People-Plus-Machine Allocation
Assign judgment to people and scalable optimization to machines
- Difficulty
- Expert
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 99%
People-plus-machine allocation begins by decomposing a complex decision into components and assigning each component to the actor with the stronger comparative advantage. Human experts handle idea generation, nuanced interpretation, synthesis, and situations where evidence is sparse or company-specific. Machines handle repetitive and complicated mathematics, behavioral-bias correction, portfolio construction, and trade execution. An integration layer then converts human judgments into structured inputs, combines them with data and algorithms, and returns actionable outputs. The goal is not to have technology rubber-stamp people or people override every algorithm. It is to design a coherent operating system in which complementary capabilities improve the final decision. Because both technologies and human practices evolve, the division of labor must be reviewed rather than treated as permanent.
Origin
Matthew Granade described Point72's systematic best-ideas book and its people-plus-machine thesis on Masters in Business.
Core principles
- 01Allocate work according to comparative advantage rather than ideology.
- 02People excel at nuance, synthesis, thin data, and idea generation.
- 03Machines excel at repetitive mathematics, bias correction, construction, and execution.
- 04The strongest system combines complementary outputs into one process.
- 05Reassess the boundary as capabilities change.
How to run it
- 1
Decompose the decision
Break the end-to-end decision into idea generation, interpretation, calculation, construction, execution, and review activities.
Pro tip Map information handoffs as well as tasks.
Watch out Automating an undifferentiated process usually preserves its hidden weaknesses.
- 2
Map human advantages
Assign nuanced interpretation, synthesis, thin-data reasoning, and context-heavy judgment to qualified people.
Pro tip Use domain experts rather than generic reviewers.
Watch out Human judgment is valuable but remains vulnerable to inconsistency and bias.
- 3
Map machine advantages
Assign repetitive mathematics, optimization, bias correction, construction, and execution to algorithms.
Pro tip Prioritize high-frequency activities with explicit objectives.
Watch out Machines can optimize the wrong target with great efficiency.
- 4
Engineer the handoffs
Structure human insights so models can consume them and present machine outputs so people can interpret and challenge them.
Pro tip Preserve the provenance of every important input.
Watch out Poor handoffs can erase the advantage of both participants.
- 5
Evaluate the combined system
Measure the quality, consistency, speed, and risk of the integrated process rather than grading humans and machines in isolation.
Pro tip Compare the hybrid system against human-only and machine-only baselines.
Watch out Do not attribute gains to collaboration without controlled comparisons.
- 6
Rebalance over time
Periodically revisit which activities belong to people and which belong to machines as evidence and capabilities change.
Pro tip Use observed failure modes to guide reallocation.
Watch out A sensible allocation today may become obsolete as technology advances.
In the wild
Discretionary investors study companies, meet management teams, and form nuanced views. Their behavior and ideas become inputs alongside external datasets. Algorithms then help select, size, construct, and execute a systematic best-ideas portfolio.
→ The combined book preserves company-level human insight while applying machine consistency and scale.
Experienced analysts interpret management credibility and unusual industry conditions, while a model standardizes ratios, tests scenarios, flags bias, and recommends exposure limits. Reviewers see both the recommendation and its supporting evidence.
→ Credit decisions gain consistency without discarding context that is difficult to encode.
Common mistakes
Automating the whole decision
End-to-end automation can discard valuable nuance when data is thin or circumstances are unusually specific.
Using machines as decoration
A model adds little value when humans can routinely ignore it without explanation or when it merely confirms existing opinions.
Ignoring handoff design
Even strong human and machine components fail when their inputs, outputs, and authority are not clearly connected.
Is it for you?
Best for
It is best for consequential recurring decisions that combine ambiguous qualitative evidence with large-scale calculation or optimization.
Not ideal for
It is not ideal for simple tasks where one reliable human or automated method already solves the problem efficiently.
From the transcript
“one of the kind of key questions we're always asking is what are the people best at and what are the machines best at?”
“the people are very good at the nuanced situation, at the idea generation, at the interpreting the thin data at the synthesis UM. And that…”
“what we're trying to do is figure out how you marry those two up in a really smart way”
From the episode
Matthew Granade Discusses Quantitative Investment (Podcast)
Matthew Granade