# How to Use AI on Minitab and SPC Data Stored on File Shares

> Source: https://netray.co/erp-help/private-ai-layer-minitab-spc-file-share-on-prem-architecture
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: ERP Help

## Short answer

Do not let a language model calculate Cpk. Export or parse SPC data from file shares into a normalized table, compute capability and control limits in code, and give the model tools that return those numbers plus the source file. A local open-weight model then answers questions and summarizes trends, and NTFS permissions decide which files each user can touch.

## Kind

advanced

## System

mfg-systems

## Erp

Minitab and SPC data on file shares

## Module

SPC, Capability, Gage data / Private AI layer

## Query

how to use AI on Minitab and SPC data stored on file shares

## Alternate Queries

- ask questions about SPC data with a local LLM
- Minitab project files AI analysis on-premise
- SPC data file share natural language query

## Applies To

Minitab Statistical Software project and worksheet files, CSV and Excel exports from gages, CMMs and SPC tools, held on Windows or NAS file shares. Applies to any air-gapped or on-prem environment.

## Steps

### Turn SPC files into a private, trustworthy assistant

**Items:**

Inventory the shares: which folders hold Minitab projects, worksheets, CSV exports and reports, who owns them, and what naming convention (if any) ties a file to a part, characteristic and date.

Decide the extraction route. Minitab's own project and worksheet formats are proprietary, so set up a scheduled export of raw data to CSV, or ask engineers to save data as CSV or Excel alongside the project. Use Minitab macros or automation for the export if your licence allows.

Load the exported data into a normalized table: part, characteristic, subgroup, timestamp, value, spec limits, machine or operator, and source file path.

Compute subgroup means, ranges, control limits, Cp, Cpk, Pp and Ppk, and rule violations in Python or SQL, and store the results with the formulas used.

Index report text, PDFs and engineering notes locally, and serve an open-weight model with vLLM or Ollama on your GPUs.

Expose tools such as get_capability, list_out_of_control_points and find_files_for_part through an MCP server; the model reports numbers returned by tools and cites the source file.

Mirror NTFS or share permissions to every query, log all tool calls, and compare the assistant's numbers with Minitab for a sample of characteristics before go-live.

## Sections

### What is on the shares and what can be read

**Body:**

Typical SPC estates are messy: Minitab projects and worksheets saved by individual engineers, CSV and Excel exports from gages and CMMs, PDFs of control charts, and perhaps a database from a dedicated SPC package. Names are inconsistent and the same data may exist in three copies.

Minitab project and worksheet files are proprietary formats, so do not assume that a generic parser reads them reliably. Plan to get raw data in an open format by export, either manually saved by users or produced by scheduled automation. Where a dedicated SPC database exists, a read-only view of its measurement tables is cleaner than files.

Start by profiling: count files by type and folder, find the active ones, and pick the top 20 characteristics that matter to customers or audits. Scope the first release to those.

**Code:**

# PowerShell: profile the share (read-only)
Get-ChildItem -Path \\fileserver\quality\spc -Recurse -File |
  Where-Object { $_.Extension -in '.mpx','.mpj','.mtw','.csv','.xlsx','.pdf' } |
  Group-Object Extension |
  Select-Object Name, Count

# Python: capability computed in code, never by the model
import numpy as np
def cpk(values, lsl, usl):
    mu, s = np.mean(values), np.std(values, ddof=1)
    return min((usl - mu) / (3 * s), (mu - lsl) / (3 * s))

### Reference architecture

**Body:**

A file crawler feeds a parser and normalizer, which write to a small analytical database. A statistics service computes control limits, capability and rule violations. A tool server wraps those functions. A retrieval index covers reports and notes. A local model sits on top. An audit log records each question, tool call and number returned.

The key design rule is separation of language from arithmetic. The model chooses which tool to call and explains results in plain language. The numbers come from tested code, so a Cpk in an answer can be reproduced and checked against Minitab.

For an air-gapped site, everything runs in the enclave: crawler, database, statistics service, index and model. Updates to weights and libraries move in by controlled transfer.

### Permissions on file shares

**Body:**

File shares have NTFS or NAS ACLs, and engineers rely on them to separate programs, customers and export-controlled work. The assistant must honor them. The crawler records each file's ACL, resolves group membership for the signed-in user at query time, and filters tools and retrieval accordingly.

Beware of the crawler account itself: it needs read access to everything it indexes, so protect it, rotate its secret and audit its use. For export-controlled programs, run a separate index and model instance inside the appropriate enclave rather than relying on filters alone.

Test by asking restricted-user questions that would reveal a protected program's measurement data.

### Write-back: usually none

**Body:**

SPC data generally should not be changed by an assistant. Raw measurement files are evidence. The assistant can draft a summary, a control-plan note or an email to a supplier, which a person reviews. If you want it to create an NCR or a corrective action request, route that through your QMS or ERP API with a human approval step.

Never let the model edit data files, adjust spec limits or re-run capability with changed parameters without a visible, logged request from a person.

### Sizing, phases and limits

**Body:**

SPC assistants are light. A single GPU with a 7B to 32B class model usually handles 50 to 300 users, and the numeric data is small. Document and report indexes are the main storage: tens of gigabytes for large archives. Use the calculator with your counts.

Phase one, three to six weeks: file profiling, export routine, normalized table, capability and control-chart tools. Phase two: report and note retrieval. Phase three: drafted notes and, if wanted, approved record creation in the QMS.

Limits: if your team already has a dedicated SPC package with a database, build on that rather than on file shares. Minitab itself is better for exploratory statistics. The assistant is for finding, checking and summarizing, not replacing statistical judgment.

## Pitfalls

- Letting the model compute Cpk or control limits from pasted numbers instead of calling tested code.
- Assuming proprietary Minitab files can be parsed reliably without export.
- Ignoring duplicates: the same dataset saved in several folders yields conflicting answers.
- Indexing every folder, including export-controlled programs, into one shared index.
- Not documenting the formulas (subgroup size, sigma estimate, constants) behind each reported number.

## Ai Angle

DataRay is built for this: chat over file shares and databases on-prem, with a local model. Netray adds the deterministic statistics tools for SPC, and ERPray links out-of-control characteristics to the ERP lots, work orders and suppliers they came from.

## Faq

### Can an AI read Minitab project files directly?

**Answer:** Not reliably. Minitab project and worksheet formats are proprietary. The dependable approach is to export raw data to CSV or Excel, by scheduled automation or as a team habit, and let the assistant work from those exports plus the reports.

### Why compute Cpk in code and not with the model?

**Answer:** Language models make arithmetic and sigma-estimation mistakes and cannot show their work. A tested function returns the same result every time, with a stated method, and you can reconcile it with Minitab output for audit purposes.

### Will it respect our folder permissions?

**Answer:** Yes, if designed that way. The crawler stores ACLs, and at query time the tool server resolves the user's groups and filters files and results. For export-controlled work, use a separate instance in the right enclave.

### Is this better than just using Minitab?

**Answer:** They solve different problems. Minitab is the tool for analysis. The assistant helps people find the right dataset across many folders, check capability across characteristics, and summarize trends. Use both.

## Search terms

minitab ai, spc data ai assistant, ask questions about spc data, minitab file share search, cpk calculation local llm, on-prem ai spc, private ai quality data, air-gapped spc analysis

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json