Where Should a Small Manufacturer Start With AI?

Start with the paperwork around production, not the machines: quoting from customer enquiries and drawings, supplier and purchasing admin, and turning know-how in people's heads into written work instructions. Pick one, run it for four to six weeks with a general assistant or the AI in software you already have, and measure the hours saved before spending on inspection cameras or sensors.

The reason is cost and data. Machine-level AI, such as camera inspection or predictive maintenance, needs sensors, clean historical data and an integrator, and it's priced in thousands. Office-side AI costs about $20-$30 a user a month and works on documents you already have. The exception: if scrap and defects are your biggest cost, inspection may be the right first project, and the scoring below will show it.

Follow me on Instagram@sagnikteaches

Where a small manufacturer can start, compared

Score your candidates before you choose

Rather than picking the most interesting idea, score each candidate from 1 to 5 on four things: hours it takes a week, how repetitive it is, how ready the data is, and how survivable a mistake would be. Here's a filled-in example for an illustrative 30-person precision engineering shop:

Connect on LinkedInSagnik Bhattacharya

RFQ triage wins narrowly, supplier chasing is a close second and an easy follow-on. Inspection scores high on hours and repetition but low on data, which usually means "not yet" rather than "never". Your own table may look very different, and that's the point of doing it.

Subscribe on YouTube@codingliquids

RFQ triage: from a customer email to a quote-ready spec

Estimators lose a surprising amount of time just working out what a customer is asking for. An assistant is good at that part. It is not good at costing machine time, which should come from your costing sheet. A prompt to adapt:

You help the estimator at a precision engineering company.
From the customer email and drawing notes below, extract:
part number and revision, quantities and quantity breaks,
material and thickness, finish, tolerances called out,
required lead time, certificates or inspection reports asked for.
Then list every question we must ask before quoting, and anything
that looks inconsistent between the email and the drawing.
Do not estimate prices or machine times.
EMAIL: [paste]
DRAWING NOTES / TITLE BLOCK TEXT: [paste]

For an email asking to "quote 250 off and 1,000 off of the attached bracket, 3 mm mild steel, zinc plated, first batch in 4 weeks, drawing rev C attached, also do you do powder coat?", an illustrative output:

Part: bracket, drawing number BR-2231
Revision: email says rev C; title block says rev D - CHECK
Quantities: 250 and 1,000
Material: 3 mm mild steel (grade not stated in email; drawing
note 4 gives grade)
Finish: zinc plated (thickness not specified)
Tolerances: +/-0.1 mm on hole centres (drawing note 2)
Lead time: first batch in 4 weeks
Certificates: none mentioned
Questions for the customer:
1. Please confirm revision C or D.
2. Zinc plating thickness and any passivate colour required?
3. Is powder coat instead of, or as well as, plating? Colour?
4. Delivery split for the 1,000 off, or all at once?
5. Material certificates required?

The revision mismatch is the valuable catch; quoting the wrong revision is an expensive mistake, and on a 1,000-off order it can mean a batch made to a superseded drawing. What you'd still fix: the AI assumed plating is done in-house because the email didn't say otherwise. If you subcontract plating, that lead time goes into the four-week question. The general approach for one-off and custom work is in quoting custom jobs faster with AI in a small workshop.

Work instructions from the people who know the job

Many small manufacturers depend on two or three people who know how a changeover or a tricky setup really works, and it's written down nowhere. AI makes capturing that much less painful. Ask the setter to talk through the job while someone records it on a phone, then have an assistant turn the transcript into a draft.

A snippet of what the setter says:

Right, first thing, isolate the machine and lock it off. Swap to the soft jaws, they're in the blue drawer, the ones marked B. Torque them up, don't overdo it or you'll mark the part. Load the program, it's the one ending 07. First part always goes to the inspection bench before you run the rest.

And the illustrative draft it becomes:

Changeover: bracket BR-2231, lathe 2
1. Isolate the machine and apply your lock-off. (SAFETY)
2. Fit soft jaws set B (blue drawer).
3. Tighten jaws to [TORQUE VALUE - setter to confirm].
   Over-tightening marks the part.
4. Load program ending 07. Check the program number on screen
   matches the job card.
5. Run one part. Take it to the inspection bench.
   Do not run the batch until first-off is signed. (CHECK)

Notice what it did well and what it couldn't: it structured the steps and flagged the safety and check points, but "torque them up" has no number, so it left a placeholder instead of inventing one. That's the behaviour to insist on in your prompt. The setter then reviews, adds photos, and it goes in the folder. A fuller method is in writing SOPs with AI from rough notes.

