Two Sigma started training neural networks on equity-returns data when most US asset managers were still arguing about whether Bayesian regressions were too aggressive. A decade later, the same firms that resisted are now embedding machine learning into every layer of the portfolio-construction stack, from signal generation to execution. According to the SEC’s staff research on AI in asset management, the technology is now deployed across most large US managers, and the regulatory conversation is shifting from whether it should be allowed to how it should be supervised.
What “AI in portfolio management” really covers
The label spans four distinct workflows. Signal generation uses machine learning to find predictive patterns in market and alternative data. Portfolio construction uses optimisation models, sometimes ML-driven, to translate signals into target positions under risk and capacity constraints. Execution applies reinforcement learning and supervised models to choose how and when to place orders. Risk management uses ML to model tail behaviour, factor exposures and counterparty stress. Each layer has its own maturity curve. Each has different regulatory implications.
The vendor stack is also varied. Some managers build entirely in-house, with proprietary models and data feeds. Others rely on third-party platforms such as MSCI Barra, Bloomberg AIM, BlackRock’s Aladdin, or specialist vendors such as Two Sigma Plus or Numerai for specific layers. Cloud providers, particularly AWS, Azure and Google Cloud, are now central to the cost structure of mid-tier US asset managers running ML in production. The build-versus-buy decision shapes both the speed of iteration and the regulatory documentation burden.
Outside the institutional quant world, AI is also showing up in private wealth portfolio construction. Direct indexing engines optimise tax-aware rebalancing across thousands of household portfolios in ways no human team could match. Robo-advisor allocation models use clustering algorithms to assign clients to model portfolios. The volume of decisions made by ML systems on behalf of US retail investors is now substantially larger than the volume of decisions made by their human advisors, even if the strategic choices still flow through a person.
Where ML has clearly added value
The cleanest wins are in execution. Trading-cost analysis at large US asset managers consistently shows that ML-driven order routing produces measurable improvements in average implementation shortfall versus older execution algorithms. The 5-15 basis point gains are not glamorous, but on a $500 billion AUM book they cover the cost of the data science team many times over. Alternative data ingestion, where ML systems convert satellite imagery, credit-card panels and web scrapes into structured signals, is the second clear win. The buy-side use of these data has become commonplace enough that the original information edge has narrowed, but the signal still pays for the cost of acquisition in many strategies.
Portfolio optimisation under realistic constraints is the third clean win. ML-enhanced optimisers can handle non-linear cost functions, regime-aware risk models and large universes of instruments in ways the classical mean-variance approach cannot. The result is portfolios that more reliably hit the manager’s stated risk targets and that adapt faster when market conditions change. For long-only US equity managers running thousands of accounts in parallel, the time savings alone justify the investment.
Risk modelling is a quieter success. Modern risk systems use ensemble ML models to forecast factor exposures, default probabilities and liquidity stress under scenarios that linear models cannot capture. The wins here are not headline alpha, but operational: a risk team that can model 10,000 portfolio scenarios overnight in a way that would have taken weeks five years ago has a different conversation with its portfolio managers about hedging.
Where AI has consistently underdelivered
Direct alpha generation, the use case most often pitched, has proved harder than the marketing suggested. Equity factor returns are noisy, regime shifts break models trained on prior data, and the rate of overfit in academic studies is high. Several public quant funds that built their pitch around proprietary ML signals have produced returns indistinguishable from the broader factor benchmarks once fees and costs are accounted for. The honest read is that ML helps incumbent quant strategies hold their edge longer; it does not invent new alpha from nothing for the average buy-side firm.
The harder failure mode is over-trust in model outputs. Several US fund failures in the past three years can be traced to portfolio managers acting on ML signals without enough validation of model assumptions. The pattern is not unique to AI: every wave of quantitative innovation has produced similar episodes. The remedy is governance: documented backtesting, independent validation, kill-switch protocols and clear escalation paths when a model behaves outside expected bounds. The firms that have invested in those controls have largely avoided the failures that hit the firms that did not.
Macro forecasting is another category where AI has underdelivered. Most large language model-driven macro tools produce plausible narratives but inconsistent point forecasts. The systems that work best in this space combine ML with human macro judgement rather than replacing it. The pattern is similar to what every other industry has discovered about the shape of generative-AI productivity: the technology lifts experts, it does not replace them.
How US regulators are framing supervision
The SEC has taken a measured posture so far. The 2023 predictive-analytics rule proposal, since narrowed, framed the supervisory question around conflicts of interest: would an AI system trained on client-specific data place the firm’s interests ahead of the client’s? The asset-management industry pushed back hard on the original draft, but the principle is now well understood. Any AI system that touches client portfolios has to be auditable, explainable to the extent practical and free from configurations that systematically disadvantage the end client.
FINRA has separately issued guidance on supervision of AI-driven recommendation systems for retail broker-dealers. The Commodity Futures Trading Commission has flagged similar concerns for derivatives trading. The pattern across regulators is convergent: there is no special carve-out for AI, but there are several specific obligations around documentation, testing and ongoing monitoring that firms are expected to demonstrate. The firms that put audit, monitoring and bias-testing infrastructure in place before the regulatory ask have an easier path through exams than the ones bolting it on retroactively. The dynamic looks much like the one driving why fintech needs to build for federal employee benefits across other regulated fintech categories.
| Indicator | Date | Primary source |
|---|---|---|
| SEC proposed rule on conflicts of interest in broker-dealer/adviser use of predictive data analytics | July 26, 2023 | SEC Press Release 2023-140 |
| Federal Register publication of the proposed rule | August 9, 2023 | Federal Register |
| SEC withdrawal of the predictive-analytics proposal | June 17, 2025 | Federal Register |
Sources linked in the right column.
Where the next phase goes
Three forces will shape AI in US portfolio management over the next three years. Large language models will absorb more of the unstructured-data workload, which lifts the ceiling on alternative-data signal generation. Cloud-based ML platforms continue to lower the cost of model deployment, which pulls smaller US managers into capabilities that used to require a quant fund’s budget. And the maturing tokenisation conversation, visible in the tokenized US Treasuries market that reached roughly $7 billion in late 2025 data, will eventually let AI systems trade against richer on-chain order-flow signals than the current off-chain feeds provide.
The talent market is the other constraint. Senior quant researchers, ML engineers with finance domain knowledge and risk officers who can supervise model behaviour are all in short supply, and compensation has risen accordingly. Mid-tier managers that cannot match the comp packages of the top hedge funds and the big tech firms have responded by leaning more heavily on platform vendors, by partnering with academic groups, and by building development programs that grow talent internally. None of those options is a perfect substitute, but together they have kept the field from concentrating entirely into the top five firms.
By the end of 2026, the question for most US asset managers will not be whether to use AI but where it sits in the chain of responsibility. The firms that integrate ML carefully, with clear human-in-the-loop checkpoints and an audit trail their regulators recognise, will compound the operational gains. The firms that treat AI as a marketing layer will continue to produce returns indistinguishable from their benchmark and quietly explain to clients why the next year is different. The technology has earned its seat at the table. What it has not done is replace the experienced human judgement that still anchors every successful US portfolio management franchise.
