Can AI Quote Machine Shop Jobs? What a Vendor's ChatGPT Test Got Wrong

Not reliably. In a vendor's demo, a custom GPT quoting a sheet-metal enclosure got the quote set-up right but came in 18 percent short on laser cut length and counted 8 bends on a part with 12. A general chatbot can draft the covering text, but it cannot measure a part. Test any tool on your own past jobs first.

Treat that as an anecdote, not a measurement. The test comes from Paperless Parts, which sells quoting software, and it is a single informal run in a marketing video: no ChatGPT model version, no published prompt, no date on the page and no baseline price. It still shows the real risk, because a wrong number that looks normal gets sent. Applied to an illustrative 25-part bracket at illustrative rates, misses of that size price the job about $79 too low.

Follow me on Instagram@sagnikteaches

What the vendor's ChatGPT test found

The test is a demo page on the Paperless Parts site, "LLMs vs. Manufacturing-Aware AI", the transcript of a video presented by Scott Sawyer, the company's co-founder and chief scientist. He built a custom GPT inside ChatGPT, called it "QuoteGPT" and described the approach as "vibe quoting". He gave it the blueprint, screenshots of the 3D model (ChatGPT does not natively accept CAD files) and a detailed prompt on how he wanted the quote built. He gave it no real list of work centres and asked it to assume shop rates and markups. It worked on the quote for about 10 minutes.

Connect on LinkedInSagnik Bhattacharya

The part was a two-piece sheet-metal enclosure assembly with installed hardware: a U-shaped cover with vent slots, captive screw features and a few small pierced holes. The custom GPT read the AMS 4016 material callout as 5052 aluminium and noted that the temper was an estimating assumption because the drawing did not make it fully explicit. That is the right instinct. Its answers were then set against the figures a human estimator using Paperless Parts would get.

Subscribe on YouTube@sagnikteaches

Read those rows the way an estimator would. The set-up items come mostly from reading the drawing's title block and notes, and the chatbot got them right. Every miss is a measurement: cut length, bends and blank size all come from geometry, and the custom GPT only had pictures of it. The flat pattern is worse than "about 10%" suggests. The area is about 9 percent smaller (158 against 174 square inches), but the long side is 3.7 inches short and the short side 1.2 inches too long, so a blank or a nest built on it is the wrong shape. The 36.87-degree "formed feature" shows the same habit in miniature: a diagonal edge in a screenshot looks like a bend, and a model that cannot measure has no way to tell.

The price line cannot be judged at all. About $300 for one enclosure and about $100 each for ten may be reasonable or not. The rates behind them were ones the chatbot was told to assume, and when Sawyer asked colleagues in his own office what the part should cost, the answers were "all over the place", so the page gives no baseline price.

Keep the caveats next to the numbers. Paperless Parts sells quoting software, its baseline is its own tool, and the demo goes on to argue for its own manufacturing-aware approach, which it calls Wingman. It is one part, one run and one prompt that was not published, with no ChatGPT model version named and no date on the page beyond a 2026 copyright. Nobody can rerun it exactly, and a newer model might do better or worse. What it does show is the type of error to look for, and that type matches how chat assistants work.

Reading a drawing is not measuring a part

Cut length and bend count are not judgement calls. They are measurements from the model of the part, and a chatbot that reads a drawing or a description is estimating, not measuring. Software built for quoting reads the geometry itself: the Paperless Parts pages describe detecting cut length, pierces and bends from the part and feeding them straight into pricing formulas. That is a different kind of tool from a chat assistant, and the difference is exactly where chatbot quotes go wrong.

If you were planning a custom GPT for quoting like the one in the demo, note that OpenAI has scheduled custom GPTs to retire on 11 December 2026 and is moving creators to plugins; ChatGPT Projects also keeps instructions and files together. Nothing in the reasoning changes. A Project can hold your rate card and your terms, which helps with wording and consistency, but it does not give a chatbot the ability to count bends on a part it cannot measure.

Why a chatbot gets these numbers wrong

• It reads, it does not measure. Given a PDF drawing or an image, it interprets what it sees. A complex flat pattern with 40 cut-outs has more perimeter than a glance suggests, and small features get missed.
• It fills gaps plausibly. If the drawing does not make a bend obvious, such as a return flange hidden in another view, the model produces a count that fits a typical part of that shape.
• It does arithmetic by pattern. Multi-step costing with rates, times and quantities is where language models slip. Why AI is bad at maths covers this and the checks to run on any AI-produced total.
• It does not know your shop. A price of a few hundred dollars for one part might be too high or too low. Without your laser rate, press brake setup times and overhead, there is no way for the chatbot to be right except by luck.

