Thread
Causal ML
Estimating who would change their behavior, and turning that into a decision about who to contact.
Predictive models rank people by what they are likely to do. Causal ML estimates what would be different if you acted, which is a harder question and usually the one that matters. It spans two disciplines on purpose: the estimation belongs to incrementality, the policy it feeds belongs to decisioning, and a model validated offline against historical outcomes has demonstrated neither.
Decisioning
-
Choosing the Right Campaign Optimization Event
The deepest event that stays reliable, timely, and frequent enough to train on. Alignment sets a ceiling you cannot buy your way past, precision decides how close you get, and only one of them improves with volume.
-
The Case for Doing Nothing
Contact fatigue, capacity, and cost mean no-action is a real treatment arm. Why it belongs in every decisioning system by default, and why organizations reliably leave it out.
-
Uplift Modeling in Practice
Where to cut the ranking, how much to withhold, and why the offline number is larger than the money. Estimating heterogeneous treatment effects when you have a randomized holdout, and validating the policy prospectively.
-
Why Agentic Marketing Still Needs Holdouts
Two agents, the same budget and the same population, differing only in what they are scored on. One improves its dashboard by 65% while destroying value, and the one doing the right thing looks worse.
- Worked example
Broad Nurture, Propensity, or Uplift?
One trial-to-paid nurture sequence, four targeting policies, and a randomized holdout. Compared on incremental conversions rather than model accuracy, which is the only comparison that changes the answer.
Systems
- System
Next-Best-Action Policy
Turning scores into decisions under real constraints: capacity, contact fatigue, margin, cooldowns, and the explicit option of doing nothing.
- System
Treatment Effects and Policy Evaluation
The bridge between a campaign's average effect and a per-customer decision: estimating how the effect differs by customer state, and scoring a candidate policy on randomized or logged data before it is trusted with volume. The component that produces the numbers next best action consumes.
- System
Policy Governance and Safety
Who may change the policy, what it may do, what stops it without a human, and how anyone knows whether it worked, whether the policy is a rules engine, a model, or an agent. The agentic case is the hardest version of the same component.
- System
Designs When Randomization Is Refused
No sales leader will accept a do-not-touch list on named accounts. What can be assigned is the marketing program itself, with sales left alone in both arms, and after that the territory, the partner, and the order of a staged rollout.
- System
Rep Capacity and Account Prioritization
Which accounts get a rep's hour, which is the scarcest allocation on the list. Scoring is easy, and the constraint that decides whether it changes anything is how many hours exist.
- System
The Prioritization Log
Which accounts were ranked, which were assigned work and at what probability, and what was actually done, with a two-sided exploration band around the cutoff so that both working and not working have support near the line. The record that lets the next ranking be evaluated before it ships.