ChatGPT Apps for Manufacturing: 7 Small-Factory Jobs and Limits

A small manufacturer can use ChatGPT to triage customer enquiries, draft work instructions, write non-conformance and corrective-action reports, summarise maintenance and downtime logs, draft supplier emails, tidy planning spreadsheets and write product documents. It should never price jobs unaided, sign off safety or tolerance decisions, or receive drawings covered by an NDA or export controls.

The reasoning is simple. ChatGPT is good with words and tables, and the factory has plenty of both. It is unreliable with geometry, shop costs and anything where a confident wrong answer leaves a scrap part or a customer dispute behind. The seven jobs below all sit on the first side of that line; pricing, tolerances, quality release and safety content sit on the second.

Follow me on Instagram@sagnikteaches

One note on products. OpenAI is retiring custom GPTs: they are scheduled to stop on 11 December 2026, and OpenAI's migration route for existing ones is to plugins. For a new reusable setup, use a ChatGPT Project with your instructions and reference files, or the equivalent in another assistant. The jobs below work the same way in a Project.

Connect on LinkedInSagnik Bhattacharya

The test that shows why quoting stays with an estimator

Paperless Parts, a company that sells quoting software, ran what it called a "vibe quoting" test. It built a custom GPT and gave it a blueprint, screenshots of the 3D model (the vendor notes ChatGPT does not accept CAD files natively) and a detailed prompt, and asked it to assume its own shop rates and markups. The part was a simple two-piece sheet-metal enclosure in 5052 aluminium, a U-shaped cover with vent slots, captive screw features and small pierced holes. The run took about ten minutes. A human estimator working in Paperless Parts provided the baseline.

Subscribe on YouTube@sagnikteaches

The assistant also read a diagonal laser cut at 36.87 degrees as a non-90-degree formed feature, which would have distorted the costing. Setup items such as part number, material, finish and the bill of materials came out acceptably, and it proposed a reasonable routing. The vendor's speaker concluded the approach was not reliable for this task, citing missing integrations, no native 3D file support, possible gaps against NDA, security and export-control requirements, slowness and errors in detail.

Read it with the right amount of salt. It is a single part, tested once, by a company that sells a competing product, and the human baseline came from its own software. It is evidence of a failure mode, not a measurement of ChatGPT in general. But the failure mode is the one that matters: the assistant sounded plausible, the setup looked right, and the numbers that drive cost (cut length, bends, flat size) were each wrong in a different direction. A quote built on those would be priced on a part that does not exist.

That test is the limit this tutorial uses. Anywhere a job needs the assistant to understand geometry, tolerances or real shop cost, treat it as out of bounds. Anywhere it only needs to read, sort, summarise or draft, it is useful.

Job 1: sort and answer incoming enquiries

What it does. Requests for quotes arrive with half the information. The assistant reads the email text (not the drawing) and lists what is stated, what is missing, and the questions to ask. It also sorts enquiries into "we make this", "we could make this", and "not for us".

Illustrative example. A ten-person sheet-metal shop gets 30 enquiries a week. The estimator spends 5 minutes on each reading and replying, 2.5 hours in total, and half the replies are the same "which material, what quantity, what finish?" The assistant drafts the clarifying reply in seconds, and sorts the "not for us" group (for instance, large castings) with a polite decline.

You help a small sheet-metal fabrication shop. Read the enquiry text below.
1. List what is stated: part type, material, quantity, finish, delivery date.
2. List what is missing that we must know before quoting.
3. Draft a short reply asking only for the missing items.
4. Label the enquiry "we make this", "we could make this" or "not for us", using the capability list below.
Do not estimate price or lead time.

Capability list: laser cutting up to 3 m x 1.5 m, bending up to 3 m, aluminium and mild steel up to 6 mm, powder coat via a local partner.

Enquiry: "We need 200 of a bracket in 3 mm steel, powder coated black, by the end of next month. Drawing to follow."

A realistic (illustrative) output marks quantity, material, thickness and finish as stated, lists tolerances, drawing revision, and whether the coating needs a specific thickness or colour code as missing, and labels the enquiry "we make this". What you fix: it may add a polite line about "competitive pricing", which you delete.

Limit. No drawing goes in. If the customer attaches one, the estimator opens it, not the assistant.

A quick sum on the enquiry job

Using the earlier illustration, 30 enquiries a week at 5 minutes each is 150 minutes. If drafts cut that to 2 minutes of reading and editing per enquiry, the weekly figure is 60 minutes, a saving of 90 minutes, or about 6 hours a month. At an estimator's loaded cost of, say, $40 an hour, that is $240 a month of time returned to quoting work. A ChatGPT Business plan starts at two seats, at $20 a seat a month billed annually or $25 billed monthly, so $40 to $50 a month. On paper the enquiry job covers that several times over, but the first month also carries setup and the time spent checking drafts, so the real case builds when the same Project takes on supplier chasers and report drafts. Start with one job and add others only as each shows a saving.