Supplier chasing and purchase paperwork

Buyers in small manufacturers spend hours on chasers, comparisons and invoice queries. Paste your open purchase orders with due dates into an assistant and ask for a chaser per supplier, grouped, polite, specific about the order line and date. Compare three supplier quotes laid out in different formats as one table. Flag supplier invoices where the price or quantity doesn't match the order. Each is covered in detail in comparing supplier quotes side by side with AI and matching supplier invoices to purchase orders automatically.

A chaser drafted from three overdue lines on the open-orders list, as an illustrative example:

Hi, three lines on our orders with you are now past their promised dates: PO 8812 line 2 (40 x 25 mm bright bar, 6 m lengths, due 18 Sep), PO 8830 line 1 (M8 flange nuts, 2,000, due 22 Sep) and PO 8841 line 4 (saw blades, 5, due 24 Sep). Could you confirm new dispatch dates by Friday? The bright bar is holding up a customer order, so that one matters most.

The value is in the specifics and the priority. A vague "any update on our orders?" gets a vague answer.

Supplier quotes are the other time sink. Three quotes for the same steel arrive as a PDF, an email body and a spreadsheet, each with different units and delivery terms. Ask the assistant to lay them out as one table and to state every assumption. An illustrative result:

For a 60 m order, Supplier B's cheaper metre price disappears: the 100 m minimum and the delivery charge make it $685 for material you mostly don't need yet, against $414 from A. The table makes that visible in seconds, but check the conversions yourself. In one illustrative case, an assistant read "$6.55" from Supplier C's quote as per metre when the line actually said per kilogram, which would have made C look like the cheapest by far. Ask for the original unit to be shown next to every converted figure, as in the second column above.

This is often the easiest second project, because mistakes are cheap to spot: a chaser with the wrong date gets a reply saying so.

The same approach covers the customer side. "Where's my order?" emails take minutes each to answer because someone has to look up the job, check with the shop floor and write a reply. Export the open-orders list each morning and paste it in with the day's queries, and the assistant drafts every reply at once. Before, a typical answer read "It's in production, should be with you soon." After, the draft says: "Your 250 brackets (order 4471) finished machining yesterday and are at the plater; we expect them back on Thursday and will dispatch Friday for Monday delivery." It is only as accurate as the list you paste, so the list has to be the one the production manager trusts.

Inspection cameras and predictive maintenance: when, not if

Machine-level AI makes sense when three things are true: the problem is expensive (scrap, returns, unplanned downtime), it repeats often enough to learn from, and you can collect the data, such as hundreds of labelled images of good and bad parts or months of sensor readings including real failures. Many small shops meet the first condition but not the third, which is why these projects stall.

If inspection scored well on your table apart from data, start collecting images now, even before you buy anything. When you're ready, AI quality control for small manufacturers explains how to begin, and what AI visual inspection costs helps you budget it.

A six-week pilot in a 30-person engineering shop

Back to the imaginary engineering shop from the scoring table, and its winning candidate. The shop receives about 25 RFQs a week. The estimator spends about 45 minutes on each, nearly 19 hours a week, and quotes take about three working days to go out.

• Week 1. Collect 20 past RFQs with the quotes that went out. Write the prompt above. Decide who owns it: the estimator, with the production manager as sponsor.
• Week 2. Run the prompt on the 20 past RFQs and compare with what the estimator actually extracted. Fix the prompt where it missed things (it initially ignored drawing notes printed in small text in the title block).
• Weeks 3-4. Live use on every new RFQ. The estimator reviews each extraction in about five minutes, sends questions to the customer the same day, and costs the job on the usual sheet.
• Weeks 5-6. Measure. Time per RFQ is down to about 25 minutes, about 10 hours a week, saving roughly 8 hours. Quotes go out in about one working day. Two revision mismatches were caught that would previously have been quoted wrongly.

Cost: two seats on a business assistant plan at about $25 a user a month on monthly billing (both ChatGPT Business and Claude Team need a minimum of two seats), or Microsoft 365 Copilot Business at $21 a user a month on annual billing if the shop already runs Microsoft 365. The win rate needs a full quarter to judge, so keep tracking it.