Here is how that looks on screen. Asked to "price 25 of the attached bracket" with a PDF drawing, a chatbot returns something like this (illustrative):

Material: mild steel, 1.5 mm, blank approx. 420 x 310 mm
Estimated cut length: approx. 2,000 mm (outer profile plus holes)
Bends: 8
Laser time: approx. 1.7 min per part
Press brake: 8 bends x 20 s = 2.7 min per part
Suggested price: $22.70 per part for 25, plus setup

Every line is laid out like an estimator's sheet, and that is the trap. Three inputs are guesses presented as measurements: the cut length, the bend count and the laser time that follows from both. The fix is mechanical. Open the flat pattern in your CAD or nesting software, read the true cut length and pierce count, count the bends on the formed model, then put those into your own costing sheet. On the bracket used below, that gives 2,450 mm and 12 bends. What you never do is nudge the chatbot's total by a few per cent and send it, because a total built on a wrong bend count is wrong in a way no percentage correction fixes.

What 18 percent and 4 bends do to a price

The demo published no estimator's price, so it cannot show what its misses cost. To see the effect, apply misses of the same size to an illustrative part at illustrative rates. None of these figures is a real shop's, and the part is not the enclosure from the demo. It is a mild steel bracket, 1.5 mm thick, with a flat blank of 420 by 310 mm, a true laser cut length of 2,450 mm, 14 pierces and 12 bends.

Batch costs total $100, and the per-part variable cost is $15.92. For one part, cost is $115.92; with a 30 percent markup the price is about $150.70. For a batch of 25, the cost is $15.92 plus $4 of batch cost per part, or $19.92, and the price is about $25.90.

Now apply the two misses from the vendor's test: cut length 18 percent low (2,009 mm instead of 2,450) and 8 bends instead of 12. Laser time drops to 1.70 minutes, $3.41 a part. Bending falls to 2.67 minutes, $4.00 a part. Variable cost becomes $13.48. At 25 parts the cost is $17.48 and the marked-up price is $22.72, so the quote is $3.17 a part or about $79 across the batch too low. On a single part the gap is only a few dollars because setup dominates.

That is the sneaky side of it. On this part the damage looks small. The wrong bend count matters more than the money suggests: 12 bends may need a different sequence, a special tool or an offset die that 8 bends would not. A quote that says "8 bends" can also hide a part that cannot be formed in the order the AI assumed. The cost model is only as good as the geometry going in, and the cost of errors rises with batch size, material price and the number of parts on the order.

It also shows why a chatbot's single-part price cannot be judged in isolation. In this illustrative shop the right price for one part is near $151. In a shop with higher setup charges or a minimum order value, twice that could be perfectly normal. A quote is right or wrong against your rates, which is the information the chatbot did not have. The same goes for the demo's figure of about $300 for one enclosure: with rates the chatbot was told to assume and no estimator's price published, there is nothing to judge it against.

Where AI does help in quoting

None of this means leave AI out. It means give it the jobs where it is strong and keep the measurements and totals in your own method. Four places where it earns its place:

Reading the RFQ email

RFQs arrive as rambling emails with attachments. AI is good at pulling out the request into a checklist. A sample prompt and output (illustrative):

Read this RFQ email and list: part name, quantity (and any break quantities), material and thickness, finish, tolerances, certifications, delivery date, drawing files attached, and anything unclear that I must ask the customer. Do not estimate prices.

[paste email]

Part: bracket, drawing BR-204 rev C
Quantities: 25 now, "possibly 100 per quarter"
Material: mild steel, thickness not stated
Finish: "powder coated, black"
Tolerances: general +/-0.5 mm per drawing note
Delivery: "end of next month"
Unclear: thickness; colour reference; whether the 100 per quarter is a commitment; whether material certificates are needed

The output is useful because it names what is missing. The customer email that asks "what thickness, and is the 100 a quarter firm?" often decides whether the job is profitable. For a fuller template that turns these emails into draft quotes, see the machine shop quote template for RFQ emails.

Drafting the covering text

Once the estimator has a price, AI can write the covering email, the notes on lead time and the paragraph on what is included and excluded. Give it the final figures and your standard terms and tell it to change nothing numerical.

Checking a finished quote for omissions

Paste the quote and the RFQ and ask which requirements in the RFQ are not covered. A reviewer prompt like this will often catch a missed finish or certificate. It is a second pair of eyes, not the author.

Summarising your own history

Give it an export of 50 old quotes with actual hours and ask which parts consistently overran. AI is fine with data you supply because it is analysing, not recalling. Always spot-check three of its claims against the sheet.

