AI and Proprietary Data: Is It Really a Moat for a Fund Platform?

The claim that proprietary data is a moat for a fund platform is repeated far more often than it is tested. Model capability is converging and repricing downwards, so the durable advantage is said to sit in the operational data a platform accumulates. That logic is sound about the model layer and careless about the data layer. Most of the relevant data does not belong to the platform, cannot lawfully be reused without a defensible basis, and sits under contracts drafted before anyone asked who owns a derived model. This article sets out where the data sits, what the contracts decide and what allocators ask.

The temptation is to treat AI as a clever answer machine and feed it everything. In a regulated fund business, the discipline is the reverse: keep the data, govern it, verify the output, and let the institution learn from its own work. That is where a real moat forms. David Lloyd, Chief Executive Officer at CV5 Capital

Executive Summary

A proprietary data advantage in fund operations is real but narrow. It exists where a firm holds a record it generated itself, has the right to reuse it, and can evidence that reuse to a regulator and an allocator. Outside that intersection it is rhetoric.

• Model capability is converging and getting cheaper, so the model layer is not where advantage sits.
• Most operational data is held by the administrator, the manager or a vendor, not by the platform calling it proprietary.
• Ownership of the derived model turns on a few contract terms, and silence favours the vendor.
• The Cayman Islands Data Protection Act (2021 Revision) applies the data protection principles by section 5(1) and holds the controller responsible for processing carried out on its behalf.
• Regulators are converging on the view that existing accountability and outsourcing obligations already cover AI.

The Model Is Rented and Most of the Data Is Borrowed

The strongest version of the argument is straightforward. Model weights are available to anyone who pays, several laboratories produce comparable output, and the cost of a given capability keeps falling. Value migrates up the stack towards data, workflow and the feedback loop.

The argument is careless about the data layer, because it treats data as undifferentiated material the firm happens to hold. Each record has an owner, a collection purpose, a confidentiality regime and a contract governing its use. The argument also assumes accumulation equals learning. Past decisions improve future output only if competent people reviewed them and both corrections and errors were captured usably. Absent that, an unreviewed loop industrialises existing habits.

Where the Data Actually Sits

The operational record is distributed across at least four parties. The investment manager holds the strategy and the trading rationale. An independent fund administrator holds the books and records, the investor register and the valuation trail. Software vendors hold whatever passes through their systems.

The final row rarely gets the attention it deserves. Corrections a compliance officer makes to a draft are the highest value signal a firm generates, because they encode judgement rather than facts. If that signal accrues inside a third party system on terms never negotiated, the firm is annotating someone else's product.

What the Contracts Decide

Ownership of the learning is a drafting question before it is a technology question. Four terms carry almost all of the weight, and in standard vendor paper each favours the vendor or is left silent.

An enterprise privacy setting substitutes for none of these. A setting is a configuration the vendor may change; a clause is a promise the firm can enforce. The same applies inside the structure: a platform intending to derive value from patterns across its funds must say so in the delegation and services agreements.

The Data Protection Constraint Is Not a Formality

Investor data is personal data. Under the Cayman Islands Data Protection Act (2021 Revision), personal data means information relating to an identified or identifiable living individual. Section 5(1) imposes the data protection principles in Part 1 of Schedule 1. Section 5(4) requires a controller to comply with those principles and to ensure compliance for personal data processed on its behalf. Delegation does not move the obligation.

Four principles bear directly on model development. The second requires that personal data be obtained only for specified lawful purposes and not further processed in any incompatible manner. Data collected to satisfy anti-money laundering obligations was not collected to train a model, so the compatibility analysis must be done rather than assumed. The third requires data to be adequate, relevant and not excessive. The eighth restricts transfer to a country that does not ensure an adequate level of protection, making processing location a legal question rather than a procurement preference.

Where a processor is engaged, Schedule 1 Part 2 requires a contract obliging it to act only on the controller's written instructions and to take appropriate security measures. Section 12 addresses automated decision making: a controller must notify a data subject where a decision significantly affecting that subject is taken solely on automated processing, and the subject may require reconsideration. Under section 55 the Ombudsman may impose a monetary penalty not exceeding CI$250,000.

The practical consequence is that "the data is inside the model, so it is gone" is not a defence. A firm relying on legitimate interest must show the interest is real and articulated, the processing necessary, and the balancing exercise against data subjects' rights performed.

Model Governance and the Regulatory Posture

Financial regulators have largely declined to write a separate rulebook for AI, which is frequently misread as permission. The Financial Conduct Authority states that its approach is principles-based and focused on outcomes, and that it does not plan to introduce extra regulations for AI, pointing instead to the Senior Managers and Certification Regime. An accountable individual owns the outcome and cannot attribute a failure to a tool.

