Can AI read shop-floor data from a custom MES?

Industry PlaybooksCanBy Keyur PatelReviewed
SHORT ANSWER

Yes. Most Indian MES systems are custom builds in PHP, .NET, or Excel pipelines. AI connects to the underlying database directly, regardless of frontend framework, and reads OEE, production, downtime, quality, and changeover data. Joined with Tally for cost view and ERP for plan, it works for sheet-driven plants too.

What 'MES' actually means in Indian manufacturing

When a global vendor says MES, they mean Wonderware, Rockwell FactoryTalk, Siemens Opcenter, or one of the other enterprise platforms. When an Indian mid-market plant says MES, the meaning is much wider. It might mean a proper tier-one platform on a few high-value lines. It often means a homegrown PHP or .NET application written by a local developer five years ago, sitting on a MySQL or SQL Server database. It frequently means an Excel pipeline where supervisors enter shift data into a workbook that gets consolidated by an MIS analyst.

The first question every shop-floor analytics conversation should answer is: which of these are we actually working with. The integration approach is genuinely different for each, and the vendors who pretend otherwise tend to spend the next four months stuck on the wrong assumption.

The four MES patterns we see in India

What we find when we walk into an Indian mid-market plant for the first MES integration call.
PatternHow AI connectsTypical effort
Proper tier-one MESVendor API or direct read of the underlying database1 week, mostly access provisioning
Custom PHP or .NET appDirect read of MySQL or SQL Server database1 to 2 weeks, depending on schema clarity
Excel pipeline from supervisorsFolder watch on the consolidation Excel, scheduled ingest1 week, plus light file-naming discipline
Hybrid - lines on different systemsAll of the above in parallel, joined on canonical SKU2 to 3 weeks, item master alignment is the bulk

The customer profile we deal with most often is the third and fourth row. A custom PHP app for the older lines, a newer ERP module for the recent ones, and an Excel sheet for two job-shop machines that never made it onto either system. AI reads all three in parallel and presents them as one shop-floor view.

What shop-floor data we actually read

Across the four patterns, the data captured is broadly the same. The frontend changes; the underlying questions a production head wants answered do not.

STANDARD SHOP-FLOOR QUESTIONS WE ANSWER
  • Production by line, shift, SKU. Output count and weight, with comparison to the planned schedule for the day.
  • Downtime by reason code. Planned (changeover, maintenance) versus unplanned (breakdown, material short, manpower short, power), aggregated weekly.
  • OEE components - availability, performance, quality. Calculated where the data supports it, presented as the three components rather than a single OEE number that hides the diagnosis.
  • Quality reject reasons. Top reject categories per line per week, with the rate trending up or down. Critical for the BOM variance discussion.
  • Changeover times. Average and outliers per SKU pair. Surfaces planning discipline issues that cost capacity quietly.
  • Operator and shift comparison. Anonymised where needed. Shift A consistently doing 8% better than Shift B is a training discussion, not a blame discussion.

How AI connects to each pattern

For a proper tier-one MES, we use the vendor's API where one exists or a read-only database connection where it does not. For a custom PHP or .NET app, we connect directly to MySQL, MariaDB, PostgreSQL, or SQL Server with read-only credentials. The frontend framework genuinely does not matter - the AI reads the database, not the application layer.

For an Excel pipeline, we set up a watched folder where the MIS analyst drops the consolidated workbook (or where the script that builds it writes the file). KolossusAI ingests on a schedule and treats each new file as the source of truth for the day. The supervisor's data entry workflow does not change - the AI reads what is already being produced.

For a hybrid plant, we run all three connectors in parallel and reconcile the production output to a canonical SKU master. The output of one query, asked in plain English, answers across all the lines regardless of which system captured the data.

What changes when you join MES with Tally and ERP

Shop-floor data alone tells you whether the line ran. It does not tell you whether the line ran profitably. Joining MES output with Tally purchase entries gives you actual material cost per unit produced. Joining with the ERP production plan gives you variance against schedule. This is where the AI layer earns its keep - the cross-system join is the analyst's full-time job today, and it is the same join every week.

