Triage Scores

The Problem

Every day, decision makers must choose what to do in high-stakes situations: Should a doctor operate, an employer offer incentives to retain employees, or a judge require bail while defendants await trial? To help them make these decisions, many professionals rely on “risk scores,” which estimate what will happen if no action is taken. For example, a risk score might predict the chances that an employee might leave their job without an incentive to stay.  

But risk scores typically do not predict the likely outcome if the decision-maker does intervene. For example, they do not predict the odds that an employee will quit their job even if they receive an incentive to stay. This limitation means that risk scores provide decision-makers with only half the information they need. Without understanding the full picture, time, money, and effort may be directed away from the cases where intervention would make the most difference.  

The Solution

To address this gap, the Access to Justice Lab (A2J Lab) at Harvard Law School will develop and test “triage scores,” a new, evidence-based AI-driven decision tool. Unlike risk scores, which predict outcomes only when no intervention occurs, triage scores estimate outcomes both with and without intervention. In the bail example, a triage score could predict the chances of a negative outcome if bail is not imposed (for example, would a defendant released without bail commit another crime?) and compare it to the chances of a negative outcome if  bail is imposed (for example, the defendant might lose their job while waiting to post bail). Being able to compare the possible outcomes with and without intervention will allow a judge to better evaluate the impact of choosing to impose bail.  

The Research

Working with the Norfolk Commonwealth Attorney’s Office and Philadelphia Legal Assistance, the A2J Lab and a team led by Professor Kosuke Imai from the Departments of Government and Statistics will develop triage scores with the goal of improving decision making across a variety of fields. In law, for example, the focus might be on public safety, justice, and child welfare by helping legal professionals better understand the likely impact of their decisions.  

The research has four components: developing the statistical methodology underlying the AI tool; building context-specific triage scoring systems; assessing the accuracy of the tool; and conducting randomized controlled trials to study the real-world impact.  

Field tests will focus on two legal settings. First, Philadelphia Legal Assistance (PLA), which provides free legal aid to low-income clients but cannot meet existing demand, will test whether triage scores can help allocate legal assistance to child custody cases where clients are likely to succeed only if they get one-on-one assistance. Second, the Norfolk Commonwealth Attorney’s Office will test whether triage scores can inform decisions about offering plea bargains versus going to trial, helping prosecutors allocate limited trial resources. 

The Impact

Even within the initial test cases, the potential reach is substantial. Organizations like Philadelphia Legal Assistance serve more than five million individuals nationwide, and state and federal prosecutors handle roughly nine million criminal cases each year. Improving decision-making in these contexts could affect millions of lives.

Beyond these settings, triage scores have the potential to replace traditional risk scores across law enforcement, healthcare, business, and other resource-constrained fields. By identifying which interventions will have the greatest impact, these tools could reduce strain on limited resources, increase efficiency, and improve outcomes across a wide range of decisions. If successful, triage scores will be a new line of work within the A2J Lab.