In the Cayman Islands the operative measure is the CIMA Rule and Statement of Guidance on Internal Controls, effective 13 October 2023. A regulated entity may rely on a service provider's system of internal control over outsourced activities, but only where the governing body is satisfied and can demonstrate to the Authority that the system meets the Rule. The Rule also requires information technology systems to be secure, independently monitored and supported by adequate contingency arrangements.

The IOSCO report "Artificial Intelligence in Capital Markets: Use Cases, Risks, and Challenges", published on 12 March 2025, lists model and data considerations, concentration and third party dependency, and human interaction with AI systems among commonly cited risks. Concentration is the one firms describing a data moat most often ignore, because a loop running on a single provider has a single point of failure.

The SEC Division of Examinations, in its fiscal 2026 priorities, said it will assess whether firms have adequate policies and procedures to supervise their use of AI technologies, including for fraud detection, back-office operations, anti-money laundering and trading. It will also review the accuracy of representations about AI capabilities. On 18 March 2024 it charged two investment advisers over false and misleading statements about their use of artificial intelligence, settled for US$225,000 and US$175,000.

What an Allocator's Operational Due Diligence Will Ask

Diligence teams have moved from asking whether a firm uses AI to asking where it sits in the control environment. The weak answers follow a pattern.

Very little of this is about artificial intelligence. It is outsourcing and change control diligence applied to a new dependency.

So Is It a Moat?

Sometimes, and less often than claimed. The advantage is real where four conditions hold:

• the firm generated the data by doing the work rather than receiving it;
• it holds a written right to reuse that data for the purpose in question;
• the reuse is lawful and consistent with confidentiality owed to investors and managers; and
• the loop is closed by people competent to correct the output, with corrections retained usably.

Where all four hold, the advantage compounds, because a competitor cannot buy a history it did not live through. Where one fails, what remains is a capability any competitor can license within a quarter. The uncomfortable observation is that the most useful categories are the most encumbered. Governance decisions and structuring precedent are the most defensible; investor files and position data, more voluminous and more tempting, are the most constrained.

One limit should be stated plainly. CV5 Capital provides regulated platform infrastructure, governance and operational coordination for third-party investment managers. It is not the investment manager of the strategies operated on the platform, so the value of any operational learning is measured in consistency, control quality and time to market, never in performance. See our notes on the institutional fund stack, on launching under one regulated platform, and on whether an autonomous agent can be an investment manager.

Key Takeaways

• Map every data domain to its owner, collection purpose, confidentiality regime and contract before calling it proprietary.
• Read the input licence, output licence, derived data clause and sub-processing terms in every AI vendor contract.
• Document a lawful basis before reusing AML or onboarding data to develop a model.
• Keep a person in the decision path for anything significantly affecting an investor, and retain the evidence.
• Build the AI inventory, the data flow map and the manual fallback now, and state AI capabilities only in terms you could evidence.

Frequently Asked Questions

Is proprietary data really a moat for a fund platform?

Only where the firm generated the data itself, holds a written right to reuse it, can reuse it lawfully, and closes the loop with competent human review. Governance and structuring precedent is the most defensible category; investor files and position data are larger but far more encumbered.

Can a fund use investor or AML data to train an AI model?

Not without analysis. Under the Cayman Islands Data Protection Act (2021 Revision), the second data protection principle requires personal data to be obtained for specified lawful purposes and not further processed in any incompatible manner. Data gathered for anti-money laundering purposes was not gathered to develop a model, so a compatibility assessment comes first.

Do financial regulators have specific rules for AI?

Largely they apply existing rules. The Financial Conduct Authority states that its approach is principles-based and outcomes-focused, and that it does not plan to introduce extra regulations for AI. In the Cayman Islands the CIMA Rule and Statement of Guidance on Internal Controls, effective 13 October 2023, already covers reliance on outsourced systems of control.

What will an allocator ask about AI in operational due diligence?

Which processes use it, whether any output reaches a NAV, filing or investor communication without sign-off, what data leaves the firm and to where, whether investor data trains any model, and what happens if the provider fails.

Can a model trained on personal data be treated as anonymous?

Not automatically. The European Data Protection Board's Opinion 28/2024, adopted on 17 December 2024, concluded that AI models trained on personal data cannot in all cases be considered anonymous. Anonymity requires that the likelihood of extraction from the model, and of obtaining data through queries, are both insignificant.