Run your own test before any AI touches a live quote

No outside test, including the vendor's demo above, tells you how a tool handles your parts. The only result that matters is what it does on your own jobs. A fair test takes an afternoon:

1. Pick 12 past jobs you quoted and ran. Include four simple parts, four medium and four awkward ones, so the test is not flattered. You must know the true cut length, bend count and run time for each.
2. Hide the answers. Give the tool only what a customer would send: the drawing or file and the RFQ text.
3. Record each output. For every job note the tool's cut length (or equivalent), feature counts, time estimate and price.
4. Score against your records. Use the table below.
5. Look at the spread, not the average. One 40 percent miss on a complex part matters more than ten small ones.

The bars are mine and illustrative; set your own to match your margin. If your margin is 12 percent, a tool that is off by 15 percent can turn a profit into a loss. Expect a general chatbot to do far better on the text tasks than on geometry. Tools built around CAD data are a different category, and the same test applies to them: ask the vendor for a trial and run your 12 jobs. That includes the company behind the demo, whose comparison was judged by the people who sell the software on the other side of it.

A quote that looked right and cost the job

Here is a made-up but typical case from a small machine shop. An estimator, short of time, asked a chatbot to price a turned stainless part from a PDF drawing. The reply gave a unit price and a material cost that looked sensible and used the right alloy. The estimator sent it with light edits.

The order arrived, and the first batch showed the problem. The drawing carried a note for a surface finish that needed a second grinding operation, and a thread callout that required a thread gauge the shop did not own. Neither was in the quote. The chatbot had read the dimensions and priced a plain turned part. The shop absorbed the grinding subcontract and the gauge, and the margin on the order disappeared.

What would have caught it: the requirement-extraction prompt above, which lists notes, finishes and callouts as separate lines, plus a rule that every drawing note maps to a line on the quote or an explicit exclusion. It is the same discipline as catching made-up figures in AI-drafted proposals: do not trust the total, trace each line.

Five-minute sums that catch a bad quote

You do not need software to sanity-check a number that a chatbot gives you. Two or three quick sums expose most big errors. Using the bracket from earlier (illustrative):

• Weight check. 420 mm by 310 mm by 1.5 mm is 195.3 cubic centimetres. Mild steel at about 7.85 grams per cubic centimetre gives roughly 1.53 kg for the blank. If the quote's material line implies 3 kg or 0.8 kg, the area or thickness was misread.
• Cut-time check. 2,450 mm of cutting at 2,000 mm a minute is 1.2 minutes before pierces. If a quote shows 5 minutes of laser time on this part, something has been added without explanation.
• Bend-time check. Count the bends yourself on the drawing, tick each one off with a pen and compare with the quote's count. Twelve bends at 20 seconds is four minutes; a quote built on eight is 33 percent short on this line.
• Ratio check. Compare the price with your last three similar parts. A quote more than 25 percent away from your own history needs a reason written next to it.

These checks take five minutes and need nothing but a calculator and the drawing. They also double as the scoring method for the 12-job test above.

What to put in a Project, and what to keep out

If you use a chat assistant for the text work, give it a Project (ChatGPT Projects, Claude Projects or a Gemini Gem, which Google is moving to skills on personal accounts from November 2026) so it applies the same wording every time. Good contents are your standard terms, a template for the covering email, a list of questions you always ask customers, and examples of your best past quotes with prices blanked. Poor contents are live rate cards presented as something the assistant should calculate with, and customer drawings covered by a confidentiality agreement unless your plan keeps business data out of training and your customer allows it. Check the privacy settings on your plan before uploading customer files; business plans from the major vendors do not train on your content by default, but consumer plans have a switch you must check.

A good test of a Project's setup: ask it to quote a part and see whether it asks for the missing facts or invents them. An assistant told to "never state a price; list what is needed to price this job" is useful and harmless. One told to "act as our estimator" will do exactly that, confidently, and that is how a believable wrong price gets made.

Three set-ups, ranked by risk

1. AI drafts text, estimator prices. Low risk. The chatbot reads the RFQ, lists the missing questions and writes the covering email. Numbers come from your spreadsheet or quoting software.
2. Quoting software reads the geometry, AI helps around it. Medium risk, worth it for volume. Geometry-based tools measure cut lengths, bends and features and apply your rates; the estimator reviews outliers. Test it with your 12 jobs first, and see what AI quoting software costs a machine shop before you ask for a trial.
3. Chatbot prices from a drawing. High risk. You are relying on a text model to read geometry and do multi-step maths with rates it was told in a prompt. The vendor's demo and the costing above show what a few misreads cost.

