Mar 1, 2026 · Lab notes · 7 min
Automation in investment research
Automation can assist investment research, but the evidence in this article does not establish that AI causes superior fund returns. Numerai’s November 20, 2025 financing announcement reports a $30 million equity financing at a $500 million company valuation. BlackRock’s October 2, 2025 announcement describes an AI-enabled portfolio commentary tool. These are different activities, and neither establishes an autonomous fund’s investment advantage.
The useful question is which parts of research can be automated while preserving the evidence, uncertainty and decision authority needed to assess a recommendation. Collecting documents, drafting a memo, constructing a portfolio and executing a trade are separate operations. A capability demonstrated in one does not establish reliability in the others.
What a performance comparison requires
A useful return comparison must align the measurement period, strategy, risk and costs. Half-year, calendar-year and since-inception results are not directly comparable. Funds investing in AI-related businesses do not establish that AI made their investment decisions. The table below sets out the evidence a defensible evaluation requires.
| Comparison dimension | Evidence required | Why it matters |
|---|---|---|
| Period | Identical start and end dates | Half-year and since-inception returns are not directly comparable |
| Fees | Consistent gross or net basis | Different fee treatment changes the result |
| Exposure | Leverage, concentration and factor exposures | A return difference need not come from the research method |
| Strategy | What the system actually automates | Investing in AI is distinct from using AI to invest |
| Selection | Full sample and failed experiments | Selected winners do not establish class performance |
| Assets | Dated AUM, financing and capacity separately | Company valuation is not managed capital |
| Verification | Identified primary record and methodology | A marketing statement is not an independent audit |
An “AI-driven” label is not a consistent research category. It may describe conventional statistical models, document analysis, an analyst assistant or delegated execution. To compare investment processes, identify the actual task, the human intervention and the decision rights.
| Evidence category | What it can establish |
|---|---|
| Financing announcement | A company-reported financing event |
| Product announcement | A described capability and named implementation |
| Performance record | Results for its stated period and methodology |
| Controlled evaluation | Results within the tested task and conditions |
Our proposed evaluation starts with a matched research task. Give the automated workflow and a simpler baseline the same information available at the decision time. Score factual accuracy, source coverage, omitted contrary evidence and review effort. Investment performance requires a separate test with exposures, costs and an out-of-sample period specified in advance. A more polished memo is not evidence of investment alpha.
An illustrative research pipeline
A research pipeline can separate document retrieval, extraction, analysis, review and portfolio decisions. At each boundary, retain the source and its availability date. A later filing or revised dataset must not silently enter an earlier decision record. This is a proposed architecture, not a claim that a particular staffing level or cost saving has been observed.
The diagram below illustrates that design. A retrieval stage gathers permitted material. Analysis stages prepare claims and supporting passages. A risk review checks the proposed decision against explicit limits. An authorized human decides whether it may proceed.
graph TB DATA["Market Feeds · News · Alt Data"] --> R["Research Agents
Source review"] R --> A["Analysis Agents
Signal extraction"] A --> K["Risk Agents
Exposure limits"] K --> P["Portfolio Construction"] P --> APPROVAL{"Authorized human approval"} APPROVAL -->|"approved"| T["Trade Execution"] APPROVAL -->|"revise or reject"| A T -.->|"performance data"| R HUMAN["Authorized human team"] -.->|"objectives &
constraints"| K HUMAN -.->|"decision"| APPROVAL style R fill:#e8e8e8,color:#0a0a0a,stroke:#0a0a0a style A fill:#e8e8e8,color:#0a0a0a,stroke:#0a0a0a style K fill:#e8e8e8,color:#0a0a0a,stroke:#0a0a0a style HUMAN fill:none,stroke:#0a0a0a
Tool and agent protocols can carry requests and results between components. They do not by themselves retain a complete research record, enforce a portfolio mandate or establish regulatory compliance. The application must implement permissions, records and review. Access to a brokerage connection is a technical capability, not standing authority to trade.
Numerai’s announcement separately reports approximately $550 million in assets under management and capacity of up to $500 million from J.P. Morgan. These company-reported figures describe different concepts from its financing and valuation. Capacity is not proof that the full amount was funded.