Telling the shop floor what the pilot is, and isn't

In a 30-person firm, word that "the office is bringing in AI" travels fast and usually arrives as "they're automating jobs". A short, specific note before the pilot starts heads that off. An illustrative version from the production manager:

For the next six weeks, the estimating office is trying an AI assistant to read customer enquiries and drawings faster, so quotes go out in a day instead of three. It doesn't touch the machines, the schedule or anyone's job. If it works, the next thing we'd look at is writing down changeover instructions, and for that we'd want the setters' help, because you know how the jobs really run. Questions to me any time.

It names the task, the time limit, what's out of scope and where the shop floor comes in. The setters who later record changeover explanations are far more willing when they heard about it first, and from someone they work with.

Customer drawings, confidentiality and contracts

Customer drawings are often covered by confidentiality agreements, and some customer terms restrict sending their data to third-party services. Before uploading anything:

• Read your NDAs and customer terms for clauses about third-party processing or cloud services.
• Use a business plan. Business tiers such as ChatGPT Business, Claude Team, Microsoft 365 Copilot and Gemini in Workspace don't train on your business content by default. Consumer plans need the model-training setting switched off.
• Keep defence, aerospace and export-controlled work out until you've checked the contract and any control requirements with the customer or your adviser. Some of that work can't go to general cloud tools at all.
• Paste the text you need, not the whole pack. For RFQ triage, the email and the title-block notes are usually enough.

Signs your first manufacturing AI project worked

• A measured time saving on the task, in hours a week, from before-and-after timings rather than impressions.
• Fewer errors of the kind the project targeted, such as wrong revisions quoted or late supplier deliveries noticed too late.
• The person who owns the task still uses it after six weeks without being asked.
• A second candidate from your scoring table is now the obvious next step.

The timing doesn't need software. For two weeks before the pilot and two weeks during it, the estimator notes the start and finish time of each RFQ on a sheet by the screen:

RFQ    Received  Started  Finished  Minutes  Sent to customer  Notes
Q-311  Mon 09:10 Mon 10:05 Mon 10:52   47     Wed               drawing rev query
Q-312  Mon 11:30 Tue 08:40 Tue 09:20   40     Wed
Q-327  Mon 09:25 Mon 09:40 Mon 10:03   23     Mon  (pilot)      rev mismatch caught
Q-328  Mon 10:15 Mon 10:20 Mon 10:47   27     Tue  (pilot)

Twenty rows of that are more convincing to a sceptical production manager than any estimate, and they show the second benefit, time to reply, which the customer notices first.

If the owner has quietly gone back to the old way, find out why before starting anything else. It's usually one missing piece of data or one output that needs too much fixing, and both are fixable.

Manufacturers' follow-up questions

Do we need an ERP with AI built in before starting?

No. The first projects described here work with a general assistant and the documents you already have. If your ERP or MRP system already offers AI features, test them on the same job and compare. Buying a new system for its AI features alone is rarely the right first move for a small manufacturer.

Who on the shop floor should own the first AI project?

Someone who does the task every week and is irritated by it, such as the estimator for quoting or the buyer for supplier chasing, with the owner or production manager as sponsor. Projects owned by whoever is keenest on technology but does not do the task tend to stall once the novelty wears off.

How much should a first AI project cost a small manufacturer?

For office-side work, the main costs are assistant seats, often around $20-$30 per user a month, and the staff time to set it up and review outputs, typically a few days spread over the first month. Inspection cameras and sensor projects are a different scale, usually needing quotes from integrators before you can budget them.

Further reads

• AI Supplier Management: Track Prices, Lead Times, and Risk — Track supplier prices, lead times and risk once chasing works.
• How to Create SOPs From Screen Recordings With AI — Turn screen recordings into procedures for office systems.
• How to Use AI to Set Reorder Points and Prevent Stockouts — Set reorder points for raw materials and consumables.
• Are Chat-With-PDF Tools Safe for Contracts and Client Files? — Whether chat-with-PDF tools are safe for customer drawings.
• How to Build a Job Costing Sheet in Excel With AI Help — Build the costing sheet your AI-drafted quotes price from.
• Does Microsoft 365 Copilot Keep Your Business Data Private? — How Copilot handles business data if you're on Microsoft 365.
• How to Run Your First AI Pilot Project in a Small Business — A general structure for any first AI pilot.
• Predictive Maintenance With AI: Realistic for a Small Workshop? — When AI predictive maintenance pays in a small workshop, the five cheaper rungs below it, and a CNC shop costed end to end.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: vendor pages for assistant plan prices (ChatGPT Business, Claude Team, Microsoft 365 Copilot Business), checked September 2026.