Six AI Applications in Manufacturing That Need No New Machines

A small factory can use AI on six office-side jobs without buying sensors or machines: drafting quotes from enquiries, writing work instructions, chasing suppliers, drafting ISO procedures, checking a scheduling sheet, and sorting complaints. Each starts from text you already hold, and you read the output before anything leaves the building.

Most writing about AI in manufacturing describes predictive maintenance, camera inspection and digital twins. Those need hardware, data pipelines and budgets a 20-person shop does not have. The jobs below sit in the inbox, the shared drive and the planning spreadsheet. They are unglamorous and they work, because every one of them is a reading-and-writing task, and reading and writing is what current assistants do best. The running example is a 20-person sheet-metal shop making brackets and enclosures to customer drawings. Every figure is illustrative.

Follow me on Instagram@sagnikteaches

What to keep out of the assistant, and how to choose where to start

Before the six, a rule that applies to all of them. Treat anything a customer marked confidential as off limits until you have read your agreement with them and the data terms of the plan you use. As a general pattern, business plans of the main assistants do not train on business content by default, while free and personal plans may allow it unless you switch it off in settings. Check the current wording on the vendor's page, because it changes. For most of these jobs you do not need the drawing at all: you need the text of the enquiry, a list of dimensions or your own notes.

Connect on LinkedInSagnik Bhattacharya

The second rule is to keep arithmetic out of the model. Prices, capacity hours and tolerances belong in a spreadsheet with formulas you can audit. Let the assistant read, summarise, draft and flag. Let the spreadsheet calculate. The third is that a person reads everything before it is sent or filed, and the person is whoever knows the shop floor.

Subscribe on YouTube@sagnikteaches

For a first job, I would pick the third or the first, because the output is text you read and a mistake costs a rewrite, not a scrapped batch. If you want a broader view of where a small manufacturer should begin, this starting guide sets out the options.

1. Quote drafts from an enquiry

An enquiry arrives by email: "Need 250 of the attached bracket in 2 mm mild steel, powder coated black, delivery in three weeks." Someone in the shop reads it, finds the missing details, works out a price and writes back. The reading and writing take 20 to 40 minutes. The price needs the shop's own knowledge, so keep that in a spreadsheet.

What the assistant does

It turns the email into a structured list, notes what is missing, and drafts the cover email. It does not set the price.

Below is a customer enquiry. Extract into a table: part description,
quantity, material, thickness, finish, tolerance if stated, delivery date,
certificates requested. List anything missing that I would need to quote
(for example material grade, finish colour code, tolerance, packaging).
Then draft a reply under 120 words that confirms what we understood,
asks for the missing items, and says a price follows within one working
day of getting them. Do not state a price or a lead time.

[paste enquiry]

Illustrative output: a table with "Bracket, 250, mild steel, 2 mm, powder coat black, tolerance not stated, delivery 3 weeks, certificates none", then missing items "material grade, colour code, bend tolerance, packaging", and a short reply. The estimator checks the table against the drawing, then prices it in the sheet.

The sheet does the sum

A worked price for the 250 brackets, with the shop's rates: material $1.84 a part, laser cutting 1.6 minutes at $1.20 a minute ($1.92), bending 0.9 minutes at $1.00 a minute ($0.90), powder coat $1.10, packing $0.15. Cost per part $5.91. A 30% mark-up on cost gives $7.68 a part, and $1,920 for 250, plus a one-off $85 programming charge. If the model had been asked to price it, a slip such as multiplying instead of adding the bend time would pass unnoticed. In the sheet, the formulas are visible.

Start: take your last 10 enquiries, run them through the prompt, and compare the extraction with what the estimator found. Watch for: the assistant filling a missing tolerance with a typical one. The instruction "list anything missing" is the guard; read the table for any value the customer never wrote.

2. Work instructions that operators will actually use

Many small shops run on knowledge in one person's head. Writing it down takes hours nobody has. The assistant shortens the writing, not the knowing: the expert talks through the job, the voice note is transcribed, and the assistant turns it into a one-page instruction in a fixed layout.

This is a transcript of an experienced operator describing how to set up
the press brake for the bracket B-204. Turn it into a one-page work
instruction with these headings: Purpose, Tools and gauges, Safety,
Steps (numbered, one action each), Checks (what to measure, with the
tolerance), If it goes wrong. Use only what is in the transcript. If a
step seems to be missing or ambiguous, add it to a list headed
"Questions for the operator" at the end instead of filling it in.

Illustrative output: nine numbered steps for the setup, a checks list ("first part: bend angle 90 degrees, tolerance plus or minus 1 degree, check with the angle gauge"), and a questions list: "Step 6 mentions 'the usual back-gauge setting'. What is the value?" That questions list is the point. The first draft is never complete, and the quickest way to finish it is to put it in front of the operator.