Job 2: draft work instructions from your own notes

What it does. Experienced operators know the setup by heart and no one has written it down. Record a voice note or type bullet points and the assistant turns them into a numbered work instruction with a materials and tools list, steps, check points and a "stop and call a supervisor if" section.

Illustrative example. A press-brake operator describes how they set up a 90-degree bend on 3 mm steel: tooling choice, back-gauge, first-piece check with a protractor, a reminder to check grain direction. The assistant produces a one-page instruction. The operator corrects the tooling reference and adds a line about guard checks. A new starter can follow it.

Limit. Safety content (guarding, lock-out, PPE, emergency stops) comes from your safety procedures and your responsible person, not from the assistant's general knowledge. Use the assistant for the layout and plain wording, then have the person responsible read every line. For the wider picture of documentation, ISO 9001 documentation: hire a consultant or draft it with AI? sets out which parts of a quality system an assistant can draft and which need a person's judgement.

Job 3: write the first draft of a corrective-action report

What it does. A customer complains. You need an 8D or similar report: problem description, containment, root cause, corrective actions, verification. The assistant structures your notes into the report format and flags where your root-cause reasoning has gaps.

Illustrative example. Notes read: "customer returned 12 of 400 brackets, holes 0.4 mm off, found drill wear on station 3, replaced bit, added first-off check." The assistant produces a draft with problem statement, containment (quarantine and sort of the 400), cause (tool wear), and corrective action, and then asks the awkward question: "Why was wear not detected by the existing check?" That prompt is where it earns its keep.

The detailed method, with a customer-complaint walkthrough, is in 8D report with AI: answer a customer complaint step by step.

Limit. Root cause is a finding, not a draft. The assistant can propose hypotheses; the people who were on the floor test them. Do not send a report to a customer that nobody has read line by line.

Job 4: summarise downtime and maintenance logs

What it does. Machine logs, maintenance sheets and shift notes are free text mixed with timestamps. You paste a month of entries and the assistant groups them by machine and cause, counts incidents, and points out repeats.

Illustrative example. A small plastics moulder keeps a shared sheet with 140 stoppage entries in a month. The assistant groups them: 38 percent "material feed" on two machines, 22 percent "mould change overrun", and a cluster of "sensor fault" entries on one press during night shifts. The supervisor now has a place to look first. Ask for a table with entry counts and the row numbers, so you can verify the counts with a pivot table.

Below are 140 stoppage log entries (date, machine, minutes lost, free-text note). Group them by likely cause, using the notes. For each group give: number of entries, total minutes lost, machines involved, and the row numbers.
If a note is unclear, put it in "unclear" instead of guessing. Do not suggest fixes yet.

Run the same sheet through a pivot table once. If the assistant's totals do not match, find out why before trusting it with next month's. For the spreadsheet side, the ChatGPT Excel guide covers building and checking formulas with an assistant.

One more caution on the downtime job. Log entries written by tired people at the end of a shift often use shorthand: "Mach 3 jam again", "same as Tues". The assistant cannot know what "again" refers to unless the earlier row is in the same paste. Include at least the preceding week, and tell it to keep "unclear" as a category. A group of 30 unclear entries is itself a finding: it tells you the log format needs a fixed list of causes to choose from, which costs a morning to set up and improves every later analysis.

Limit. It finds patterns in what people wrote. It does not know the machine. A "sensor fault" may be a loose connector; only a technician can say.

Job 5: draft supplier emails, chasers and purchase queries

What it does. Late deliveries, wrong material certificates, price-increase notices and requests for a revised delivery date. The assistant drafts the message in the tone you want from the facts you give.

Illustrative example. A material supplier is five days late on a 2 tonne order of 5052 aluminium sheet. The buyer pastes the order number, the promised date, the production start date and the cost of idle time per day, and asks for a firm but polite message requesting a confirmed delivery date within 24 hours and a part-shipment if possible. The draft is 80 words. The buyer adds a phone call to follow.

Limit. Commercial terms, penalty clauses and anything that could be read as a legal notice come from the purchasing manager. Keep supplier price lists and negotiated terms out of the prompt unless your plan and contracts allow it.

Job 6: sanity-check planning spreadsheets

What it does. Many small shops plan on a spreadsheet: jobs, machine hours, due dates. The assistant reads the sheet and reports conflicts: two jobs on one machine at the same time, a due date before the earliest finish, a job with hours but no machine.

