Most Enterprise AI Programs Define the Technology Before They Define the Problem

Foundation models are appropriate when the task is primarily generative and when the organization's proprietary data does not represent a structural advantage in the task being performed. Custom model development makes sense when the prediction problem is specific to the organization's data, when the performance requirements exceed what a general model can achieve on the specific task, or when regulatory or data privacy requirements prevent the use of external model providers. In most enterprise programs, the answer is a combination: foundation models for generative and language tasks, custom or fine-tuned models for prediction and operational intelligence applications where proprietary data and domain specificity determine model quality.

The data requirement depends entirely on the use case. What organizations consistently underestimate is not the volume of data required but the quality. Data that has been collected for reporting purposes often lacks the structure, consistency, and labeling required for model training. A data readiness assessment before the program begins identifies these gaps and scopes the work required to close them.

A focused, well-scoped AI use case with data that has been assessed as ready can move from scoping to initial production deployment in twelve to twenty weeks. Programs that skip the data readiness phase and discover gaps during build consistently take longer and cost more than programs that front-loaded that work.

Pharmaceutical AI programs at DAM are designed with regulatory requirements as a primary design constraint, not a secondary consideration. For programs that touch HCP data, adverse event data, or promotional content, the relevant regulatory frameworks are incorporated into the data handling and output review architecture. The compliance team is a stakeholder in the program design, not a reviewer brought in at the end.

AI governance is not a policy document. It is a set of operational controls embedded in how AI programs are built, monitored, and managed: a defined model inventory, a monitoring framework that detects model drift, explainability requirements matched to the decision context, a human-in-the-loop design for high-stakes decisions, and a process for handling model errors. Governance is designed into the program from the start.

The return on an AI program should be measured against the business metric it was designed to move, not against the model's technical performance. This requires that the business metric and the measurement methodology are agreed before the program begins, that the measurement baseline is established before deployment so the counterfactual is clear, and that the business outcome is tracked in production rather than only during pilot.