The incumbents respond
A documented institutional example is useful when its scope is clear. Portfolio commentary supports communication about analytics. It should not be described as autonomous security selection, portfolio construction or execution unless the source establishes those capabilities.
BlackRock’s announcement says Aladdin Wealth’s Auto Commentary converts portfolio analytics into commentary and names Morgan Stanley as its first implementation. It is dated October 2, 2025. This supports a narrow claim about an advisor tool.
A review of any claimed institutional deployment should identify the responsible organization, the task, the deployment date and the evidence of actual use. Announced plans, pilots and production implementations belong in separate categories. A firm-wide technology budget does not measure the resources or results of one AI workflow.
The same distinction applies to retail products. An AI-themed fund invests in a theme; an AI-assisted research tool helps prepare information; a discretionary portfolio service makes investment decisions under a mandate. These categories carry different claims.
The regulatory reckoning
The SEC’s 2026 examination priorities, issued in November 2025, include emerging technology and AI. They discuss review of automated investment tools and the accuracy of firms’ representations about AI. These are the published priorities for that examination year, not a certification of any product or an assertion that the document remains the latest edition.
The SEC’s 2019 investment-adviser fiduciary interpretation describes duties of care and loyalty. The scope of an adviser’s obligations depends on the relationship and applicable law. Automating part of the workflow does not demonstrate that those obligations have been met.
FINRA’s 2026 Regulatory Oversight Report discusses agentic AI risks, including authority, supervision and controls. Its scope is member-firm oversight; it is not interchangeable with the SEC’s adviser framework. A research implementation should document who can approve an action and how exceptions are reviewed.
| Regulatory Body | 2026 Focus | Key Concern |
|---|---|---|
| SEC | Automated investment tools, AI representations | Fiduciary alignment of AI-driven recommendations |
| FINRA | Autonomous agent oversight | Scope of authority, human-in-the-loop protocols |
| SEC, 2019 interpretation | Investment-adviser fiduciary duty | Duties of care and loyalty in the advisory relationship |
Adoption surveys also need a defined population, question and date. Using an assistant in a research task is different from delegating a portfolio decision. Without those definitions, apparently conflicting adoption percentages cannot be reconciled. This article therefore makes no numerical forecast of industry-wide autonomous investment deployment.
The trajectory
Four questions organize a practical evaluation. They connect the investment hypothesis to the operating process without assuming that automation improves returns.
First, is the result comparable? Use the same observation period, fee basis and benchmark. Examine leverage, concentration and factor exposure before attributing a return difference to the research method. Retain unsuccessful experiments and specify the evaluation before inspecting the result. A selected winner cannot establish how a whole class of systems performs.
Second, what does an accepted result cost? Count data licenses, model calls, engineering, monitoring and human review. Track corrections and rejected outputs, not only successful demonstrations. A hypothetical reduction in drafting time becomes an operational saving only if the completed work meets the same quality standard.
Third, is the record reproducible? Retain the source document, retrieval time, relevant passage, transformation and model or workflow version. Preserve earlier records when a claim changes. Protocols can simplify integration, but providers still differ in data semantics, licensing, permissions and error behavior.
Fourth, who has decision authority? Keep research access separate from trading permissions. State the mandate, risk limits and approval requirements before enabling execution. A reader should be able to distinguish a generated recommendation, an approved order and an executed transaction in the record.
The thesis: automation is worth testing where it improves research coverage or reduces the cost of producing a verifiable result. That is a workflow hypothesis. It is not a conclusion that an autonomous fund will outperform or that human investment responsibility can be removed.
The next useful evidence is a bounded, reproducible evaluation: a defined task, a baseline, a source record, measured errors and a review decision. Keep simulated results distinct from live performance. Until those tests support a stronger conclusion, describe the system by what it demonstrably does.
Factual review: October 26, 2026. Original publication date retained. Primary sources are linked alongside the revised claims. Removed unsupported performance, adoption, staffing, deployment and enforcement claims. The architecture is presented as an illustrative design. Company financing, capacity and assets under management are distinguished.