Estimators are already testing tools like ChatGPT, Gemini, and Claude to see if they can speed up quoting. And in some ways, they can. But it’s only natural that a highly fluent chatbot starts to break down when handling highly specific, highly nuanced tasks. Watch the video to see how a chat bot handled a real sheet metal assembly quote, and why AI for estimating needs to be built around the realities of manufacturing.
Video Transcript
Quoting for custom part manufacturers is a real challenge. You have to make sense of 3D CAD models and engineering drawings, figure out what a part will cost to make, decide how you should price it, and do all of that quickly.
Naturally, estimators are asking: how can AI help?
In response, we’re seeing a lot of content being posted making bold claims about AI. Here at Paperless Parts, we’d like to walk through what’s actually possible today and the approach we’ve taken to create manufacturing-aware AI that makes quoting easier without sacrificing speed or accuracy.
I’m Scott Sawyer, Co-Founder and Chief Scientist here at Paperless Parts. My job is to bring AI capabilities into our product to help you automate the annoying tasks you’re eager to stop doing.
What We’ll Cover
First, we’ll look at one possible way to apply state-of-the-art AI technology to estimation. Then we’ll talk about our approach at Paperless Parts. Finally, we’ll dive into some of the things we’ve built and how we’ve gotten as far as we’ve gotten.
Can You Use ChatGPT to Quote Parts?
To get started, let’s talk about one way we’re seeing folks use top-of-the-line AI models to quote parts, and that’s “vibe quoting.”
In software engineering, there’s a thing called vibe coding, where you prompt AI models with a description of what you’re trying to build and have it generate the code for you. You check whether the software seems to do what you want, but you never really bother to review what code it wrote or how it did it.
Vibe coding can be a great way to build a quick demo. But today, it’s not really a viable option for writing robust, secure, high-quality enterprise software.
So let’s suppose that instead of vibe coding, you want to do vibe quoting. How might that work?
Let’s imagine “QuoteGPT.” And we can actually do better than imagine it, because I tried it out for you.
Big disclaimer: don’t do this with your customer’s data unless you really understand your NDAs and any compliance requirements like DFARS or CMMC.
I created a custom GPT inside of ChatGPT. For this example, I uploaded a simple sheet metal assembly. I had a blueprint, and I also took screenshots of the 3D model, since ChatGPT doesn’t natively accept CAD files. Then I wrote a detailed prompt describing exactly how I wanted it to generate a quote for my part.
Then I hit enter, and it started thinking. It actually thought for quite a while. It worked for about 10 minutes on this quote, and then it returned some truly interesting information.
What ChatGPT Got Right
ChatGPT presented a lot of good information about this assembly.
First, it said, “This is not a single part. It’s a two-piece sheet metal enclosure assembly with installed hardware.” That’s correct. It found and understood the BOM table without specific prompting around that, which is fantastic.
Next, it said, “It’s 5052-type sheet suitable for bending. The exact alloy temper is not fully explicit, so this is an estimating assumption.”
That’s not bad. It found AMS 4016 and translated that to aluminum 5052. I think it could have made a stronger statement than this, but it’s not wrong.
Next, it said, “From the detail sheet in your CAD screenshots, this is a U-shaped cover with vent slots, mounting captive screw features, and a few small pierced holes.”
This is arguably the coolest thing it did. That’s a really good description of the part, which is amazing because a lot of those words are not present in the blueprint.
Where ChatGPT Missed Important Details
Then it said there’s one non-90-degree formed feature called out at 36.87 degrees. It said the 36.87-degree formed feature is fine, but it adds setup sensitivity.
That is the first thing it got distinctly wrong.
That is not a non-90-degree formed feature. Although that angle callout is present, this is actually just a diagonal cut on the laser. That’s not quite right, and it could have the wrong impact on the costing for this part.
I then asked it to make a bunch of assumptions about shop rates and markups, and it went ahead and quoted a price for quantity one and quantity 10.
How did it do?
It came back with about $300 for one and about $100 each for quantity 10.
I was tempted to reach out to some of our estimating customers to quote this, but I didn’t want to waste their time. However, I don’t mind wasting my own team’s time.
So I asked around the office and said, “What should this part really cost?”
The answers were all over the place. Some team members were fairly aligned with ChatGPT. Others were thinking more about prototype quantities and the general pain factor of dealing with really small quantities for parts. Others were thinking more at production scale.
How ChatGPT Compared to Human Estimators
Let’s grade ChatGPT against our smart humans.
Both estimators, meaning the human estimator and QuoteGPT, nailed quote setup. Stuff like part number, material, finish, bill of materials, and proposing reasonable routing. Everyone was fine at that.
I didn’t give ChatGPT a real list of work centers or anything like that, but it qualitatively understood what the manufacturing steps would be.
But when you move into non-trivial quantities, other things really start to matter.
For example, let’s talk about laser cut length. A human estimator using Paperless Parts for assistance would have correctly come up with 155.5 inches of cut length. ChatGPT predicted 128 inches, so it was 18% off and an underestimate.
Definitely not perfect.
As far as bends, the human estimator using Paperless Parts correctly gets to 12, which is the correct bend count. ChatGPT sees eight of those based on the data I gave it. That’s a third off.
Finally, Paperless Parts correctly unfolds the part. It’s 18.9 by 9.2 inches as a flat pattern. ChatGPT estimated that at 15.2 by 10.4 inches. That’s off by about 10%.
More importantly, if you ordered sheet metal based on that unfolded size, you would be short.
These things — cut length, bends, unfolded size — start to matter at non-trivial prototype quantities.
