josecustom.ai josecustom.ai Book

AI for Small Manufacturers: Where to Start and What to Keep Out

Where AI pays off first in a small manufacturing company, why the front office usually beats the shop floor, what data must stay out of public tools, and how to test a quoting pilot on jobs you have already quoted.

For a small manufacturer, AI is most useful right now as help with the paperwork around production: reading RFQs, drafting quotes, turning experienced people’s knowledge into written work instructions, and making old records searchable. Machine vision and predictive maintenance can also pay off, but they need clean data, some hardware, and more setup. For most shops under 100 people, the office is the better place to start.

The other thing to settle early is what stays out. A manufacturer’s most sensitive data is often not personal information. It is a customer’s drawings, a controlled specification, or your own process parameters. That changes which tools you can use.

Where does AI pay off first in a small manufacturing company?

Start where people are reading and typing, because those jobs need no sensors or machine integration, and a person can check every output.

RFQ intake and quoting

An RFQ arrives as an email with a PDF, a drawing, and a list of notes. Someone has to pull out material, quantity, tolerances, finishes, certifications, and delivery date before they can even begin estimating.

AI handles that extraction well. It can also compare a request against your past jobs, flag requirements that are missing or unusual, and draft the quote letter. The estimator still sets the price and owns the decision. AI should not invent cycle times or material costs. It should pull those from your own rate sheets and job history, and show where each number came from.

Work instructions and training material

Many shops depend on what a few senior people know, and some of it has never been written down. Record a setup walkthrough on a phone, transcribe it, and have AI turn it into a structured draft. The person who does the job corrects it, and it goes through your normal document control. That is faster than asking your best machinist to write a manual, which usually never happens.

Quality paperwork

Nonconformance reports, corrective action records, and supplier follow-ups follow a pattern, which makes them good candidates for first drafts. AI can also search your past reports for similar problems before someone starts from scratch. Root cause still belongs to the people who looked at the part. If you are certified to a standard such as ISO 9001 or AS9100, AI drafts are subject to the same review and record rules as anything else.

Finding things in your own records

“Did we ever run this alloy for that customer?” “What fixed this alarm last time?” A private search tool over approved job records, manuals, and maintenance logs can answer these questions in seconds, with a link back to the source document. It needs careful permissions, because a tool that can read everything will show everything to anyone who asks.

Purchasing and supplier email

Matching a purchase order to a packing slip and an invoice, chasing confirmations, and flagging a changed delivery date are routine and rule-based. Keep the approval of orders and payments with a person. Any tool that can act on email deserves the approval and permission checks in an AI security audit.

What about predictive maintenance and machine vision?

Both are real, and both have prerequisites that small shops should check before buying.

Predictive maintenance needs machine data over time and a history of actual failures. If you do not capture vibration, temperature, current, or controller data today, your first project is collecting it. If a machine rarely fails, there are few examples to learn from. Basic condition monitoring with sensible alert thresholds often gets you most of the early value, and it is a reasonable first step either way.

Visual inspection works best on repeatable parts, under controlled lighting, with a clear definition of a defect and enough labeled images of good and bad parts. If your defects are rare, you will have few examples, and if you run high-mix work with short batches, each new part may need its own setup. Ask any vendor to test on your parts, under your lighting, and to report false rejects and missed defects separately. A missed defect and a scrapped good part cost you different amounts.

A simple test: can someone show you the data today, in one place, for one machine or one part? If not, the data work comes first.

What must stay out of public AI tools?

Before anyone uploads a file, sort your data into groups. This is the same habit I describe in my AI vendor risk assessment: decide what the tool will see, then judge the tool against that.

Customer drawings and specifications. These usually come with NDA or contract terms limiting who may receive them. Sending them to an AI vendor can be a disclosure to a third party, whatever that vendor does with the data. Read the terms, and ask the customer if the contract is unclear.

Export-controlled technical data. If you handle drawings or data controlled under ITAR or the EAR, public AI tools should be off-limits unless your export compliance owner has reviewed the specific tool and setup. Where data is stored and who can access it both matter. A general assurance from a vendor’s marketing page does not answer that.

Controlled Unclassified Information. Defense suppliers covered by DFARS 252.204-7012 have specific safeguarding and cloud requirements for this information. If you are working toward CMMC, a tool that handles CUI becomes part of what you have to account for. Bring the AI question to whoever owns your compliance program before anyone tries the tool.

Your own trade secrets. Process parameters, formulations, supplier pricing, and margins may not be regulated, but they are what makes your company hard to copy. Trade secret protection generally depends on your taking reasonable steps to keep them secret, so uploading them to personal accounts works against you.