What a shop gains is mostly in training and cover. A new starter can read one page and then watch the job, rather than shadowing for a week. If an experienced operator takes two days off, the instruction exists. In an illustrative case, the shop spent about 40 minutes per instruction against about three hours writing them from scratch, and wrote 12 in the first month.

Start: the three jobs that fail most often when the usual person is away. Watch for: safety steps. An assistant working from a transcript will leave out anything the operator did not say, and operators leave out what they do without thinking, such as guarding or lock-out. A safety-trained person must read every instruction before it is issued.

3. Supplier emails that get an answer

Chasing late material, querying a price, rejecting a delivery: a purchasing person may write fifteen of these a week, each one slightly different and each easy to put off. The assistant drafts from the purchase order line, so each email carries the facts.

Draft a short, firm, polite email to a supplier. Facts:
PO 4471, line 3: 40 sheets 2 mm mild steel, 1250 x 2500
Promised delivery: 12 November. Today: 14 November. Not received.
This material is for customer job J-1182 due 28 November.
Ask for: a confirmed delivery date by end of tomorrow, and whether a
partial shipment is possible. Mention the knock-on to our job without
blaming. Under 100 words. Sign off [your name].

Illustrative result: "Hi [first name], PO 4471 line 3 (40 sheets of 2 mm mild steel) was due on 12 November and has not arrived. The material is for a job due on 28 November, so a confirmed date by end of tomorrow would help us plan. Could you also tell me whether a partial shipment is possible? Thanks, [your name]". It states the facts, makes one request with a deadline and leaves out blame.

The quick sum: 14 chasers a week at six minutes each is 84 minutes. At two minutes each with a draft, it is 28. That saves under an hour a week, but the greater gain is that the emails go out on the day the delivery is late, not three days later. Keep a log of which suppliers needed chasing and feed it into a supplier scorecard; the chase sheet sets up the tracking behind it.

Start: take this week's three latest deliveries and draft the chasers. Watch for: dates and quantities. Check every number against the purchase order before sending. Never let the assistant assume a PO number from context.

4. ISO documents drafted from what you actually do

If you hold or are working towards ISO 9001, the quality management standard, you keep documented information that matches how you work and evidence that you evaluate the system, which in practice means procedures, a document register, internal audit records and management review notes. ISO published a new edition, ISO 9001:2026, in September 2026 and withdrew the 2015 edition; if you are certified to the old one, ask your certification body about the transition. Small shops often have a template pack bought years ago that nobody follows. The assistant can help turn real practice into text, which is a better starting point than a generic template.

• Procedures: give it your notes on how goods-in inspection really works and ask for a procedure with purpose, scope, responsibilities, steps and records kept. Ask it to list gaps between your notes and the steps.
• Internal audit checklist: paste a procedure and ask for ten audit questions that check the practice and the record, not just the wording.
• Management review minutes: give it your bullet notes from the meeting and ask for a clean record in a fixed layout.
• Document register: paste a folder listing and ask for a draft list with document number, title, owner and revision, flagging duplicates.

An illustrative case: the shop had a goods-in procedure that said all material is inspected on arrival. Practice was that only certificated material for aerospace customers was checked. The assistant, given the real notes, wrote a procedure with two routes, and listed a gap: no record existed for the second route. That gap, not the neat prose, was the useful output. A certification body will compare documents with what staff say and what records show, so document what happens, then improve it.

Start: one procedure, the one your last audit found weakest. Watch for: a polished document that nobody follows, which is worse than a plain one that matches the floor. Have the person who does the job read it and mark every sentence that is not true. The assistant helps with the writing; the audit trail remains yours, and it is wise to ask your certification body or consultant about their expectations for documents drafted with AI.

5. A scheduling sheet that flags overload before Monday

A small shop schedules in a spreadsheet: jobs down the side, machines across the top, a due date for each. The weak point is capacity. A job is booked because the due date is free, and the machine is already full that day. A model is not a scheduling engine, but it is good at reading a sheet and saying "this looks wrong". Let the sheet total the hours, and let the assistant read it.

In the sheet, a SUMIFS formula gives laser load for Monday as 3.5 + 2.8 + 3.1 = 9.4 hours. The laser has 7.5 productive hours a day (8 hours less a 30-minute allowance for setup and checks). The prompt below asks for the reading, with the sums supplied.

Here is next week's schedule as a table, with a column for daily capacity
per machine (laser 7.5 h, press brake 7.0 h) and a column of daily
load that my spreadsheet calculated. Identify every machine-day where load
exceeds capacity. For each, suggest which job to move and to which day,
without missing its due date. Use only the loads in the sheet; do not
recalculate them. Say if no move is possible.

