A model trained on observed data learns the associations present in that data. Once the model is used to choose an action, however, the data-generating process can change, and the predictive accuracy measured before deployment may no longer correspond to the quantity that matters for decision-making. This is a central bottleneck in deployed AI: the problem is often not weak prediction, but prediction answering a different question from the one the decision actually poses.
Our research develops estimators for that decision-relevant question, together with guarantees that remain valid when flexible machine learning is used for the complex components that cannot realistically be specified by hand. The four bottlenecks below are where this distinction becomes consequential in practice.
A randomized trial is often not an option — too slow, too costly, or wrong to withhold from people. And a trial you do have rarely covers the population you must decide about. We provide methods that recover the effect of a decision from records not designed to measure it, and carry findings from one population to another.
Knowing that something works on average does not tell you who to give it to. We estimate how the effect differs across people and turn that into a rule for who to act on — one that can respect a budget, a capacity limit, or a fairness requirement you set in advance.
Deploying a model can change actions, outcomes, and the population it faces, creating distribution shifts that pre-deployment accuracy may not capture. We develop methods for reliable learning under these shifts and for more robust extrapolation beyond the conditions represented in the training data.
Operational records are rarely complete. People drop out partway through, outcomes are still unfolding when the data is pulled, and what you care about is often measured only indirectly. We provide methods that work with these records as they are, rather than discarding whatever is incomplete.
Estimate which parts of a care program or digital therapeutic actually drive outcomes and who is likely to stay engaged — evidence that shapes what to build before a full trial is run.
Test whether ad performance or survey responses measured in one audience carry over to another, so a result from one campaign or panel can guide the next without running a new large-scale experiment.
Compare operating policies, charging protocols, or component designs across plants and process conditions, and identify which choices are likely to work best in a specific target environment.
Make stress tests more credible with explicit “what if” market scenarios, and build synthetic financial twins of assets, individuals, or firms for lower-cost testing in areas such as portfolio analysis and credit assessment.