Pranav Dhakulkar
Principal AI & DS Consultant
If you run the historian, the LIMS, the ERP integration or the network between the plant floor and the business, this is for you. When an AI project in your plant stalls, it rarely stalls because of the model. It stalls on data, and the data sits with you.
The pressure is already on
Leadership has read the same headlines you have. Yield optimization, predictive maintenance, energy, quality release. Vendors are lining up with platforms that promise to connect to your PLCs, SCADA, historian, LIMS, MES and ERP and start producing value within weeks.
The promise assumes something that is rarely true in a working dairy plant: that the data behind those connections is complete, aligned and trustworthy enough to learn from
What it actually looks like from the IT and OT side:
Two clocks that never agree:
Process data lives in the historian by the second. Production and cost live in the ERP by the batch or the shift. Lab results in the LIMS carry the time a technician entered them, not the time the milk was sampled. Lining the three up is a manual job every time someone asks.
Tags without meaning:
Thousands of historian tags, many named by an integrator years ago, some pointing at instruments that were moved, replaced or recalibrated with no record. Nobody is sure which ones still mean what their name says.
Lab results with no batch attached:
Composition and quality results exist, but they are not joined to the batch, silo or vat that produced them. Without that join, no model can connect a quality issue to a process condition.
Gaps nobody measures:
Compression settings, network drops, a data collector that restarts overnight. Every plant has missing data. Very few know how much, or where.
Spreadsheets as the integration layer:
When systems do not talk, people do. The monthly yield or solids report is assembled by hand, and by the time it reaches the people who can act on it,
the product has already shipped.
A boundary that was built on purpose:
OT networks are segmented for good reasons. Moving plant data out safely, to anywhere a model can use it, is a project in itself and usually an unplanned one.
Why the pilot always looks fine:
A pilot runs on one line. Two people clean that line's data by hand for two weeks. The model performs and everyone is pleased. Then it meets line two and the hand cleaning does not scale.
The pilot did not prove the data was ready. It proved two people could clean it.
This is not only a dairy problem. In McKinsey's 2025 survey of manufacturing COOs, 46% named limits in their own data or IT and OT systems as a barrier to scaling AI,
and only 2% said AI was fully embedded across their operations.
What to do before the next AI project
Before anyone connects a model, get a clear, evidence based answer to five questions:
1) Instrumentation Are the stages that move yield and quality actually measured, and calibrated?
2) Capture Is that data kept, at a useful resolution, for long enough to learn from?
3) Alignment Do process, lab and production data share a clock and a batch ID?
4) Completeness Do you know how much data is missing, and where?
5) Ownership Does someone own the data across the IT and OT boundary once it is in use?
If any of those questions produces a pause, that pause is the finding. It is far cheaper to have it now than six weeks into an implementation, with a budget committed and a vendor waiting on your data.
The model was never the hard part. What you need first is a data assessment.
About Findability Sciences
Findability Sciences is an enterprise AI company that helps dairy processors turn plant data into
better decisions, through the following tools and services:
Data Assessment :A readiness check of the plant data you already have, before any AI project begins.
Lacta AI™ : An industrial AI platform built for dairy plants.
Lacta Insights™ : AI-powered operational intelligence for dairy plant performance.
Lacta BPC™ : Conversational AI for enterprise knowledge and business workflows.
Learn more at Findabilty.ai / Dairy AI.
