AI for Manufacturing (2026): What a $2M to $50M Shop Can Actually Use This Quarter, Before Anyone Says "Smart Factory"

AI for manufacturing: the short answer for a small shop

There are two "AI for manufacturing" conversations and the search results only have one of them. The plant-floor one is predictive maintenance, machine vision, and a scheduling engine: a capital project, with sensors and a vendor. The office one is a $20-to-$25-a-month business seat doing the paperwork around your shop: quotes drafted from the RFQ, order-status emails answered, work instructions written and translated, maintenance and quality logs summarized, suppliers chased. As of September 2026, a shop under about $50 million in revenue gets its first and cheapest return from the office one, and it can start this quarter. Here is that version, from someone who installs it for owner-led businesses, with the two things the assistant must never touch and what it costs.

Why the vendor answer is not your answer yet

Ask an AI assistant which tools a small manufacturer should use and, as of this week, it names MRPeasy, Katana, Odoo, and MachineMetrics. Those are real products for real problems: inventory, MRP, and machine data. They are also a second decision, not a first one. A ten-to-fifty-person shop that has never used an assistant on its own paperwork is not ready to evaluate a platform, because it does not yet know which hours it is trying to buy back.

The paperwork is where the hours are. An estimator who spends half of every day formatting quotes. An office manager who answers "where is my order" forty times a week. A shop lead who writes work instructions once, in English, and then explains them verbally to a crew that reads Spanish. A maintenance log that lives on a clipboard and gets read after the breakdown. None of that needs a sensor. All of it is drafting, summarizing, translating, and sorting, which is exactly what a general assistant does well.

The five office jobs a $20 seat does for a manufacturer

1. Quotes and estimates, drafted from the RFQ

Paste the request for quote, the customer's history, and your pricing rules and format. The assistant returns a first-draft quote in your template, a list of what the RFQ left ambiguous, and the questions to send back before you price it. The estimator still sets the price, the material buy, and the delivery promise. What they stop doing is the forty minutes of assembling the document.

A worked example, not a statistic: a shop that turns 20 quotes a week and saves 30 minutes of drafting per quote recovers about 10 estimator hours a week. Your quote count and your drafting time decide your number, which is why the first thing I ask a shop for is last month's RFQ folder and a stopwatch.

2. Order status and customer replies

"Where is my order," "can you move the ship date," "resend the cert." Give the assistant your order-status template and a copy of the schedule, and it drafts the reply; a person reads it and sends it. The rule is the same one I install everywhere: the assistant drafts, a human sends. That one rule is what keeps a wrong ship date from reaching a customer.

3. Work instructions, SOPs, and the Spanish version

Record the shop lead explaining the setup once, transcribe it, and have the assistant turn it into a numbered work instruction with the tooling list and the checks. Then ask for the same document in Spanish, or whichever language the floor reads. The lead reviews both. For a lot of shops this is the single highest-value job on the list, because the instructions finally exist in writing and in the crew's language, and training a new hire stops depending on one person's memory.

4. Maintenance and quality logs someone can read

Paste a month of maintenance entries or nonconformance notes and ask for what repeated, what is overdue, and which machine shows up most. This is not predictive maintenance; it is reading your own log, which is the step most shops skip. When the same spindle appears four times, you have a maintenance decision with a number attached, and that number is what a floor vendor will eventually price against.

5. Supplier follow-up and the purchasing inbox

Open purchase orders, late confirmations, price changes buried in a supplier's email. The assistant drafts the follow-ups from your PO list and summarizes the inbound thread so the buyer sees the change in price or lead time without reading twelve messages. Same rule: a person approves before anything is sent.

The two things it never touches

• The price and the promise. Margin, material buys, and delivery dates are the estimator's and the owner's decisions. The assistant drafts around them; it does not make them. A model will happily produce a plausible price, which is exactly why it must never be asked to.
• The floor. Nothing on this list controls a machine, releases a job, or signs off an inspection. Those are controlled processes with their own systems and their own approvals. If a vendor tells you a chat assistant should be in that loop, ask them to put it in writing.

The plan decides where your drawings go

This is the decision most shops make by accident, when someone signs up for a consumer account to try it. As of September 2026, the business plans state that customer data is not used for training by default, and they give the owner an admin console to add and remove people:

• Claude Team: $20 per seat per month billed annually, $25 monthly, for 2 to 150 seats. No training on Team data by default. Details on Claude Team plan.
• ChatGPT Business: Standard seats at $20 per user per month annually, $25 monthly, 2-seat minimum. No training on Business data by default.
• Microsoft 365 Copilot Business: $21 per user per month on an annual subscription, added to a Microsoft 365 business plan, for shops whose office already lives in Outlook and Excel.

Then write the never-paste list, and write it before the first seat is assigned: customer drawings under an NDA, anything ITAR or export-controlled, your pricing files, and employee data. Most quote drafting works from the RFQ's text and your rules without the CAD file ever leaving your server. If a customer contract restricts where their data can go, that contract wins over every vendor's terms, and your assistant policy should say so. The one-page template I hand clients is on the AI acceptable use policy.

Which job first

Pick by hours times repetition times low judgment, which is the same scoring I use for every business in what to automate first. For most shops that puts quoting or order status at the top, because both happen every day and both are drafting, and puts work instructions second, because the payoff is large but the volume is lower. Maintenance-log reading is the one to run monthly from day one, because it produces the number that decides whether the plant-floor conversation is worth having.

What it costs, as of September 2026

Five office seats on annual billing are about $100 to $125 a month before setup. The whole bill, seats plus usage plus setup, with worked budgets for teams of 1, 10, and 25, is on how much does AI cost.

What to do this week

1. Time the quotes. Pull last month's RFQs, count them, and time how long the drafting takes on the next five. That number is your business case, and it takes an afternoon.
2. Write the never-paste list. One page: what goes into the assistant, what never does, and who approves what goes out. Do it with whoever owns your customer NDAs.
3. Buy business seats for the office people who will use it daily, give them the quote and order-status templates, and run it for thirty days before anyone proposes a platform.

If you would rather have the job list, the order, and the policy written for your shop before you spend anything on software, that is the AI Readiness Assessment: $2,500 flat, two weeks, a written roadmap, and the fee becomes your deposit if we build it together. The smart-factory conversation is real. It is also the second conversation, and the first one costs about the price of a lunch per seat.