6 - 10 hrs
Analyst time per week
Joining MES, Tally, and ERP for the manual MIS pack
2 hrs
With AI doing the join
Analyst spends time on diagnosis, not data plumbing
1 week
Detection loop
Versus 4 to 8 weeks when the join only happens at month-end

See AI for Indian manufacturers for the full multi-system pattern, or AI for custom CRMs if your shop-floor app is built on the same custom stack as your sales CRM.

Security model for shop-floor reads

The shop-floor systems we read often run inside the plant network on a local server. KolossusAI deploys a small read-only connector inside that network. The connector authenticates with credentials your IT team controls, reads only the tables we have agreed during onboarding, and never writes back to the source. The schema we read is documented and reviewed with your team before connection.

For air-gapped plants, KolossusAI runs fully on-premise - the model and the connector both live inside the plant LAN and no data leaves your boundary. This is the deployment shape we use most often for defence and pharma customers, and it works equally well for a paranoid auto-component plant that simply does not want production data on the public internet.

The honest limits

AI does not magically read data that is not being captured. If your shop floor is genuinely paper-only and the MIS analyst types numbers into Excel from a clipboard at the end of the shift, the variance and downtime quality is bounded by what the supervisor wrote down. The AI reads what the Excel records. It does not invent the missing fields.

For plants that want to upgrade data capture quality at the same time, KolossusAI's onboarding includes an honest review of which shop-floor capture points are weakest and what minimum discipline change pays back fastest. We do not sell MES upgrades, but we will tell you when one is worth doing before the analytics will be trustworthy.

FREQUENTLY ASKED

Questions readers actually ask.

Does AI need a special API on our custom MES?

No. The connector reads the underlying database directly with read-only credentials. We do not need the application developer to expose an API or modify the codebase. Most custom Indian MES apps are built on MySQL, MariaDB, or SQL Server, all of which we read natively. PostgreSQL and a handful of older MS Access databases are also covered. The only thing we need from your IT team is read-only access and a one-page schema walkthrough during onboarding.

What if our MES vendor went out of business years ago?

Common situation, especially for systems written for specific industries by single-developer shops in the early 2010s. The application is unmaintained, but the database is still capturing data every shift. We treat the database as the source and bypass the application entirely. The production head keeps using the same data entry interface on the shop floor; the AI reads the data the application is writing. No vendor cooperation needed.

What is OEE and can AI compute it from custom MES data?

OEE - Overall Equipment Effectiveness - is the product of availability, performance, and quality, expressed as a single percentage. AI can compute OEE from custom MES data if the underlying capture has the three components: planned run time and downtime (availability), produced units versus theoretical maximum (performance), and good versus rejected units (quality). KolossusAI presents the three components separately by default, because a single OEE number hides which of the three is dragging the score down.

Can it work with shop-floor data captured only on paper?

Indirectly. AI cannot read paper. But every plant we have seen with paper capture also has an MIS analyst typing the data into Excel daily or weekly. The Excel becomes the source. KolossusAI watches the consolidation file and ingests on schedule. The honest limitation is that data quality is bounded by what the supervisor writes on the sheet and the analyst types into the workbook. We work with what is captured today and flag the capture gaps the owner should close to make the analytics richer.

How do you handle multiple MES systems across plants?

Each plant typically has its own setup - a custom MES at one plant, an Excel pipeline at another, a tier-one platform at the headquarters plant. KolossusAI connects each plant's MES separately and maintains a plant-to-SKU map. Queries can scope to a single plant, a region, or the full network using the same phrasing. Production head asks the question once; the AI runs the query against every connected plant and returns a consolidated answer with per-plant drill-down.

How long until shop-floor data is queryable in plain English?

For a single-plant deployment with a custom MES on MySQL or SQL Server, the typical timeline is two to three weeks. Week one is the schema walkthrough and read-only connector setup. Week two is the SKU and reason-code alignment with Tally and the ERP. Week three is real use - the production head and shift supervisors ask questions in plain English and get answers grounded in shop-floor data they trust. See how the manufacturing deployment works.