The FCA’s AI Findings: What They Mean for Private Fund Managers

On October 19, 2026, the Financial Conduct Authority (FCA) published AI in financial services: good practice and areas for improvement. Drawing on supervisory work, industry engagement, and testing, the examples illustrate how the existing framework applies to AI, including software used only for internal operations. They introduce no new regulatory requirements.

For private fund managers using AI to maintain a fundraising tracker or coordinate a limited partner (LP) transfer, the findings raise practical questions about delegated authority and supervision. The controls should let managers delegate routine work while retaining oversight of consequential decisions; if reviewers must reconstruct every task from scratch, much of the benefit is lost.

1. Institutional investors do not remove the governance question

The FCA’s publication is relevant to FCA-authorized fund managers serving professional and institutional investors. While the Consumer Duty applies only to retail market business, including distribution-chain activities that determine or materially influence retail customer outcomes under the FCA Handbook’s application provisions, broader governance obligations do not depend on having retail clients.

Managers serving institutional LPs must still meet FCA requirements on systems and controls (SYSC), risk management, and accountability under the Senior Managers and Certification Regime (SMCR). An internal AI tool can affect the fund by misstating a commitment or recording a closing condition as satisfied without evidence. As the FCA notes in section 3.3 of the publication, senior managers must take reasonable steps to ensure the business areas they oversee are effectively controlled.

Following the FCA’s good practice on named ownership and central agent inventories, managers should maintain an internal record for workflows like a fundraising tracker. That record should name the person supervising the agent, specify the data it may access and the updates it may make, and identify who can restrict or suspend the agent if errors threaten the tracker’s accuracy. The controls should be proportionate to the agent’s authority and the consequences of its errors.

2. Define where the agent supports, automates, or abstains

In its Consumer Duty examples, the FCA highlights choosing where AI should "support, automate, or abstain" as good practice. Applied to an LP transfer, an agent might be authorized to chase an outstanding document at agreed intervals. Communicating the general partner’s (GP) consent has different consequences: it requires evidence that the authorized decision-maker has granted consent, and the communication must accurately reflect any conditions.

In practice, those boundaries should be enforced through access permissions and required approvals. An email asking the agent to skip an approval gives it no additional authority; conflicting instructions or missing evidence should trigger escalation to an authorized person. Managers should establish these limits before delegating the work.

3. Counter automation bias and keep evidence with the record

Clear permissions still leave a question about the quality of human review. The FCA’s examples of "automation bias" describe reviewers increasingly accepting AI recommendations without scrutiny. At the same time, the FCA warns against excessive reliance on human-in-the-loop controls alone, highlighting the need for technical checks alongside them so preventable errors are caught before reaching a person. Where human judgment is required, reviewers need the expertise and time to examine proposed changes, the authority to reject incorrect updates, and a route for escalating recurring failures.

In a fundraising tracker, one way to support meaningful review is to surface uncertainty and link proposed updates directly to source evidence. The agent should flag ambiguity, such as when an email is unclear or an entity name only partly matches the investor register, and the supporting evidence should sit beside the update. For a commitment change, the reviewer needs to identify the investing vehicle and compare the recorded amount with the source correspondence. A log saying only "AI updated investor" leaves the reviewer searching the inbox to reconstruct the task before checking the result.

4. Judge errors by their consequences and use deterministic rules where needed

An overall accuracy percentage can conceal differences in risk. An inaccurate activity summary and an unauthorized commitment change might each count as one error, even though their consequences differ. Testing should examine failures that matter to the fund: an investor using multiple investing vehicles, an old email forwarded into a current thread, or fresh comments reopening a document previously marked as agreed. These cases test whether the agent preserves the correct entity, chronology, and document status.

The FCA identifies performance thresholds with predefined responses as good practice. An ambiguous email might require review of one entry, whereas repeated confusion between investing vehicles should trigger a pause in automated commitment updates until the defect is fixed. The consequences of an error should also guide the choice of software. In our view, fixed permissions, approval requirements, and rule-bound compliance calculations belong in deterministic code.

5. Supervise the provider and refine the workflow after deployment

Delegating work to a provider leaves the regulated firm responsible for its compliance obligations. Both section 4.2 of the publication and the FCA’s outsourcing guidance emphasize third-party oversight, and governance obligations can apply outside the technical definition of outsourcing. Managers should establish where investor data is processed, which model providers support the service, how material changes are communicated, and how access restrictions and data security are maintained. These questions also apply when counsel or an administrator uses AI to deliver services to the manager.

An underlying AI model can change while the interface remains identical. Repeating workflow tests helps managers assess how an updated model handles the fund’s exceptions, consistent with the FCA’s examples of managing model changes and its accompanying speech on AI. Managers should also verify fallback and exit arrangements so they can access current records and continue work if a service becomes unavailable shortly before a closing. When a rejected update reveals a recurring defect, such as assigning emails to the wrong vehicle, correcting the entry leaves a second job: revising the workflow rules that produced the error.

How Navys puts these controls into practice

We built Navys and Sam, our AI associate for fund operations, around this division of authority and supervision:

• Granular permissions and escalation: Navys lets managers define what each participant, including Sam, can access and do across vehicles and workflows. Sam handles routine coordination and follow-ups through email, while steps requiring specialist sign-off or updates with incomplete support are held for an authorized person to review.
• Source-linked audit trails: Navys records who made each update and when across the investor register, LP transfers, and the Intelligent Fundraising Tracker. Email-driven updates link to the source correspondence alongside Sam’s recorded steps, confidence levels, and explanations, so reviewers can check the fund record without rebuilding history from separate threads.
• Deterministic code for fixed rules: Navys’s publicly traded partnership (PTP) analysis engine uses deterministic code: the same transaction parameters, tax classifications, and calculation rules produce the same mathematical result every time. Professional review establishes that the inputs and tax treatment are appropriate, while the code executes the calculation reproducibly.
• Exception testing and ongoing refinement: Before a workflow goes live, our legal and engineering teams test Sam against the exceptions fund managers actually encounter, including multi-vehicle commitments, forwarded threads, and reopened documents. In production, we record what reviewers change and why, using those corrections to refine Sam’s instructions, checks, and test suites while the manager retains approval authority.