By: Eric Swick
Private fund managers are increasingly using artificial intelligence platforms such as Anthropic’s Claude and OpenAI’s ChatGPT to support investment research. Portfolio managers and analysts may use these tools to review issuer filings, compare companies, summarize market information, test investment ideas, write code, and assist with quantitative analysis.
At the same time, the same platforms can be used to draft investor communications, support compliance, prepare internal memoranda, and perform other administrative work. This creates an important question for managers considering the use of soft dollars: How should the cost of a general-purpose AI platform be treated when only some of its use relates to investment research?
When Can AI Be Paid with Soft Dollars?
Section 28(e) of the Securities Exchange Act of 1934 provides a safe harbor that permits an investment manager to use client commissions to obtain eligible brokerage and research services under certain conditions.
Under the SEC’s Section 28(e) guidance, the product or service must constitute an eligible brokerage or research expense, provide appropriate assistance in the manager’s investment decision-making and order execution, and be reasonable in relation to the value of the brokerage and research services received. The manager must make this determination in good faith, and its best execution obligations continue to apply.
The analysis should focus on what the AI platform is being used to do. Potentially eligible research uses may include:
- Analyzing securities, issuers, industries, and economic trends;
- Comparing companies and evaluating portfolio strategy;
- Synthesizing issuer filings, market data, and other investment information;
- Supporting portfolio analysis, investment models, and quantitative research; and
- Providing pre-trade market or execution-strategy analysis.
By contrast, marketing, investor relations, compliance, recordkeeping, human resources, and general administrative functions do not qualify for the Section 28(e) safe harbor. The fact that an investment professional uses the platform does not make every use investment research.
The Mixed-Use Challenge
General-purpose AI platforms will often be mixed-use products. A portfolio manager might use Claude to analyze a bank’s earnings in the morning, draft an investor letter in the afternoon, and prepare an internal personnel review later that day. Only the first activity may qualify as investment research.
The SEC’s guidance requires managers to make a reasonable allocation when a product or service serves both eligible and ineligible functions. The manager must maintain adequate books and records supporting the allocation and pay for the ineligible portion with its own funds.
This is where AI differs from a traditional research subscription. The service is not necessarily tied to one publication, dataset, or research function. Its purpose can change from one prompt to the next. An allocation based only on employee title, the number of licenses assigned to investment personnel, or a general estimate of time spent may not fully capture how the platform is used. While the allocation does not require mathematical perfection, it must be reasonable, documented, reproducible, consistently applied, and executed in good faith. Managers should also revisit the methodology as usage changes and more reliable information becomes available.
Given these challenges, some managers may choose to have the adviser pay the full cost during a limited pilot period. This allows the firm to understand how the platform is actually being used, confirm what activity and cost data are available, develop its classification and review procedures, and coordinate payment mechanics before client commissions are used. Once the methodology is operating effectively, the manager can consider using client commissions for the portion that qualifies under Section 28(e), while separately determining whether any fund-related use may be paid directly by the fund under its governing documents. An adviser-paid pilot is not required, but it may be a practical approach where the cost is modest and the allocation methodology is still being developed.
A Practical Example: Using Claude Data to Support the Allocation
One hedge fund manager recently considered whether Claude’s administrative data could support a mixed-use allocation. The portfolio managers, analysts, and traders used the platform for both investment research and other functions. The question was whether the firm could identify the eligible research portion with enough precision to support payment through soft dollars.
Anthropic offers different reporting tools depending on the Claude product and plan. Its Compliance API can provide authorized Claude Enterprise customers with access to activity records and, with the required access and permissions, chats and session content. Its Analytics APIs provide aggregated usage and cost information, including per-user reporting for certain Enterprise plans. For Claude Platform API activity, the Usage and Cost Admin API can report usage by time period, API key, or workspace, while cost reporting is generally provided on a daily basis.
