AI should enter a system because it is appropriate to a defined problem, not because a transformation programme needs an AI workstream. Start by identifying the decision being supported, the cost of an incorrect result and where human judgement remains necessary.
Define purpose and decision rights
Specify users, intended outcomes and prohibited uses. Assign responsibility for data, model behaviour, production operation and exceptions. A human-in-the-loop statement is only meaningful when the human has enough context and authority to intervene.
Establish data discipline
Document provenance, permitted use, retention, representativeness and known gaps. Model performance cannot compensate for a poorly defined target or data that does not reflect the operating environment.
Engineer observable behaviour
Monitor input quality, changes in distributions, output confidence, operational failures and user overrides. Choose explanations that help the actual user make a decision rather than adding technical artefacts that nobody uses.
Operate change deliberately
Models, dependencies and business processes change. Version inputs and models, test meaningful failure modes, make rollback possible and use production feedback to decide whether retraining is justified.
Applied AI is one component of digital transformation architecture. Its value depends on the workflow, controls and maintainable system around it.