Start from the decision, not the model
Ask which decisions are made frequently, with incomplete information, and with material cost consequences. That short list is the AI agenda.
Data readiness is a capability, not a prerequisite
Waiting for perfect data delays value. Sequencing use cases so that each one improves the data estate is more effective than a multi-year data programme with no output.
Agentic systems and the operating model
As systems begin to execute steps rather than only recommend them, governance, escalation rules and accountability must be designed explicitly.
Measuring what changed
Every deployment should carry a defined performance metric — cycle time, scrap rate, forecast error, working capital — measured before and after.