow of fresh data.
For smaller or medium sized manufacturers that deal with high-mix/low yield, automatic assessment of job opportunities for feasibility and profitability holds tremendous promise. Not having to depend on vendors or other third parties to use AI could change the game for those businesses. But could they marshal the resources to deploy the technology effectively?
We spoke with Xometry CEO Randy Altschuler to discuss their technology and what manufacturers can learn from the company’s experience. (This interview is edited for length and clarity.)
IndustryWeek: Has AI been at the core of Xometry’s operations since the very beginning?
Randy Altschuler: We’ve used AI since the very beginning. … We have a number of patents, a whole IP around this, a portfolio of proprietary technology.
IndustryWeek: AI requires training data to build its core algorithm. How long did it take you to get enough data to really tweak the algorithm to the point that you could trust it the way you do today?
Randy Altschuler: We had what’s called a cold start. That was very ugly. In the very beginning we ran some of our own jobs, we just collected manually a bunch of data from suppliers. But manufacturers already have lots of data on jobs they’ve accepted. They have the prices, the costs, the margins.
They probably still have the actual geometries of those parts, maybe the CAD files or the drawings. The bigger challenge will be taking that data and structuring it and finding algorithms. But I think they probably have a surprising amount of the information they need.
IndustryWeek: How does a manufacturer decide whether or not it’s worth investing in the IT resources necessarily to even attempt using this sort of technology?
Randy Altschuler: It’s a great question, how do you figure out the ROI on this? I would take something that you’re already doing today, build a team around it, figure out how to collect the data and structure the data and build the algorithms. You just start small. Do a test and see how that fares.
What’s the ROI on that? I’ve got two people devoted to this, they’re costing me X dollars, I’ve generated this amount of incremental profit, maybe it’s because I’m getting more customers, maybe it’s because I’m getting better, higher margins on these with the prices I selected. And then you can figure out how much additional work those two programmers can do, what the bill is for where I’m storing the data.
IndustryWeek: Does your AI look at part data submitted for a quote and also determine what the best manufacturing modality is for that part if there are multiple options?
Randy Altschuler: It’s even one step further. It gives the customer the option. It makes available to them all the different ways we can make it. But rather than saying here’s what you can do, it’s more like we give you every option that we can and exclude what we can’t.
IndustryWeek: How much did you have to retrain your AI and dial in the accuracy every time you added a new manufacturing method to the options?
Randy Altschuler: There absolutely is a curve. The trough of the curve is now higher going out of the gate. We’ve gotten smarter and better about how to do it. That also applies, by the way, to geography. Even though we’re using effectively the same algorithms, training up on the new data, things may be more or less expensive [in Europe or Asia].
IndustryWeek: So what can a manufacturer learn from Xometry’s experience in terms of what they need to know to adopt AI?
Randy Altschuler: Capturing the data is critical. It’s not necessarily been top of the mind at companies historically, particularly manufacturers, but make sure you’re capturing your data even if you don’t know yet what you’re going to do with it.
IndustryWeek: But how do you know which data you’ll need to train AI at the very beginning?
Randy Altschuler: Some of these processes, as we’ve been launching them, we’ve been doing them manually before, we hadn’t been auto quoting it. We’ve created prices for customers. They said yes, they said no, we’ve seen what the profitability is. We’ve seen what suppliers take and then we’re using that data to train up an auto quote AI model.
IndustryWeek: Manually curating data, to get started.
Randy Altschuler: Exactly. And if you’re a manufacturer, you have that manual data as long as you’ve collected it. So you almost have an advantage as to when we started the company, where we had no data. … Manufacturers will already have that data available. It’s just a question of accessing it.
IndustryWeek: How does a manufacturer prepare to scale more complicated data for more complicated AI? Do they have to work with third parties? Can they possibly do it in-house?
Randy Altschuler: Start simple, take a simpler process that you’re doing where you feel confident about what data is already clean, what’s representative, and start there. If you’re getting success, then you can invest. Maybe it’s using external, maybe it’s hiring people internally, but start small. Prove it out. Show that you can get it done or what the ROI is on that. Then you can expand.
IndustryWeek: So let’s say a manufacturer starts using AI for simple tasks, they build in-house expertise in machine learning, and now they want to use AI for something more complicated. How much do the learnings on those comparatively simple tasks apply to more complex AI? Are you looking at entirely new training, big adjustments, or does the existing expertise cover the more complicated tasks?
Randy Altschuler: I think the technology, the techniques and knowledge are very extensible across different industries but you’re going to need an SME to help the AI programmer. If you’re trying to use AI to help people decide what to eat for dinner versus using AI to help you make manufacturing decisions, I think it’ll still be the same AI programmer but the SME might be someone who serves school lunches versus somebody who’s been in a machine shop for 30 years.