Illustrative reply: "Laser, Monday: load 9.4 h against 7.5 h, over by 1.9 h. Moving J-1188 (3.1 h, due Thursday) to Tuesday brings Monday to 6.3 h and Tuesday to 4.7 h, and still meets Thursday." The sums check: 9.4 - 3.1 = 6.3 and 1.6 + 3.1 = 4.7. The assistant spots the clash and proposes a move, and the sheet confirms the arithmetic. Link this with a work-in-progress view, as described in the job shop WIP tracking tutorial.

Start: next week's real schedule, with capacities written in a column. Watch for: hidden constraints the sheet does not show, such as an operator on leave, a tool that is on order or a customer who needs a call before a change. The assistant only knows what the table says.

6. Complaint triage that puts safety first

Complaints arrive by email, phone notes and delivery forms. They get answered in the order they are noticed, and the patterns go unseen. Triage means giving each a category, a severity and an owner, and drafting a first reply, so the right person acts on the right one first.

Classify each complaint below. For each give: category (dimension,
finish, damage in transit, wrong item or quantity, documentation,
other), severity (1 = safety or line-stopping, 2 = unusable parts,
3 = cosmetic or minor), suggested owner (quality, dispatch, sales) and
a one-line reason. If the text mentions injury, fire, a failure in use,
or a regulatory issue, mark severity 1 and write "ESCALATE TO A PERSON".
Then draft a first reply of under 80 words that thanks the customer,
confirms what we understood, and says who will contact them and when.
Do not promise a refund or replacement.

Illustrative output across twelve complaints in a week: three finish defects (severity 3), four dimension issues (two severity 2), two transit damage, one wrong quantity, one documentation, and one flagged for escalation because the customer mentioned a bracket cracking in use. The escalation line is the one that matters. A human sees it first, that day.

Count categories weekly. After eight weeks you will know whether "finish" is a one-off or the most common fault, which feeds a corrective action, sent to the supplier when the cause lies there. The complaints and defects pattern tutorial shows how to read the counts. Once a complaint becomes a return, your returns (RMA) process picks it up.

Start: last month's complaints, already resolved, so you can compare the assistant's sorting with what you did. Watch for: severity placed too low. Read every severity 1 and 2, and a sample of 3s. If the assistant ever rates a safety complaint as cosmetic, add the words that triggered the miss to the prompt.

A quote that went out wrong, and why

A small fabricator let an assistant draft a quote from an enquiry and, to save time, asked it to include a price using a rate card pasted into the chat. The reply looked right. It had applied the per-minute laser rate to hours, and the quote went out about 40% under cost on a 500-part order. The customer accepted. The error was found at invoicing, and the shop absorbed the loss. The fix was a rule: the assistant extracts and drafts; the spreadsheet prices; a second person looks at any quote above a set value. The rule cost one extra step per quote and removed a whole class of error. (Illustrative.)

Two other shops, the same six jobs

The jobs travel. A wholesale distributor that assembles kits from bought-in parts has no laser and no press brake, yet four of the six apply almost unchanged. Its quote drafts read a customer's parts list and flag items it does not stock. Its supplier emails chase late inbound pallets, using the same facts-first prompt. Its complaint triage sorts "wrong item picked" from "damaged in transit", which sends the first to the warehouse lead and the second to the carrier claim. Only the work instructions and the ISO documents need rethinking, and the instructions become pick-and-pack sheets for the kit builds.

A B2B software reseller with a small hardware assembly bench, say a firm that configures and ships pre-loaded devices, finds a different mix. Its quotes are the longest job, because each enquiry mixes licences, devices and support. It uses the extraction prompt to produce one table per enquiry, and the sheet to price. Its work instruction is a build checklist for each device model, with a check that each unit boots, updates and is wiped of test data. Complaint triage matters most for "arrived dead" and "wrong licence", where the owner differs. The point is not to copy this list. It is to see that the same pattern, read, structure, draft, review, fits the paperwork of almost any small maker or assembler.

Before and after: the same quote reply

A before-and-after shows what review adds. The first draft an assistant produced for the bracket enquiry, and the version the estimator sent:

All three changes remove an assumption. That is what the reviewer is for. Over a few weeks, add each recurring invention to the prompt as a rule ("never state a colour code the customer did not give") and the first drafts improve.

A thirty-day order of work

1. Days 1 to 3: decide the review rule. Who reads what before it leaves, and what is never uploaded. Write it on one page.
2. Week 1: supplier emails. Draft this week's chasers, send them after checking the numbers, and log the replies.
3. Week 2: quote drafts on ten past enquiries, then live ones with the estimator reading every table.
4. Week 3: complaint triage on last month's complaints, then on new ones, with a person reading every severity 1 and 2.
5. Week 4: one work instruction and one scheduling check. Leave the ISO documents for month two, when your review habit is steady.

