What data do you need before deploying AI agents?
Enough to answer the specific decision the agent takes, at the frequency the decision is taken, with a clean feedback loop back into the source system. In practice that means live machine telemetry over OPC UA or Modbus, current ERP master data, working MES batch genealogy or work-order flow, and a named domain owner responsible for keeping the source data clean. Without those four, an agent runs advisory-only at best.
Minimum viable dataset by agent role
Maintenance agent: vibration, current, temperature or cycle-count telemetry on the target assets, plus historical CMMS work-order feedback. Quality agent: live inline vision or gauge output, plus MES batch genealogy for upstream traceability. Procurement agent: live consumption from the MES, on-hand and open-PO from the ERP, and vendor lead-time master data. Scheduling agent: order book with due dates, machine state, and a maintained setup / changeover matrix.
Each also needs the failure-mode catalogue or spec catalogue owned by the domain expert on your side; the agent runs against those, it does not invent them.
The feedback loop that matters more than the data itself
An agent that makes decisions but never learns which ones were right stops being useful within a quarter. Every agent role has a natural feedback signal: was the predicted failure real, was the quarantined batch a true reject, was the PR quantity right for consumption, was the recommended schedule executed as written. If that signal is not captured back into the source system, the agent goes stale.
Data-quality prerequisites we insist on before go-live
Sensor calibration cadence with a logged record. BOM change-control with a named owner. CMMS work-order feedback actually entered by technicians. MES batch-genealogy completeness on the target line. Vendor master-data hygiene on lead times and MOQs.
If any of these is missing, the pilot ships in advisory-only mode until the plant closes the gap. This is not a technical limitation; it is what "responsible" looks like.
Related questions
Ajinkya’s reference plants run on OPC UA / Modbus / MQTT for machine telemetry and SAP S/4HANA (RFC / BAPI) for ERP integration. Concrete data-quality-gap findings by pilot: [NEEDS DEPLOYMENT NUMBER].
Reviewed by Amey Kadle, Founder, Ajinkya Technologies. Last reviewed: 2026-08-29.