Depending on the available data, a manager could develop a process that:
- Identifies the user, workspace, conversation or session, and timestamp;
- Classifies the activity as eligible research or non-research based on written criteria;
- Matches the classification to available usage or cost data; and
- Calculates the portion of the invoice eligible for payment with client commissions.
A manager could use a rules-based or AI-assisted tool to perform the initial classification. For example, prompts involving issuer analysis, security comparisons, or portfolio modeling could be placed in the research category, while prompts involving investor communications, compliance, or personnel matters could be placed in the non-research category.
The classification tool should not make the final legal determination. Compliance or another designated reviewer should test samples, resolve exceptions, document overrides, and apply a conservative treatment where the purpose cannot be supported. A conversation that begins with issuer research and later moves to investor reporting may need to be divided rather than classified as a single activity.
The manager will also need a reasonable convention for matching activity to cost. Where one usage interval includes several prompts, the allocation might be based on relative token usage, the number of prompts or conversations, or another consistently applied measure. Dedicated research workspaces, projects, or API keys can make the process easier and reduce the number of judgment calls.
Payment Mechanics Matter
The broker’s payment process should be addressed before the arrangement begins. In the example above, the broker could pay Anthropic directly but could not reimburse the manager for the research portion of an invoice the manager had already paid. The broker instead contemplated paying the vendor only for the approved research portion, while the manager simultaneously paid the vendor directly for the remaining ineligible portion using its own funds. Managers should confirm how the broker will pay the vendor, how the mixed-use allocation will be communicated, and how the split payments will be coordinated with the vendor. The workpapers should reconcile the full vendor charge to the amount paid with client commissions and the amount paid by the manager.
There is also a separate fund-expense question. A use that does not qualify under Section 28(e) may still be payable directly by a fund if the fund documents authorize the expense, the use benefits the fund, and the cost is allocated fairly. For example, fund documents may permit certain research, technology, investor communication, or fund-related marketing expenses. If so, that portion should be separately identified and supported as a direct fund expense. A non-soft-dollar cost should not automatically be treated as either a fund expense or an adviser expense without reviewing the governing documents and the purpose of the use.
Information-Security and Privacy Considerations
Using conversation content to support an allocation creates additional risks. Prompts and outputs may contain material nonpublic information, personal information, proprietary models, or privileged communications. Managers should determine who may access exported content, how long the information will be retained, and whether the review process creates an unnecessary copy of sensitive data. Managers may also wish to use metadata, dedicated workspaces, or other less intrusive information where it provides sufficient support. Any automated classification process should be covered by the firm’s AI governance, access controls, vendor diligence, and record-retention framework.
Practical Considerations for Private Fund Managers
Before implementing a soft-dollar arrangement for an AI platform, managers should consider the following:
- Consider beginning with a limited adviser-paid pilot while usage patterns and allocation controls are developed;
- Define eligible and ineligible use cases in writing before beginning the allocation;
- Use dedicated seats, workspaces, projects, or API keys where practical;
- Determine what activity, usage, and cost data are actually available under the firm’s product and plan;
- Establish how ambiguous or mixed conversations will be treated;
- Test the classification results and document reviewer overrides;
- Confirm the broker’s vendor-payment and reimbursement process;
- Reconcile each invoice to the soft-dollar and non-soft-dollar amounts;
- Review whether any separately identified amount may be charged directly to a fund under its governing documents; and
- Update soft-dollar, expense-allocation, AI, privacy, and recordkeeping policies as needed.
Takeaway
AI platforms may provide valuable investment research, but their multi-use nature makes a blanket soft-dollar classification difficult to support. A use-based allocation supported by available activity and cost data is likely to be more defensible than an approach based solely on employee title or license count.
Managers considering this approach should start with clear classification rules, understand the reporting capabilities of the platform, confirm the broker’s payment mechanics, and maintain workpapers showing how each invoice was allocated. As AI use evolves, the methodology should evolve with it.