← Back to Field Notes

November 24, 2025 · The Bloomfield Team

How AI Reads Shop Floor Data (And What It Can't Do Yet)

An AI system reads 14,000 job records from a shop's ERP in about 90 seconds. It cross-references material types, cycle times, setup durations, scrap rates, and delivery performance across every one of those records. Then an estimator asks: "What did jobs like this one typically cost us in labor hours?" The system returns an answer with a confidence range, citing the seven most comparable historical jobs as evidence.

That same system cannot tell you whether the new operator on second shift should run the titanium job or the aluminum one first. That decision requires understanding the operator's experience level, their comfort with tight tolerances, and whether the titanium tooling is staged. That knowledge lives in people.

Understanding the boundary between what AI does well and what it does poorly with shop floor data is the difference between a successful implementation and an expensive disappointment. For the full framework, see our complete guide to AI for manufacturers.

What AI Does Well With Manufacturing Data

Pattern recognition across large data sets. An estimator with 20 years of experience might remember the last three times the shop ran a similar part. AI reads every job in the system and finds all of them, including the ones from 2018 that the estimator forgot about, including the one that ran 40% over on labor because of a material issue that nobody documented in the quote.

Structured data extraction. AI reads PDFs, engineering drawings, spreadsheets, and ERP exports and converts unstructured information into structured, searchable data. A stack of 500 RFQs in PDF format becomes a searchable database of material types, tolerances, quantities, and delivery requirements in hours. Manually, that work takes weeks.

Anomaly detection. When a machine's spindle load increases 12% over three weeks while cutting the same material at the same parameters, AI flags the trend. That trend means something, probably bearing wear or a failing spindle component. The data was always there in the machine's output logs. Nobody had time to review three weeks of spindle load data manually.

Cost comparison and variance analysis. Comparing quoted costs to actual costs across 200 jobs and identifying which job types, materials, or operations consistently run over estimate is exactly what AI is built for. The patterns that emerge from this analysis, that aluminum jobs quote accurately while stainless jobs consistently run 15% over on labor, give estimators specific corrections to apply on future quotes.

What AI Cannot Do

Make judgment calls about relationships. A long-standing customer sends a rush order at a price the numbers say you should decline. The relationship is worth $400,000 annually. You take the job at a lower margin because the lifetime value of the customer justifies it. AI sees the margin. It does not see the relationship. These decisions belong to people.

Interpret ambiguous drawings. A tolerance block calls out +/- 0.005" on a feature that geometrically cannot be held to that specification with standard machining. An experienced estimator recognizes this as a drawing error and calls the customer to clarify before quoting. AI reads the tolerance at face value. It does not know that the drawing is wrong.

Replace tacit knowledge. The machinist who knows that a specific Haas machine pulls 0.0003" to the left on long bores and compensates with an offset adjustment carries knowledge that exists nowhere in any system. AI cannot access what was never documented. The most valuable thing AI does with tacit knowledge is create a system where it gets captured for the first time.

Guarantee accuracy on sparse data. A shop with 200 job records in their ERP and no documented setup times will get vague, unreliable answers from any AI system. The technology amplifies the quality and quantity of data it has access to. It cannot manufacture data that does not exist. Shops with fewer than 1,000 structured job records should expect a longer ramp-up period before the system delivers confident answers.

The Honest Middle Ground

AI in manufacturing works best as a research assistant that never forgets and never gets tired. It reads everything your operation has produced in digital form and delivers relevant answers when your team asks questions. The answers are starting points for human decisions. They compress the research phase of every operational task from hours to seconds.

The shops getting real value from AI are the ones that implemented it with this understanding. They did not expect autonomous decision-making. They expected faster access to better information. That expectation was met, and the results show up in faster quotes, better scheduling decisions, and fewer quality escapes.

The shops that will be disappointed are the ones expecting AI to replace experienced people. It will not. What it will do is make every experienced person on your team faster and more consistent by giving them access to every relevant piece of data your operation has ever generated, in the moment they need it.