Before discussing a model, I would want to know which commercial decision or workflow the proposed AI solution is meant to improve. The question sounds basic, but it changes the entire investment conversation. A technically impressive pilot can still be the wrong use of capital if the underlying work is not ready to change.
Start with the business problem
Consider a illustrative proposal to generate faster customer insights. Faster than what? Used by whom? To change which decision? If the answer is 'better productivity,' the team has not yet described a measurable outcome. Define the current workflow, its bottleneck and the decision owner before evaluating vendors or models.
Ask whether AI is necessary
Could a clearer process, a simpler analytics model, better data definitions or a rules-based system solve the problem? The alternative is not 'AI or nothing.' It is the best available intervention for a specific commercial need.
Check the data before the demonstration
My public writing on AI data readiness asks practical questions: who owns the source, whether consent permits the proposed use, whether market-specific residency or transfer rules apply, whether the data is current and complete, and who can access outputs. These are build prerequisites, not cleanup tasks. In a multi-country pharmaceutical organization, the answers can differ by market.
Decide what work changes
Who will use the output, what existing step will it replace, what training is needed and who is accountable when it is wrong? If the workflow remains unchanged, a faster model may simply create a faster queue of recommendations nobody acts on.
Agree on exit conditions
A pilot needs success thresholds and reasons to stop. Set expectations for quality, appropriate use, time saved or decision improvement, cost, risk and adoption. Distinguish a test that needs redesign from one that should end. Stopping a weak investment is a valid outcome.
A worked decision
Suppose a team proposes AI-assisted summarization of approved internal commercial reports. First establish the time currently spent, the review and confidentiality requirements, the expected quality standard and who validates the summary. Compare with a simpler reporting redesign. Test with permitted data and measure whether decision preparation actually improves. Do not scale on the strength of a compelling demonstration alone.
In my view, the strongest AI investment process is not the one that approves the most pilots. It is the one that makes clear why a particular intervention deserves resources—and what would change our mind.
Follow the decision
Who can challenge the recommendation, and what work will actually change? This question connects to two other parts of the operating model:
Illustrative example; no employer implementation is described. The views I share are my own and do not represent those of my employer.
Put this into practice
Fund the evidence, not the demonstration.
Use the AI Investment Decision Card to specify proof, alternatives, controls and conditions for stopping.
Open the AI Investment Decision Card