Can ChatGPT Read Engineering Drawings? What to Protect and Check

Often yes, but reading is the smaller question. ChatGPT can read many clean drawings and pull out dimensions, tolerances and notes, though it misreads small or dense details. Whether you may upload a customer's drawing depends on the NDA, export-control status and the plan's data terms. Check those first, and verify every number it gives you.

Most articles stop at "yes, it can read them". For a shop that quotes from other people's drawings, the expensive mistakes sit elsewhere: uploading something an NDA or export rule forbids, or trusting a tolerance the model misread. So settle permission before the upload, redact what the customer wants hidden, and treat every reading as a draft to check against the sheet.

Follow me on Instagram@sagnikteaches

Sort the drawing before you touch the upload button

Every drawing falls into one of three tiers, and the tier decides what is allowed. Make the sort a habit at the point a drawing arrives, before anyone opens a chat window.

Connect on LinkedInSagnik Bhattacharya

Tier 3 is the one with legal consequences beyond a lost customer. Drawings that carry export-control markings, or that describe items on a control list, are governed by rules such as ITAR (defence articles) and EAR (dual-use and commercial items with controls). In general terms, both treat releasing controlled technical data to a foreign person as an export even when it happens at home, inside one company (a "deemed export"). Both also exclude sending or storing such data through a service only when it stays end-to-end encrypted with approved cryptography, which is hard to square with a tool that has to read the file to answer you. Whether a particular service could ever meet those conditions is a legal question for your compliance adviser, not something to assume. If a drawing has any export-control marking, a distribution statement, or comes from a defence or aerospace programme, treat it as Tier 3 until your compliance contact says otherwise.

Subscribe on YouTube@sagnikteaches

The sort also protects you from your own staff. A short written rule such as "Tier 3 never goes into a chat tool; Tier 2 only after the checklist" works better than a long policy. The tutorial on classifying business data for AI offers a ready structure you can adapt.

What an NDA usually says that matters here

You will probably not find the words "artificial intelligence" in an older NDA. Look for the clauses that bite on any outside tool:

When the answer is unclear, ask. Here is wording you can adapt (illustrative, and have your own adviser review it before you adopt it):

Subject: Permission to use an AI assistant when quoting drawing [number]

Hello [first name],

To speed up quoting, we sometimes use an AI assistant to read drawings and list features and tolerances. We would upload [the drawing / a version with your name, part number and title block removed] to a business-grade service whose provider states it does not train on business data by default. Files stay in our workspace and we delete them after quoting.

Are you happy for us to do this for [project name]? If you would rather we did not, we will review the drawing by hand.

Kind regards,
[your name]

Keep the reply with the job file. A customer who says yes in writing has given you cover that no policy page can. A customer who says no has saved you a difficult conversation later. For the wider question of what happens to client files inside AI tools, the list of questions to ask before buying AI that touches client data is a good checklist to put to any vendor.

Plans, training and retention: what the vendor says

There are two separate questions: does the vendor train its models on your files, and how long does it keep them. They are answered differently by plan.

• Training. OpenAI's enterprise privacy page says it does not train its models on business data by default for ChatGPT Business, ChatGPT Enterprise and its API platform, unless a customer opts in to share feedback. On consumer plans, a setting in data controls (labelled "Improve the model for everyone") governs whether chats are used for training; check that it is off, and read the current label yourself, since names change.
• Retention. A deleted chat or a Temporary Chat is not gone at once. OpenAI says Temporary Chats may be kept for up to 30 days for safety, and that deleted conversations on Business and Enterprise are removed from its systems within 30 days unless the law requires longer. Deleting a chat also does not delete a copy of the file saved in your Library, which has to be removed separately. On Business and Enterprise, workspace admins control how long data is kept. Read the current help pages before promising a customer deletion "immediately".
• Who can see it. ChatGPT Business adds an admin console and SAML single sign-on, which matter if your NDA says "need to know". A consumer account used by three people sharing a login does not.
• Location. OpenAI offers data residency, meaning customer content stored at rest in a chosen region, to eligible ChatGPT Enterprise, Edu and API customers. If your customer's contract names a place, check whether your plan offers it before you upload.