Most shops should start at the first and consider the second once quote volume justifies it. If you buy parts as well as sell them, comparing supplier quotes side by side with AI is the safe mirror image of this, because you supply the quotes and the AI only tabulates them.

Putting a person in the loop without losing the time saving

The point of AI in quoting is speed on the dull parts. Keep that by splitting the job:

• Minutes 0-5: AI turns the RFQ into the checklist. The estimator answers the questions or sends them to the customer.
• Minutes 5-25: The estimator prices using the shop's method.
• Minutes 25-30: AI checks the quote against the RFQ and drafts the email. The estimator reads it and sends.

Used like that, a half-hour quote might drop to 15 to 20 minutes without any change to the pricing method. Write a rule down: no number in a customer-facing quote that came only from a chatbot. If a number came from a tool, record which tool and what the estimator checked. The near-twin of this tutorial for plastics, quoting an injection moulding job, applies the same logic to tooling and cycle time.

If you would like help working out where in your quoting flow AI is safe and where it is not, that is what an AI implementation consultation covers; you can book one here.

Questions from job shops about AI quoting

Can ChatGPT price a part from a drawing?

Not reliably. It can read a drawing's notes and suggest what to ask, but it does not measure geometry the way a CAD-based estimator does. Cut lengths, bend counts and areas must come from software or a person, and the price must use your own rates.

What did the vendor's ChatGPT test get wrong?

In a Paperless Parts demo, a custom GPT quoting a two-piece aluminium enclosure gave a laser cut length of 128 inches against 155.5, counted 8 bends against 12 and read a diagonal laser cut as a formed feature. Part number, material, finish, bill of materials and routing were fine. It is one informal vendor run with no baseline price.

Which numbers in a chatbot quote need checking first?

Cut length, pierce count and bend count, because they are measurements from the part and drive machine time. Then material weight and every drawing note: finishes, threads, certificates. Applied to an illustrative 25-part bracket, misses the size of the vendor's test (cut length 18 percent short, four bends missed) make the batch about $79 too cheap while the quote still looks normal.

Is there any safe use of AI in quoting?

Yes: reading RFQ emails, listing missing information, drafting the covering text and checking a finished quote for omissions. The numbers should come from your estimating method, with a person approving the total.

How many past jobs should I test on?

Use 10 to 20 jobs you quoted and then actually ran, so you know the true time and cost. Include simple and awkward ones. A tool that cannot match your own history on those should not price new work unsupervised.

Custom GPTs are being retired. Does that change anything?

Not the lesson. OpenAI has scheduled custom GPTs to retire on 11 December 2026 and is moving creators to plugins. ChatGPT Projects, Claude Projects or Gemini Gems can hold your rate card and standard terms, which helps with consistent wording, but none of them adds the ability to measure a part.

Further reads

• AI Myths That Stop Small Businesses Getting Started — Common beliefs about AI that stop owners testing it properly.
• AI Freight Quoting: Turn Customer Emails Into Draft Quotes — Another trade where customer emails become draft quotes.
• How to Automate Purchase Orders and Supplier Emails With AI — Automate the purchase orders that follow once a quote is won.
• How to Match Supplier Invoices to Purchase Orders Automatically — Check the other end of the job: supplier invoices against orders.
• AI Supplier Management: Track Prices, Lead Times, and Risk — Keep track of outside processors, material prices and lead times.
• Can AI Draft a Lockout Tagout Procedure? What Must Be Verified — A shop-floor document where AI drafts and a person must verify.
• AI Tools for Welding Shops: Quotes, Drawings and Job Scheduling — A welding shop owner's guide to AI on the office side: a drawing-based quote check, cut lists, a schedule stress test and where weld sign-off stays human.
• How to Price Welding Jobs: Build a Rate Sheet With AI — A welding rate build-up, a fixed-price gate quote, per-inch pricing from timing, and an AI check of quoted against actual hours on twelve past jobs.
• Can AI Quote Powder Coating Jobs? What It Gets Right and Wrong — AI can total a powder coating quote from a part list, but it misses prep, masking and oven loads unless you feed it your own timings.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Paperless Parts demo page "LLMs vs. Manufacturing-Aware AI", a video transcript by co-founder and chief scientist Scott Sawyer, undated (copyright 2026), https://www.paperlessparts.com/demos/llms-vs-manufacturing-aware-ai/ (read 5 November 2026); Paperless Parts sheet-metal fabrication page on geometry-based costing variables; OpenAI's custom GPT retirement and migration FAQ and Projects help page; Google's Gems help page; OpenAI, Anthropic and Google Workspace pages on business-data training. The test figures are the vendor's own; the bracket, shop rates and prices in the costing example are illustrative.