Thesis
Prediction answers what is likely. Decision Intelligence answers what should we do — under cost, risk, capacity, and incomplete information.
Heuman Labs builds DI as a bridge across machine learning, causal reasoning, and managerial decision science. The artifact is not a chart. It is a policy: a system that ranks interventions, simulates counterfactuals, and — where appropriate — automates the next best action with an audit trail executives can defend.
Beyond prediction
Traditional predictive analytics stops at the score. Decision Intelligence continues into intervention design: which lever, for whom, at what intensity, subject to which constraints.
That requires models that handle uncertainty, counterfactuals, and optimization — not only supervised ranking. We select the family by the decision shape: heterogeneous treatment effects when offers differ by person; sequential policies when today's action changes tomorrow's state; multi-criteria frameworks when trade-offs are political as well as numerical.
Causal & what-if models
Causal models estimate how a specific business action changes an outcome — not merely how outcomes co-occur with features. They are the backbone of credible “what if we intervene?” analysis.
Causal Forests (CF)
Does — Heterogeneous treatment effects — how an individual responds to an intervention.
Best at — Personalizing offers, outreach, or pricing when average A/B lifts hide winners and losers.
Structural Causal Models (SCM)
Does — Cause-and-effect pathways via graphs and structural equations; counterfactual scenarios.
Best at — Executive what-if design when levers interact and naive regression misleads.
Probabilistic & uncertainty models
Decisions are made before certainty arrives. Probabilistic models map states, dependencies, and risk so policies degrade gracefully instead of failing silently.
Bayesian Belief Networks (BBN)
Does — Joint and conditional probabilities across variables linked by dependence structure.
Best at — Risk and diagnostic decisions where evidence arrives piecemeal and beliefs must update.
Markov Decision Processes (MDP)
Does — Optimal action sequences when outcomes are partly random and partly controlled.
Best at — Lifecycle and retention policies that unfold over time with state-dependent rewards.
Prescriptive & optimization models
Prescriptive models recommend — or directly select — actions. They often fuse learned value estimates with classical optimization under hard constraints: budget, inventory, headcount, SLA.
Reinforcement Learning (RL)
Does — Policies that maximize cumulative reward through trial-and-error in an environment.
Best at — Sequential personalization and control loops where the system learns from its own actions.
Mixed-Integer Linear Programming (MILP)
Does — Optimal allocations with integer and continuous decision variables.
Best at — Resource routing, capacity planning, pricing packs, and constrained campaign assignment.
In practice we frequently compose: a propensity or uplift model estimates value; MILP or constrained bandits assign scarce interventions without violating business rules.
Strategic multi-criteria models
Not every executive decision collapses to a single KPI. When alternatives must be ranked on conflicting quantitative and qualitative criteria — margin vs. brand vs. regulatory comfort — we deploy structured evaluation.
Multi-Criteria Decision Analysis (MCDA)
Does — Ranked or selected alternatives under weighted, often conflicting criteria.
Best at — Board- and ELT-level choices where transparency of trade-offs matters as much as the pick.
Propensity × Decision
The highest-leverage systems couple both focus areas. Propensity models estimate who is likely to act. Decision Intelligence estimates which action changes that path and how to allocate limited interventions across the population.
Example pattern: churn propensity ranks risk → causal forests estimate uplift of three retention treatments → MILP assigns treatments under budget and agent capacity → MDP / RL refines sequencing as customer state evolves → agents execute the assignment across CRM and messaging. The enterprise receives a closed loop, not a spreadsheet of scores.
Enterprise engagement
We propose DI engagements that begin with a single high-stakes decision surface — pricing, retention, lead routing, inventory intervention — and expand only after the first policy is measurable in production.
- 01Name the decision, owner, constraints, and the counterfactual you refuse to ignore.
- 02Inventory data and experimental history; choose causal vs. probabilistic vs. optimization primary.
- 03Prototype a shadow policy that logs recommended actions without forcing them.
- 04Validate lift against holdouts or stepped rollout; publish uncertainty, not just point estimates.
- 05Automate with human gates where accountability requires it; instrument drift and regret.
- 06Extend the policy library once operators trust the first loop.