None of this makes an upload permitted. It only tells you whether the vendor-side terms could fit your NDA. A business plan with a no-training default is a prerequisite for Tier 2 work, not a licence for it. If you have not yet set the training switches on the accounts your team uses, the guide to stopping AI tools training on your business data shows where they are.

Redaction that actually removes the information

For Tier 2 drawings where the customer agrees to an upload but wants the identity removed, or where you choose to limit exposure, redact before you upload. The common failure is a redaction that only looks done: a white box drawn over a name, with the name still in the PDF's text layer.

1. Decide what to remove. Customer name and logo, drawing and part numbers, project names, addresses, signatures, revision history that names people, and any note that reveals the end use. Keep what the AI needs to read: geometry, dimensions, tolerances, material and process notes.
2. Cover or crop it. Use a redaction tool that removes the content, or crop the title block out of the page.
3. Flatten. Print or export to a fresh PDF, or to an image, so the hidden layers and the text under the boxes are gone.
4. Strip metadata. Open the file properties and clear author, title, company and software fields, which often carry the customer's name.
5. Test it. Open the new file, try to select and copy text over the redacted areas, and search for the customer's name. Nothing should appear.
6. Keep the original apart. Store the redacted copy under a neutral name, such as a job number you generated yourself.

Redaction removes names, not the part. A distinctive geometry can still identify a customer's product, so redaction reduces risk and does not replace permission. For a Tier 2 drawing where the geometry itself is the secret, the NDA question stands even when the title block is gone.

What ChatGPT actually reads, and where it slips

Given a clean, high-resolution PDF or image of a mechanical drawing, a current ChatGPT model can usually identify the views, read the title block fields, list dimensions and tolerances and summarise notes. OpenAI's upload help says ChatGPT can analyse both the text and the images in a PDF attached directly to a chat, but that embedded images may not always be included, and suggests uploading them separately as PNG or JPEG files when they are not. For drawings, a high-resolution PNG of each sheet is the surer route. I have not published a measured accuracy figure, and a general one would mean little for your drawings. Test it on a drawing you already know thoroughly. The mistakes below are the types to look for, with an illustrative case for each.

• Decimal and unit slips. A tolerance of 0.05 read as 0.5; a dimension in inches read as millimetres because a note was missed.
• Fit callouts. A hole marked as Ø8 H7 read as Ø8 H9, a looser fit, which changes the process and cost.
• Geometric tolerancing. Feature control frames, with their symbols, datum letters and modifiers, are small and dense. A position tolerance of 0.1 at maximum material condition read as 0.1 regardless of feature size loses the bonus tolerance.
• Threads. M8x1.25-6H read as M8x1-6H, a different pitch.
• Section and detail views. A dimension from a detail view attached to the wrong feature.
• Notes that apply everywhere. "Unless otherwise stated, break all edges 0.2 max" read as a note for one corner, or dropped.
• Revision confusion. Two revisions on one sheet and the model quotes the superseded value.
• Scans. A fax-quality scan or a photograph taken at an angle degrades every item above.

A sample extraction, with the errors marked

Say you upload a redacted drawing of a small aluminium bracket and ask for a feature list. The model returns something like this (invented for illustration):

Material: 6061-T6 aluminium
General tolerance: +/-0.1 mm
Features:
1. Hole 1: dia 8 H9, depth through
2. Hole 2: M6x1, depth 12
3. Slot width 10 +0.1/0, length 30
4. Surface flatness 0.05 on datum A
Notes: Anodise, clear. Break all edges.

Compare with the drawing and you find that hole 1 is H7, not H9; the slot is +0.05/0, not +0.1/0; the flatness is 0.02, not 0.05; and the anodise note continues with "10 microns minimum", which was dropped. Four errors in a list of seven items, from a drawing a person would read in two minutes. That is entirely plausible for a dense drawing, and the correction costs you less than a bad quote does. The AI was still useful, since it produced a structure to check against, but the reading cannot be accepted as it stands.

