AI for small and mid-sized manufacturers: where to start with your records

Where should a smaller manufacturer start with AI?#

A smaller manufacturer should start with AI where it already has years of records behind a frequent decision: what to quote, why parts fail inspection, when machines need attention or what to run next. The decision gives the project a measurable goal, and the records decide whether the goal is reachable.

Starting from a tool reverses that logic. A quoting assistant or scheduling optimizer that looks impressive in a demo learns little from a shop whose quotes live in email and whose actual hours were never booked against jobs.

The owner's job at this stage is choosing, not building. Name the decision, name the person who makes it today, and ask which system holds the history of that decision. If nobody can answer the last question, the project is not ready.

Use cases and the records each one needs#

Each common use case depends on a specific set of records, and the readiness check differs by use case. The table pairs the use cases most smaller plants consider with what each requires and a sensible first step that needs no new software.

Five readiness checks for your records#

Five readiness checks tell you whether a set of records can support AI work, internal or external. Run them on one record family at a time, starting with the one behind your chosen use case.

A record family that fails one check is not useless. Inconsistent codes can sometimes be rebuilt from free text, and broken links can be restored from job numbers, but each repair adds effort that should be weighed before a project starts.

• 1Digital: the records are in a system or structured files, not on paper travelers or whiteboards.
• 2Linked: each record connects to the job, part, customer or machine it concerns, and to its outcome.
• 3Coded: reason, defect and status codes are used consistently, not left blank or set to other.
• 4Deep: enough history survives past migrations to show patterns, not only the current year.
• 5Owned: the records are yours to use, with customer-owned drawings and personal details identified.

Common record problems in job shops and mid-sized plants#

The most common record problems in job shops and mid-sized plants come from work that happens outside the system. Quotes built in spreadsheets and sent by email, rework done without a nonconformance, and machine fixes made without a work order all leave gaps where the most useful information should be.

Migrations cause the other big gap. A shop that changed ERPs may have converted only open items, leaving older quotes and jobs on a server nobody opens. Before deciding that history is too short, find out what the old system still holds.

Tribal knowledge is the hardest gap to close. Senior estimators and setup leads carry rules that never reached a record. Asking them to explain past decisions while reviewing real quotes or jobs captures some of that knowledge in a form the company can keep.

Buying AI tools: what to check about your data#

Buying AI tools is the realistic path for most smaller manufacturers, and the vendor's data terms deserve as much attention as its features. Many ERP, quality and maintenance vendors now include AI functions, and their agreements say whether your records can be used to train models that serve other customers.

Your quote and process history is part of your competitive position, so decide consciously whether to give it away. Terms can also change after you sign: in August 2023, after backlash over earlier changes to its terms, Zoom added a sentence saying it would not use audio, video or chat customer content to train its AI models without consent. Read the data use section of the contract and put these questions to the vendor in writing.

• 1Is training on our records on by default, and can we opt out in the contract rather than only in a settings screen?
• 2Are our records used only to serve us, or also to improve models for other customers?
• 3If records are used for training, are they aggregated or de-identified first, and who decides how?
• 4Which subprocessors or outside model providers receive our records, and on what terms?
• 5Can we export our records and the tool's outputs in full if we leave?
• 6How will we be told about changes to the data terms, and can we refuse them?

How to judge whether the first project worked#

The first project worked if the decision it targeted got measurably better, not if the tool was adopted. For quoting, that means estimates closer to actual hours or a better hit rate on the work you want; for maintenance, fewer unplanned stops on the machines you chose to target.

Set the baseline from your own records before anything changes. The same history that trains or configures the tool gives you the before picture, which is one more reason to fix record capture early rather than after the project starts.

Illustrative: a sheet metal fabricator picks its first project#

Illustrative: a fictional sheet metal fabricator wants to use AI for quoting because its senior estimators are nearing retirement. A records check shows that quotes moved into the ERP several years ago, are linked to the resulting jobs and carry won or lost status, and actual labor is booked by operation.

Maintenance records fail the check, because most repairs were never logged. The owner chooses quoting as the first project, starts logging repairs as work orders so maintenance can follow later, and asks the senior estimators to annotate a sample of past quotes with the reasons behind their pricing. The same quote history later goes through a metadata-only licensing fit check.

The records you clean up have a second use#

The records you clean up for internal AI have a second use, because the same linked, consistently coded history is what AI developers look for when they build and evaluate systems for industrial work. A quote history with outcomes, or a maintenance history with diagnoses and fixes, can serve the shop and a licensing review at once.

SourceX assesses that history through the SourceX five-step transaction: Supply, Rights, Preparation, Approval and Delivery, starting with metadata so nothing is shared during the initial assessment. The SourceX Enterprise Data Value Framework guides which record families are worth taking further, the data is licensed rather than sold, and the company keeps ownership.