The practical risk is ordinary people trying to be helpful. An estimator uses a personal ChatGPT account to read a difficult drawing, and nobody knows. Give people an approved tool and a short written rule, and walk them through it. My posts on an AI acceptable use policy and training employees on AI cover both.

If you have controlled or highly sensitive data and still want these capabilities, a private AI deployment in your own cloud tenant may be an option, but it does not automatically satisfy ITAR, DFARS, or your customer contracts. The configuration still has to be checked against those requirements.

What should AI not decide?

Keep a person responsible for the following, even when AI does the drafting:

  • Final quotes and commitments. A wrong delivery date or a missed certification requirement becomes your obligation.
  • Part acceptance and release. AI can flag a part. A qualified person or a validated process decides.
  • Safety procedures. Lockout steps, guarding, and chemical handling instructions need review by someone who knows the machine and the rules that apply. A confident, wrong sentence here can hurt someone.
  • Machine control. A suggestion on a screen is very different from software that changes a live program or parameter. Keep them apart unless you have engineered and validated that link.

How do you test AI on quoting without risking a real job?

Test against quotes you have already finished, since you already know the right answer.

  1. Pick 20 closed RFQs. Mix easy and difficult ones, and include a few you lost or that went badly. Leave out anything controlled or restricted by contract.
  2. Use an approved tool. Run the test in a business plan whose data terms you have checked, not a personal account.
  3. Run extraction only. Have the tool pull material, quantity, tolerances, finishes, certifications, and dates from each request.
  4. Score it against your own records. For each field, mark it correct, wrong, or missing. Treat a wrong tolerance or a missed certification as a failure, however good the rest looks.
  5. Time it. Record how long intake takes by hand, and how long it takes with AI when someone checks every field.
  6. Do the math with your numbers. Minutes saved per quote, multiplied by quotes per month, multiplied by your loaded estimator cost. Compare that with the cost of the tool and the review time.

If the tool misses critical requirements, you learned that without risking a job. If it holds up, move to live RFQs with an estimator checking every output, and stay there until you have a record you trust. Then do the same with the next workflow. My guide to implementing AI covers that step-by-step approach.

Where can a small manufacturer get help for free or at low cost?

Before you pay anyone, including me, call your state’s Manufacturing Extension Partnership (MEP) center. The MEP National Network is a NIST-supported program with centers across the US, and it says its centers provide AI readiness assessments, training resources, and process improvement help for small and mid-sized manufacturers. Services and pricing vary by center, so ask yours what is available. NIST also publishes material on industrial AI for smaller manufacturers.

Your existing software vendors are another good call. ERP, CAM, CMMS, and quality systems may already offer AI features. Run them through the same data checks, and turn on only what you need.

Frequently asked questions

How can a small manufacturer use AI?

The most practical first uses are RFQ extraction, quote drafting, work instructions, quality report drafts, and searching internal records. These require no new hardware, and a person can review every output. Predictive maintenance and machine vision can follow once the necessary data is being collected.

Can manufacturers put customer drawings into ChatGPT?

Not without checking first. Customer drawings are often covered by NDAs or contract terms, and they may be export-controlled. Personal accounts should be ruled out, and even a business plan needs to be checked against the contract and any regulations that apply to the drawing.

Is AI safe for ITAR or CUI data?

Only in an environment reviewed for those specific requirements, and public consumer tools should be treated as off-limits. Where the data is stored, who can access it, and which cloud security requirements apply all matter. Have your export compliance or CMMC owner approve the setup before any controlled data goes in.

Does predictive maintenance make sense for a small shop?

It can, on machines where downtime is expensive and you can collect condition data over time. If you lack machine data or failure history, start by capturing the data and setting simple alert thresholds. Full failure prediction needs examples of failures to learn from, which a well-maintained machine may rarely provide.

Will AI replace estimators or machinists?

That is not what I see these tools being good at. They remove reading, typing, and searching, while the judgment on price, process, and part quality stays with experienced people. Their knowledge is what makes the output worth checking, so a good first project usually makes them more valuable.

The one-sentence version

Start with the paperwork around production, test it on quotes you have already finished, keep customer, controlled, and trade secret data out of unapproved tools, and leave price, part acceptance, safety, and machine control with people.

If you want an AI setup where those data rules are built in rather than left to memory, that is what my secure AI work covers.


Jose Lugo is a CISSP-certified security engineer with 12 years of U.S. Army intelligence experience. He builds secure AI work environments and fast, maintainable websites for businesses at josecustom.ai. See his portfolio of 13 live client systems at portfolio.josecustom.ai.