Three set-ups for Tier 2 drawings, compared

When a customer allows AI use, or the NDA is silent and you have asked, there are three ways to do it. Choose by the sensitivity of the geometry and the effort you can afford.

• Redacted upload to a business plan. Cheapest and best-reading. Fits an NDA that permits service providers or a customer who has agreed in writing. Weak point: the vendor still receives the geometry, and deletion takes time.
• A model run on your own computer. The file never leaves your network, which answers most NDA and retention questions. Reading quality on drawings is usually weaker than the large hosted models, and setting it up needs someone comfortable installing software. Test it on three drawings you know before trusting it.
• No AI on that job. Some customers, and some drawings, rule it out. Read it by hand and use AI only on your own notes afterwards, for example to tidy the quote wording once all customer detail is stripped.

Say a shop has three live enquiries: a catalogue-style spacer (Tier 1), a customer bracket under a standard NDA with a service-provider clause (Tier 2), and a housing from an aerospace programme with a distribution statement on the sheet (Tier 3). The spacer goes straight through the checking routine. The bracket is redacted, uploaded to the shop's business plan after a one-line email to the customer, and checked feature by feature. The housing is read by hand by the programmer, and the compliance contact is asked whether any tool at all may be used. Three jobs, three routes, and the sort took a couple of minutes at the point of arrival.

A checking routine that fits a quoting workflow

1. Ask for a structured table. Request one row per feature with columns for feature, nominal, tolerance, unit, the view it appears in, and a confidence flag. Rows are easier to check than paragraphs.
2. Add a count. Ask for the number of dimensions and the number of GD&T frames per view. Count them yourself. A mismatch means something was missed.
3. Zoom in on risk. Tight tolerances, fits, threads and finishes drive cost. Check those by eye on the original, whatever else you skip.
4. Ask for a second pass. "List only dimensions with tolerance tighter than +/-0.05 mm and any feature with a surface finish below Ra 1.6." Compare with the first list. A feature that appears in one and not the other deserves a look.
5. Have a second person check the tight ones before the quote goes out, as you would with a hand-read drawing.
6. Keep a log of misses. Note which types the model misreads on your drawings. After ten jobs you will know where to look first.

A prompt that supports this:

This is a redacted mechanical drawing. List every feature as a table with columns: feature, nominal, tolerance (as printed), unit, view, and confidence (high, medium, low).
Rules:
- Copy tolerances and fit codes exactly as printed. If you cannot read a character, write UNREADABLE rather than guessing.
- List all general notes and title block fields separately and in full.
- State how many dimensions and how many feature control frames you found on each view.
- Do not infer values that are not printed.

Telling it that "UNREADABLE" is acceptable cuts the confident guesses. If you work in a spreadsheet, using ChatGPT with Excel shows how to paste such a table straight into a quoting sheet.

What a machine shop does with the readings

Once the drawing has been read and checked, the useful jobs begin. A small shop might use the table to:

• Sort features into easy, needs-care and needs-a-specialist before pricing.
• Flag which operations need a particular machine, fixture or inspection method.
• Draft questions for the customer: "The drawing shows H7 on the bore but the general tolerance block is +/-0.1; is the fit intended?"
• Turn the note list into a work-instruction checklist for the operator.
• Compare two revisions of a drawing and list every change, which is a strong use because the model only has to spot differences between two images and a person verifies each one.

Take a quoting example. A job has 12 features, of which four are tight. The AI draft takes five minutes; your check of the four tight features takes ten. That is 15 minutes against, say, 30 for a full manual read, and you have lost nothing in the checking that matters. If it saves 15 minutes on each of 20 quotes a month, that is five hours, which the shop can put into a second look at the hard features. These figures are illustrative; measure your own.

A missed margin note that turns into scrap

Picture a quote going out for 40 brackets at a price based on a general tolerance of +/-0.1 mm, when the drawing has a block note, outside the main view, tightening all machined faces to +/-0.05. The model listed the title block and the main views but not the lower-left note. The shop finds out at first-article inspection, after cutting the parts, and either scraps them or reworks them at a loss. The cause is the missing instruction to list every note in full, and the fix is the "all notes, separately and in full" line in the prompt above, together with a rule that a person reads the whole sheet margin to margin once before pricing, whatever the AI says.