Illustrative example. A weekly plan lists 22 jobs across four machines. The assistant reports that the powder-coat partner's collection day appears before three jobs finish, and that machine 2 is booked for 52 hours in a 40-hour week. The planner moves two jobs and phones the partner.

Limit. It checks the arithmetic and the logic of what you wrote down. It does not know that the press is down for repair or that a customer will change priority tomorrow. If you are choosing between a chatbot over your own data and a planning add-on to your inventory system, AI systems for manufacturing: MRP add-on or chatbot on your own data? sets out the trade-offs.

Job 7: write product and compliance documents from facts you supply

What it does. Datasheets, care instructions, packing lists, website product text, tender questionnaire answers, certificate checklists. You supply the facts; the assistant writes the clean version.

Illustrative example. A small maker of stainless-steel food trays has a one-line spec: "304 stainless, 1.2 mm, dishwasher safe, made to order." The assistant produces a product page, a one-page datasheet, a care sheet and a short answer to a buyer's questionnaire. The owner checks every number against the spec and removes a sentence about "food-grade certification" the owner never claimed.

That last correction is the real lesson. An assistant fills gaps with plausible claims: certifications, load ratings, lifetimes. Tell it to use only the facts you provide and to mark anything else "to confirm", then read the output as a sceptical customer would.

Standing instructions for a factory Project, filled in

A Project holds instructions and reference files that apply to every chat inside it, so you write the rules once. Here is a filled-in example for the enquiry job, which you can adapt. The point is that the limits sit inside the setup, not in anyone's memory.

You assist a small sheet-metal shop's estimator. You read enquiry emails and draft replies.
Always: list what is stated, list what is missing, draft a short reply asking only for the missing items.
Never: give a price, a lead time, a tolerance, a material recommendation or a statement about certification.
If an email mentions a drawing, model or specification number, say "estimator to review" and do not describe its contents.
If an enquiry mentions defence, aerospace, controlled data or an NDA, stop and write only: "Flag to the manager before replying."
Tone: plain, polite, no sales language.
Use only the facts in the email and in capability-list.pdf. If something is not in either, write "to confirm".

The stop rule for controlled work is worth the effort. It turns "remember not to paste that" into a default the assistant itself follows, though you should still not rely on it as the only safeguard: the person pasting remains responsible for what goes in.

What the quoting test says about the other six jobs

It is easy to over-read one vendor's test. It shows an assistant going wrong on geometry-dependent numbers when given screenshots and a blueprint. It does not show that the assistant cannot read an email, restructure your notes or group your downtime log, because those jobs never ask it to measure a part. The honest reading is narrower: when the answer depends on something the assistant cannot perceive accurately, such as the length of a cut or the count of bends on a drawing, it will still produce a confident number. The question to ask of any new job is "what does the assistant have to perceive or know that is not in the text I give it?" If the answer is "nothing", you are in the safe zone. If the answer is "the shape of the part", "how the machine behaves", or "what the customer meant by that drawing note", you are in the zone the test exposed.

Apply it to each of the seven. Enquiry triage needs only the words of the email. Work instructions need the operator's words. Corrective-action drafts need your notes. Downtime logs need the log. Supplier emails need the order facts. Planning checks need the sheet. Product documents need your specification. None needs the assistant to see the part. That is why they are the seven.

What to keep out of ChatGPT altogether

• Drawings and models under NDA. A non-disclosure agreement usually limits who may see the material and where it may be stored. Uploading it to a third-party service may breach that, even if the service says it will not train on it. Read the agreement, or ask the customer in writing.
• Export-controlled technical data. Some parts and their data are subject to export rules that limit who can access them and in which places they may be processed. If your customer tells you a part is controlled, do not put its data into a general assistant. Take advice from a qualified adviser on what is permitted.
• Customer-specific rate cards and margins. Confidential by contract and commercially sensitive.
• Employee personal data beyond what a task needs, such as medical or disciplinary details.
• Proprietary process recipes where the know-how is the business: special heat-treatment cycles, mix ratios, tooling designs.

A practical way to hold the line is a one-page list taped near the office: green (public information, your own templates), amber (anonymised job data, supplier names), red (anything above). Every new job for the assistant gets a colour before it starts.

Jobs a factory should never hand over

These are the jobs where a plausible wrong answer is worst. The quoting test is the template for why.

1. The price on a quote. A quote is a commitment. Geometry-driven costs (cut length, bends, machine time, setup) are exactly what the test got wrong. Use the assistant for the email around the quote, never the number in it.
2. Tolerance, fit and material decisions on safety-relevant parts. An engineer signs these.
3. Quality release. Whether a batch passes is decided by measurement and a qualified person.
4. Safety instructions and risk assessments. The assistant can improve the wording of a draft you wrote. It must not originate the content.
5. Anything that would have to be explained to an auditor. If someone must be able to say "I decided this, and here is the record", then that person decides.