At the end of thirty days, check the one number per job set out further down this page. Keep what saved time without adding rework, change the prompt on what produced edits, and drop anything that did neither. A shop that does this will have two or three jobs it relies on, and a clear idea of the next. It is also the right moment to ask whether any of them should become an automation, so a draft appears when an enquiry arrives instead of when someone pastes it.

Who reads the output, and how long it takes

Each job needs a named reader. Quotes go to the estimator, instructions to the operator who does the job and then the supervisor, supplier emails to the buyer, ISO documents to the quality lead, schedules to the planner and complaints to whoever answers the customer. A rough time budget helps people accept the extra step: two to three minutes of reading for an email or a quote, ten for a work instruction, fifteen for a procedure. If a reader finds themselves rewriting more than a third of a draft, the prompt needs work, not the reader.

Measure it in a way that survives scrutiny

Pick one number per application and track it for six weeks, before and after. Keep it modest:

• Quotes: minutes from enquiry to reply sent, and percentage of enquiries answered within one working day.
• Work instructions: number written, and how many times a new starter asked a question the page should have answered.
• Supplier emails: late deliveries chased on the day they were late, as a share of all late deliveries.
• ISO documents: hours spent preparing for the last audit against this one, and findings that related to documents.
• Scheduling: days in the week where load exceeded capacity, counted from the sheet.
• Complaints: time from receipt to first reply, and severity 1 items seen by a person within the hour.

Resist converting these into money at first. A measured fall in reply time is a fact. A claimed saving of $400 a month usually involves assumptions you cannot defend. If you do want a figure, the AI ROI worked example shows a sober way to do it.

What this list leaves out on purpose

Predictive maintenance, camera-based inspection and machine data analytics are not here, because they need sensors, cameras or machine connections, and cost accordingly. If you are curious what they involve, the visual inspection cost breakdown gives a realistic picture of one of them, and analysing a downtime log you may already keep is a halfway point that still needs no sensors. Start with the six here, get used to reviewing AI output, and then decide whether anything heavier is worth the money. If you would rather map the order of those steps with someone, that is what my AI implementation consultation covers; get in touch if so.

Questions from small factory owners

Which of the six should a small factory start with?

Start with supplier emails or quote drafts. Both begin with text you already have, the output is read by you before it is sent, and a mistake costs a rewrite and not a scrapped batch. Leave scheduling and ISO documents for the second month, when you trust your review habits.

Do I need to upload customer drawings to an AI tool?

Usually not. Most of these jobs work from the text of an enquiry, a part list or your notes. If a drawing is covered by a confidentiality agreement, read your agreement and your AI plan's data terms first, or type the few dimensions you need into the prompt.

Can AI price a quote on its own?

It should not. Keep prices in a spreadsheet with your machine rates and material costs, and let the assistant read the enquiry and draft the covering email. Models make arithmetic and rate mistakes, and a wrong price on 500 parts costs far more than the drafting saved.

Will AI-written procedures pass an ISO audit?

Only if they describe what your people actually do. Auditors compare documents with practice and with records. Use AI to draft from your notes and observation, then have the operator who does the job read it and mark what is wrong before it enters document control.

Further reads

• AI Quality Control for Small Manufacturers: Where to Start — Quality checks beyond the office side.
• How to Clean Messy Data in Excel: Step-by-Step Guide — Tidy the spreadsheets these jobs feed on.
• Shift Handover Template: Let AI Turn Operator Notes Into a Report — Turn operator notes into a shift report.
• Daily Production Report: Summarise Counts, Scrap and Stops With AI — Summarise counts, scrap and stops each day.
• Stock Count Variance: Find Why Your Count and System Disagree — Find why your count and system disagree.
• Proforma Invoice vs Purchase Order: Let AI Catch the Mismatches — Catch mismatches between orders and invoices.
• Supplier Corrective Action Request: Draft a SCAR With AI — Write the request when a supplier caused the fault.
• AI for Industrial Manufacturing: Read Downtime Logs to Find Repeat Stops — Read a downtime log you already keep, still without sensors.
• AI Systems for Manufacturing: MRP Add-On or Chatbot on Your Own Data? — Compare native MRP AI replenishment tools against desktop chatbots running over CSV exports, with BOM thresholds and shop-floor tests.
• AI in Plant Engineering: Breakdown Logs and Maintenance Plans — Turn a messy breakdown log into ranked repeat faults and a draft preventive maintenance plan, with the safety items left to a named engineer.
• Manufacturing Cost Estimating: How AI Helps Price a New Product — One new steel product costed end to end with AI help: materials, labour minutes, overhead, scrap, tooling, margin and what moves the price.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI and Anthropic help pages on model training for business and personal plans; ISO's ISO 9001:2026 and ISO 9001:2015 pages (checked November 2026).