Keeping it on the right side of the line

Draw the boundaries clearly and write them down in one page:

• Which tier each kind of job falls in, and who decides.
• The plan and account that may be used for Tier 2, and that personal accounts are not allowed.
• The redaction steps and who checks the result.
• The customer permission wording and where replies are stored.
• The checking routine and who signs the quote.
• That Tier 3 never goes into a chat tool.

Review it when a customer's terms change or the vendor changes its plan names or data terms. AI quality control for small manufacturers is the natural next step, once reading and quoting are under control. If you would like help drafting the one-page policy and the redaction routine for your shop, that fits my AI implementation consultation, and a 1:1 call can be booked here. This is not legal advice: confirm NDA and export-control questions with a qualified adviser.

Questions machine shops ask about uploading drawings

Does my NDA forbid me from using ChatGPT?

It depends on the wording, so read it. Look for limits on disclosure to third parties, rules about subcontractors and processors, a restriction to a named purpose, and return-or-destroy duties. If it is unclear, ask the customer in writing for permission. A short email costs less than a breach claim.

Is a business plan safer than a personal one?

For training, usually yes: OpenAI states that business data is not used to train its models by default on its business plans and API. But that addresses training only. It does not by itself satisfy an NDA, an export-control rule or a customer's own security terms, so check those separately.

What if the drawing is export-controlled?

Do not put it into a general-purpose chat tool. Controlled technical data can be restricted by who may access it, including staff and vendors, not only by where it is stored. Ask your compliance adviser or export-control officer first, and ask the vendor in writing whether the product is offered for that purpose.

Can ChatGPT check a drawing against our process capability?

It can list the tight tolerances and the features likely to need a special process, which is useful for a first review. It cannot know your machines' real capability or your inspection method. A programmer or quality engineer must make that call and sign the quote.

Can I use a locally run model instead?

A model running on your own computer keeps files off a vendor's servers, which can solve the NDA and export questions. Quality varies, and you still need to check every reading. Test it on a drawing you already know before relying on it.

Further reads

• Is It Safe to Put Customer Data Into ChatGPT? — What happens to data pasted into a chat tool.
• Can Solicitors Use ChatGPT Without Breaching Confidentiality? — How another confidentiality-bound profession handles the same issue.
• Where Should a Small Manufacturer Start With AI? — See where drawing review ranks among a maker's AI options.
• AI Systems for Manufacturing: MRP Add-On or Chatbot on Your Own Data? — Chatbot on your own data or an MRP add-on.
• AI Compliance Checklist for Small Businesses: What Applies to You — A wider compliance checklist for small businesses using AI.
• Cabinet Shop Drawings With AI: Box Sizes, Reveals and Hardware Checks — A pre-cut routine for AI-drafted cabinet drawings: box widths from the site measure, a gap budget for reveals, and hardware clearances read from data sheets.
• AI in Steel Detailing: Gains for a Small Office, and Tasks to Avoid — Four admin jobs a small steel detailing office can give to AI, the tasks that stay with detailers and checkers, and a pre-send checklist.
• AI Drawing Review: Check Architectural Sets Before You Issue Them — Where AI is reliable on a drawing set (index, notes, references), where it fails (dimensions, line types), and the checklist and sign-off that close the gap.
• AI for Technical Sales: Answer Spec Questions From Your Datasheets — Build an internal spec-answer routine with versioned datasheets, cited replies, abstentions and a 30-question accuracy test.
• Can ChatGPT Write G-Code? What a Machine Shop Must Check First — ChatGPT can draft simple G-code. Here is a faulty draft taken apart, a ten-program test log, what simulators miss and the sign-off before the spindle turns.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI's enterprise privacy page (updated 8 January 2026), its help pages on data controls and file uploads, its data-residency announcement and the ChatGPT Business pricing page; the EAR (15 CFR 734.13 and 734.18) and ITAR (22 CFR 120.50 and 120.54) definitions of export and their encrypted-data provisions. All checked 5 November 2026. Drawing examples and readings are illustrative.