A mistake that came from a tidy-looking answer

A pattern worth guarding against, shown as an illustration. A planner asks the assistant to estimate machining time for a batch from a description of the part and a list of operations. It returns a clean breakdown: 4.2 minutes cycle time per part, 50 parts, about 3.5 hours plus setup. The planner books the machine for half a day. On the day, the part needs a second fixture and a deburring step the description never mentioned, and the real cycle is nearly twice the estimate. The job overruns, the next job slips, and the customer gets a late-delivery message.

Nothing in the answer looked wrong. The failure was asking the assistant for a number that depends on a physical process it cannot see. The fix is a rule: times and costs come from your own history (last time you ran something similar, from the job record) and from the person who will run it, and the assistant may format and sanity-check, not originate.

How to start without a project

Pick one job from seven. Use Job 1 or Job 7 if you want something safe to learn on, Job 4 if you already keep a messy log. Set up a ChatGPT Project (or the equivalent in your assistant) and put your standing instructions into it once: the capability list, the tone, the list of things never to include, and the rule that missing facts are marked "to confirm". Run it for two weeks alongside your normal process, and note time saved, corrections needed and any wrong output you caught.

At the end, write three lines: what it does well, what it gets wrong, what we will not use it for. That is your factory's own version of the Paperless Parts test, and it will be worth more to you than any published one. If it saves time, add a second job. If you would like to compare where AI could pay off across quality, planning and quoting, where should a small manufacturer start with AI? and AI quality control for small manufacturers take it further.

Checking that it is actually working

Pick three measures before you start, and write them down. For the enquiry job, minutes from receipt to first reply. For work instructions, the number of questions a new starter asks the supervisor in their first week. For corrective-action reports, the days from complaint to first report sent. Measure for two weeks before introducing the assistant and two weeks after. If the numbers have not moved, either the job was the wrong one or the prompt needs work; do not add more jobs on top of one that has not shown a gain.

Also count corrections. If a person edits more than a third of each draft, the assistant is making work, not saving it, and the standing instructions need rewriting with real examples of good output. Add two or three of your best past emails or reports to the Project as models. That one change often does more than any amount of clever prompting.

What small manufacturers ask before they start

Can I upload customer drawings to ChatGPT?

Not unless your agreements allow it. Many drawings sit under a non-disclosure agreement, and some parts fall under export-control rules that restrict where technical data may go. Check the contract and ask the customer. If in doubt, describe the part in your own words without the drawing, or use a tool your customer has approved.

Is ChatGPT accurate enough to quote jobs?

Not as the quote. In one vendor test of a sheet-metal part, the assistant under-counted bends by about a third and laser cut length by about 18 percent. It can help read requests and list questions, but prices come from your costing method and an estimator's judgement.

Which ChatGPT plan should a small factory use?

Use the plan that matches what you will paste in. Business plans do not train on business content by default and need at least two seats. Individual plans let you turn training use off in settings. Whichever you use, keep customer drawings and confidential rates out unless your contracts allow them.

Do I need ChatGPT apps or can I paste text?

Pasting or uploading is enough for every job here and gives you control over what the assistant sees. Connected apps can read your email or files directly, which is convenient but exposes more than you intend. Try the manual version for a month before connecting anything.

Further reads

• Is Cin7 ForesightAI Forecasting Worth It for a Small Maker? — Forecasting inside inventory software versus a general assistant.
• Automation in Clothing Manufacturing: What a Small Workshop Can Automate — What a small workshop in another trade can automate.
• How Much Does AI Visual Inspection Cost a Small Manufacturer? — What camera-based inspection costs when ChatGPT is not enough.
• ChatGPT Apps for Logistics: 7 Jobs for a Small Carrier or 3PL — The same exercise for the carrier that ships your goods.
• How to Use Microsoft Copilot for Data Analysis in Excel — Analyse production exports if you work in Microsoft tools.
• Best AI Tools for B2B Marketing at a Small Manufacturer or Distributor — AI tools for marketing a small manufacturer to trade buyers.
• Daily Production Report: Summarise Counts, Scrap and Stops With AI — Three messy shift sheets become one page: units against plan, scrap by reason, stops by cause, and the one thing the owner must act on.
• 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 'LLMs vs manufacturing-aware AI' (vendor test, single part); OpenAI ChatGPT Business pricing page, data controls help article and custom GPT retirement and migration FAQ (November 2026).