Capital Allocation in the Age of Autonomous Systems

The Shifting Frontier

Markets have always attracted automation. From telegraph-enabled stat arb to HFT, the pattern is the same: whatever work is systematic enough to codify will eventually be automated, and the returns to that work will compress to the cost of infrastructure.

The current wave is different in scope if not in kind. AI systems are now capable of automating not just execution but analysis, synthesis, and — at the margins — judgment. The question isn't whether this changes capital allocation. It's how.

What Gets Automated

The clearest candidates:

Systematic signal generation. Pattern recognition across large datasets is pure ML territory. The edge in factor investing has been compressing for years. Expect it to compress further and faster.

Document processing. Reading, extracting, and normalizing information from filings, transcripts, and reports is being automated now. The advantage of being faster at this will diminish.

First-pass screening. Opportunity triage — is this worth more time? — is increasingly tractable with AI. The cost of reviewing large deal or investment universes drops.

What Remains

The more interesting question is what doesn't automate cleanly.

Judgment under genuine uncertainty. Markets are adversarial. When everyone has the same information processed by similar models, the edge moves to interpretation — which requires a model of what other participants believe, which requires a kind of social reasoning that current AI doesn't do well.

Relationship and access. Private markets still run on trust, access, and judgment that is slow to formalize. The information advantage in private credit or venture isn't primarily about processing speed.

Novel frameworks. When the world changes in a way that historical data doesn't capture, the models fail. Building the new framework — recognizing that the old one no longer applies — requires the kind of flexible intelligence that remains distinctly human.

The Allocation Opportunity

There's a meta-point here that gets less attention: the infrastructure buildout for AI in finance is itself a capital allocation opportunity. The picks-and-shovels bet on AI adoption — compute, data, tooling — is large and still early.

More specifically: the firms building proprietary datasets, domain-specific models, and AI-native workflows will operate at structural cost and quality advantages that compound over time. The investment opportunity is in identifying which firms are actually building durable infrastructure versus which are doing AI theater.

Conclusion

Human capital allocators will remain relevant, but the nature of the work will shift. The commodity parts — processing, screening, systematic signal generation — will be automated. What's left is the harder stuff: judgment, relationships, novel frameworks, and the meta-work of choosing what to automate and how to design the systems that do it.

That last part — system design — may be the most underrated skill in finance for the next decade.