How should humans and AI systems divide decision-making tasks?

Puranam treats "Human–AI Collaborative Decision-Making" (HACD) as a problem for organization design rather than machine learning: the human-algorithm pair is "an organization—i.e., a multi-agent, goal-oriented system" whose design consists of "choices about division of labor (task division and task allocation) and integration of effort." He distinguishes three task types: Type A, where "algorithms equal or outperform humans" (image or handwriting recognition); Type B, where humans outperform and "must remain in the hands of humans" (judging an applicant's integrity); and Type C, the more interesting case, "where despite no clear superiority of either human or algorithm, the combination through aggregation may outperform either alone." Type C rests on a feature specific to decisions and absent from physical production: accuracy "can sometimes be improved through pooling and error cancellation," via the wisdom-of-crowds intuition for continuous predictions and "Condorcet's jury theorem" for discrete ones.

The mechanism he gives for Type A/B gains is familiar — specialization, "not fundamentally different from the calculus of outsourcing or offshoring or gains from trade." The less familiar half is the learning dimension: because both humans and ML systems are "adaptive systems that change how they make decisions over time through learning from experience," HACD designs must also specify a learning configuration, not just a task allocation. He surveys mechanisms from organization-design research reinterpreted for HACD: specialization paired with group-level feedback risks "superstitious learning from false negatives and false positives," mitigated by common priors, vicarious learning, or stable personnel; unspecialized parallel decision-makers who only see aggregate feedback ("learning by participation") depend on how "contrarians" who voted against the majority shape future decisions; and serial versus parallel architectures produce different learning dynamics because serial designs censor the inputs later agents can learn from.

This typology sits in tension with When do human-AI teams actually outperform either alone?: that review finds decision-task combinations underperform the best of human or AI alone on average, while Puranam's Type C argument is precisely the claim that aggregation should sometimes beat either alone on decisions — the theorem gives a condition (independence, modest and roughly equal accuracy) that may be exactly what average field results violate. It runs parallel to Do humans and AI divide labor differently in knowledge work?, another typology of who does what in human-AI decision work, though Puranam's axis is task type and learning configuration rather than task- and method-choice. And it offers a conceptual label for what How should AI agents and humans divide research tasks? shows empirically — a Type B-leaning division where final decisions stay with humans even as proposal and revision are delegated.

The excerpt is explicitly a conceptual "Point of View article," not a validated framework: Puranam offers typologies as "basic building blocks" for a design space that "may not yet be completely or densely populated," and the learning-configuration models he cites use "adaptive reinforcement learning algorithms to simulate human decision-makers" rather than studying real human-algorithm dyads. He also flags that the Type A/B/C boundary is not fixed — it depends on "the prevailing state of technology" at a point in time, so a task's classification, and therefore the appropriate division of labor, can shift as algorithms improve. The implication is that organization design for HACD needs its own empirical research program, not a one-time taxonomy.