Where AI can help on the shop floor

Start with the repetitive work

An AI project in manufacturing does not have to begin with a robot, a camera system, or a connection to production equipment. A more manageable place to start may be the information surrounding the work: incoming requests, document preparation, and the questions people answer repeatedly.

Look for a task with a recognizable input and a useful output. Someone receives an RFQ, gathers details, checks what is missing, and passes it to an estimator. Someone else turns approved notes into a customer update. These are candidates to investigate—not promises that a tool can do the job reliably without supervision.

Walk through a few recent examples with the person doing the work. Notice where they copy information, search folders, or wait for clarification. The best first opportunity might be a simple form or integration rather than AI. If fixed rules solve the problem, use fixed rules.

Three places to investigate

Prepare an RFQ for review

In a hypothetical machine shop, a request arrives with a drawing, an email, and a spreadsheet. An assistant could draft a structured intake summary: customer, part identifier, quantity, requested date, and missing information. The estimator would check that summary against the original files before using it.

Keep pricing, material suitability, tolerances, and delivery commitments with the qualified people responsible for them. A summary that saves reading time is useful only if omissions and incorrect details are caught. Give the reviewer a direct route back to the source document instead of a confident-looking answer with no evidence.

Find approved information

A searchable collection of approved procedures can be a more focused experiment than a chatbot connected to every company file. For example, a person might ask where to find the current packing checklist and receive a link to the controlled document.

Decide who owns that collection, how revisions enter it, and how obsolete copies are removed. The assistant should not become a second, unofficial authority. If it cannot locate an approved answer, returning no answer is better than improvising one.

Draft routine communications

An assistant might turn a supervisor's approved notes into a draft handoff or customer email. Treat that as writing support. A person still checks the facts, removes confidential information, and approves the wording before it leaves the business.

Start with communications that have a clear template and modest consequences if a draft needs correction. Do not begin by allowing a system to send commitments, revise production instructions, or change records without review.

Check the inputs and the boundaries

Generative AI can produce plausible but incorrect information. NIST identifies this risk as confabulation and also highlights data privacy risks in its Generative AI Profile. A fluent answer is not proof that a drawing, quantity, or specification was interpreted correctly.

Before a pilot, write down what the system may read and what it may do. Use authorized, non-sensitive examples initially. Check the provider's applicable data terms before uploading customer drawings or proprietary process information. Keep access limited to the materials the task actually needs.

Run a small, measurable pilot

Pick one task, one owner, and a limited collection of representative examples. Include ordinary cases and awkward ones: incomplete requests, conflicting attachments, and revisions. Keep the existing manual route available.

Record the time to finish the work, not just the time to generate a draft. Include checking, corrections, and exception handling. Track missing details and incorrect details separately; a fast summary that overlooks an important revision is not a success.

Agree in advance on what would justify continuing. That might be more consistent intake, less repeated searching, or shorter preparation time without added errors. If the reviewer spends longer repairing the output than doing the task manually, change the approach or stop the pilot. A smaller, dependable improvement is a better foundation than a broad system nobody trusts.