← Back to Field Notes

March 29, 2026 · The Bloomfield Team

How to Get Your Shop Floor Data Into an AI System

The average manufacturing operation with 50 to 200 employees generates useful data in six to ten different systems and locations. ERP job records. Quoting spreadsheets. Supplier emails. Quality inspection logs. Setup sheets in binders on the floor. Tooling databases maintained by one person. Machine monitoring outputs. Customer correspondence. All of it informs decisions, and none of it talks to each other.

Getting this data into an AI system does not require replacing those systems. It requires connecting them, structuring the outputs, and building a data layer that AI can read across all of them simultaneously.

Start With What You Have

The first step is an inventory of where data lives and what format it takes. In a typical shop, the breakdown looks something like this:

• ERP system (JobBOSS, Epicor, Global Shop Solutions, E2): job records, work orders, material purchases, labor hours, shipping records. Structured data with decent historical depth. Usually 3 to 15 years of records.
• Quoting files: Excel spreadsheets, sometimes a quoting module inside the ERP, sometimes standalone files on a shared drive. Semi-structured. Naming conventions vary by who created the file.
• Customer correspondence: emails with RFQs, drawing revisions, spec changes, delivery confirmations. Unstructured. Scattered across inboxes.
• Quality records: inspection reports, CMM data, nonconformance reports. Mix of structured data in a QMS and unstructured notes in files or paper records.
• Operator knowledge: setup notes, tooling preferences, process adjustments. Mostly unwritten. Lives in people's heads and occasionally in notebooks.

AI in 2026 can work with all of these formats. Structured ERP data is the easiest to ingest. Unstructured emails and documents require natural language processing. Operator knowledge requires a capture process before it can become data at all.

The Three-Layer Architecture

A well-built manufacturing AI system connects data through three layers.

Layer 1: Extraction. This is the plumbing. APIs pull structured data from your ERP on a scheduled basis, typically nightly for historical data and every 15 to 60 minutes for active job data. Email integrations scan incoming messages for RFQs and customer communications. Document processing reads PDFs, drawings, and spreadsheets. The extraction layer does not change your source systems. It reads from them.

Layer 2: Normalization. Raw data from different systems uses different formats, naming conventions, and structures. Your ERP might call it "Part Number." Your quoting spreadsheet calls it "PN." Your customer's RFQ calls it "Item." The normalization layer maps these to a common schema so the AI sees one unified dataset. This is the step most people underestimate. A shop with 10 years of ERP data and inconsistent part numbering can spend 4 to 8 weeks on normalization alone.

Layer 3: Intelligence. This is where the AI model lives. It reads the normalized data, identifies patterns, and delivers outputs to your team through whatever interface fits the workflow. A quoting tool. A scheduling recommendation engine. A knowledge search system. The intelligence layer is the visible product. The two layers beneath it are what make it accurate.

Common Data Problems and How to Handle Them

Inconsistent naming. The same part shows up in your ERP as "HYD-MANIFOLD-4500," "Hydraulic Manifold 4500," and "4500-HM" across different job records. AI can learn to match these variants, but the initial mapping requires human review. Budget 1 to 2 weeks for a shop with 5,000 or more unique parts.

Missing fields. ERP records where the setup time field was never filled in, or material cost was entered at the job level but not broken out by operation. Missing data does not block AI from working. It limits the accuracy of specific outputs. A quoting tool that lacks historical setup time data will estimate setup less precisely until that data accumulates going forward.

Paper records. Job travelers, inspection sheets, and setup instructions that exist only on paper need to be digitized. OCR technology handles typed forms well. Handwritten notes require more manual effort. The practical approach: digitize forward, scanning new records as they are created, and selectively backfill critical historical records.

Data in people's heads. The most valuable data in most manufacturing operations is undocumented. The senior machinist who knows that a specific aluminum alloy warps during heat treatment unless you stress-relieve it first. The estimator who knows that Customer X always negotiates 8% off the initial quote. Capturing this knowledge requires a structured interview and documentation process that turns experience into searchable data.

What You Do Not Need

You do not need perfect data. AI works with imperfect data and improves as the data improves. Waiting for your ERP to be perfectly clean before starting an AI project means waiting forever.

You do not need a data warehouse. The AI system builds its own data layer. You do not need to invest $200,000 in a standalone BI platform before building an AI tool.

You do not need to replace your ERP. The AI connects to whatever you have. JobBOSS, Epicor, Global Shop, E2, QuickBooks with a separate production tracking system. The integration layer adapts to your stack.

You do not need an IT department. Most shops in the 50 to 200 employee range do not have dedicated IT staff. The AI builder handles the integration work. Your team handles the domain knowledge.

The Practical Timeline

For a typical first AI project connecting to an ERP, a quoting spreadsheet, and customer email, the data work breaks down as follows:

• Weeks 1 to 2: Data audit. Inventory all sources, assess quality, identify gaps.
• Weeks 3 to 6: Extraction and normalization. Build connections to source systems, map data to common schema, handle inconsistencies.
• Weeks 7 to 10: Model training and testing. AI learns from your historical data, team validates outputs against known jobs.
• Weeks 11 to 12: Deployment. Tool goes live alongside existing process. Team uses both in parallel until confidence builds.

Twelve weeks from data audit to deployed tool. The data work is the foundation, and it takes about half the total timeline. That investment pays dividends on every AI application you build afterward, because the data infrastructure serves all of them.