What We Learned from “QuoteGPT”
In summary, QuoteGPT in 2026 is not integrated with your other tools and workflows. It doesn’t natively accept 3D models. It may not be compliant with your NDAs, security requirements, or export control requirements. It’s kind of slow. It’s wrong on a lot of details.
In conclusion, it’s about as fun and effective as chatting casually with an estimator for a minute.
I don’t share this to make fun of AI. People love to dunk on AI, but the truth is I use it extensively every day, and I’m way more effective because of it. This is clearly the direction the world is headed in.
But what doesn’t quite work today is the concept of vibe quoting, or expecting these models to just get everything right on the first shot.
Paperless Parts’ Approach to AI
Our team has put a lot of attention into tracking what’s happening in the AI ecosystem and bringing the best innovations into Paperless Parts to help you get your job done better.
We call this approach Wingman. These principles are on our website if you want to read more, but I’ll highlight three of them here.
First, we’re secure and compliant. Paperless Parts handles sensitive data, and we want to make sure you can use our AI features with any data you would upload to the rest of our application. That means our AI features are CMMC compliant and housed within our FedRAMP boundary.
Next, our AI features are transparent and human-centered. We make it very clear what information is coming from AI, and we make it very easy to review.
Finally, our AI features are measured, tested, and always improving.
What I’d really like to do is explain how we achieve this and why it’s different from just uploading parts to ChatGPT.
Generative AI vs. Discriminative AI
To do that, let’s make a clear distinction between two very different applications of AI technology.
On the left, we have generative AI. These are applications that produce new text, images, videos, or audio based on a prompt.
To be clear, I’m not saying generative AI is bad. I use it all the time. I use it to help me improve emails, write code, and organize my thoughts.
On the right, however, we have discriminative AI, which we’ve found to be much more useful inside Paperless Parts.
Here, we’re using some similar underlying technology, but we use it to identify, extract, and interpret information rather than generate novel new content.
Generative AI does have challenges. It can create training data disclosure and copyright issues. It can give confidently wrong answers. It has lots of general knowledge, but it lacks the domain-specific details that someone like an estimator would have.
It can also be expensive. Many of these tools are usage-based, and we’re starting to see AI usage become less subsidized than it was when many tools first launched.
Finally, generative AI is hard to benchmark and measure. If there’s no clear right or wrong answer, how do you know if the models are getting better?
With discriminative AI, we can avoid many of these challenges while still taking advantage of advanced technology.
This lets us not only detect and extract text like traditional optical character recognition, or OCR, but also understand the wider context of what’s happening in these documents.
For example, the model is able to classify things like dimensions, control frames, and view captions without confusing them with other text or numbers. It does this in a flexible way that can adapt to all the variations we see in real industrial prints.
How Paperless Parts Built AI for Quoting
To tell that story, I’d like to walk you through our journey at Paperless Parts over the last three years of bringing AI into our product.
We’ve taken a crawl-walk-run approach, adding capabilities gradually over the last few years.
One of the first things we created was a model with a simple mission: it took a file name, like a .STEP or .PDF file, and its only job was to label each character in that file name as part of the part number, part of the revision, or neither.
In our experience, most buyers put that information in the file name, but there’s no standard for it.
Instead of having a software engineer spend lots of time looking at all the different variations and writing code to parse each common format, this was a great opportunity for machine learning. It was also low risk.
That model could take those file names and automatically infer which parts were relevant for part number and revision.
We started that journey in 2023.
Then, in 2024, we focused on making quote setup easier. That’s the process of going from an RFQ to a new quote created in Paperless Parts with the line items and files assigned.
One of the challenges is that part buyers include information written in plain language in the email body. They might say, “Hi, I need a quote for quantity 10 of part 1234.” This is a perfect application for a large language model.
Now, we can’t just send these emails to public AI models like ChatGPT, and that’s due to security and compliance issues. But Amazon Web Services, which is the government cloud we’re built on, offers LLMs inside its GovCloud FedRAMP environment.
These aren’t always the latest and greatest models. They tend to lag six to 12 months behind state of the art. But LLMs have been good enough to read an RFQ email for a while.
So that’s exactly what we do. We use a slightly older LLM, run it in a FedRAMP environment, and make sure that no data presented to it is used to train a model or stored as a chat record.
This LLM is able to infer part numbers, quantities, and related information, and we’ve gotten some really great feedback on this feature.
From Wingman 1.0 to Wingman 2.0
Last year, we unveiled PDF callout extraction. This year at our POWER user conference, we announced a significant upgrade of that capability.
When we went from Wingman version one in 2025 to Wingman version two in 2026, we actually removed thousands of lines of code doing detailed extractions and handed those off to a larger AI model.
These solutions still have code around them to get data in and out and do different things with these files. But if you look at just the extraction code, we have less code in the new model, much more model capability, and a huge increase in performance on an internal callout benchmark we created to measure our AI capabilities.
AI is advancing at an exponential pace, and that means any model or software solution that’s not getting better fast is actually falling behind.
At Paperless Parts, we’re able to ride this larger wave of AI improvement by using frontier models to design systems, implement code, train and fine-tune models, and integrate that work into an application.
That lets us keep pace with the larger technology shift.
This has translated into real results. In the 11 months between our release of Wingman 1.0 and Wingman 2.0 for document analysis, we improved extraction performance by 14 times.
Final Takeaway
In summary, chatbots and agents are powerful, but they can be wrong and noncompliant.
Paperless Parts AI features are custom-built for estimators, and their capabilities and accuracy will continue to improve as the technology gets better.
If you’d like to learn more, check us out at paperlessparts.com and request a demo today.
