Next-best-action, built four times
I build products that turn complex data into the next best action for a specific person, in the workflow where that action needs to happen. I've built it four times, in four different settings, with different technology each time.
- Where
- Komen, CareTogether, iCAD, SigmaFlow
- Span
- 2013 to present
- Methods
- Rules, machine learning, LLM agents
- Headline result
- Screening completions up 200%+ year over year
- DataRecords, claims, check-ins, engagement
- StateScore each person's situation
- DecisionRank the few actions that matter
- ActionDeliver it where the person works or lives
- OutcomeLog what happened; refine the rules
Outcomes
- Komen: constituents reporting a completed screening after our outreach grew more than 200% year over year (constituent surveys).
- Komen: a second engine for donors helped lift acquisition, retention, and revenue to 12-year highs.
- iCAD: AI moved from showing findings to triaging the radiologist's worklist, across 4,000+ provider sites.
- SigmaFlow: a separately priced predictive module helped drive 10x revenue growth and an exit.
Komen: next steps for patients and navigators
The engine scored each patient's situation, including diagnosis stage, financial need, engagement, and steps completed, and pushed the next step to the navigator console and the patient portal. It was rules-based, and we logged every state, action, and outcome so it could evolve toward machine learning.
| When the data showed | Next best action |
|---|---|
| A newly diagnosed patient's first contact | Plain-language education on their diagnosis |
| Trouble paying rent, transportation, or utilities | Financial assistance, with the application prefilled |
| An application started but stalled | Outreach in the patient's preferred channel |
| Barriers beyond money | A navigator who sees the full context first |
| A clinical profile matching open studies | Clinical trial matching |
| Due or overdue for screening | A reminder and help finding affordable screening |
CareTogether: the right alert to the right person
AI check-ins with seniors become prioritized signals, such as missed medications, refill needs, reported falls and fall risk, changes in routine, and low mood. Each goes to the right family member or care team member, with pharmacist escalation for medication issues and immediate crisis routing when someone needs help now.
iCAD: from findings to decision-ready triage
Early on, the AI marked lesions and scored each case. That was useful, but it didn't change what radiologists did. The unlock was making it decision-ready: comparing each exam with the patient's priors, flagging lesions that had grown beyond an acceptable range, and triaging the worklist so radiologists read the highest-priority cases first. The AI stopped being information and became the next action.
SigmaFlow: machine learning that ranked the next move
For energy operators, machine learning models ranked the next best action for each operating asset against production, revenue, operating capital, and lease deadlines, with feature-level explanations so teams could trust the recommendation. The module was priced separately and won competitive deals.
How I build it
- Name the decision and who owns it before choosing a model.
- Start with transparent rules, log every outcome, and earn the right to use machine learning.
- Put the action where people already work, not in another screen.
- Recommend a few actions that matter, not everything that could be done.