Which AI agent could make an investment decision

You must establish and maintain a decision-rights register for each AI-assisted workflow, recording which agent may act, who owns the decision, what review occurs and how the decision is retained. Without this register, your firm cannot demonstrate who authorised an investment decision or reconstruct how it was made.

An analysis of 7 October 2026 of decision-rights registers for AI agents in investment workflows argues that private-capital firms should record which AI systems can make decisions, who is responsible for reviewing or approving their outputs and how those decisions are documented for accountability. The analysis uses investment decision-making as its concrete case, where limited partners, regulators and transaction reviewers may later need to understand how a decision was made. In our assessment, this exposes a critical gap in current practice: an AI policy can mandate human oversight without creating the operational evidence needed to demonstrate it actually occurred for a specific decision by a specific agent.

Why policy and evidence are not the same thing

A written AI policy states that oversight must happen. It does not show that a particular agent stayed within its authority for a particular decision, or that an identifiable person tested the output. In investment workflows, this distinction is especially material. The original decision context may no longer be available when a limited partner questionnaire arrives, a transaction undergoes due diligence, or a regulator asks how a decision was reached. The register bridges this gap by creating a living operational record tied to each workflow, not to each supplier or model.

What must the register contain and how is it structured?

The register is a per-workflow document, not a per-supplier one. A single language model or tool can appear in multiple workflows with different authorities; only at workflow level is it clear what that agent may do. Each entry must record which agent may take or recommend which decisions, the named owner accountable for that decision, what check the reviewer performs and how the decision is retained. The register should also document the production inventory, identity and access map, permissions, logs, revocation procedures, incidents, exceptions and remediation owners. This operational evidence is what makes the register actionable rather than aspirational.

Which governance obligations does the register address?

• Delegated authority without evidence — an AI policy mandates oversight but creates no record of whether it occurred for a specific decision.
• Distributed review without clear ownership — multiple people may review an output, but no single person is named as accountable.
• Model and tool changes without authority re-testing — agents are modified in production without re-evaluating whether their authority still applies.
• Recommendation and execution rights conflated — the register does not distinguish between the right to suggest a decision and the right to execute it.
• Reconstruction failure under scrutiny — when a transaction, LP questionnaire or regulatory review demands an audit trail, the firm cannot show the decision path.

What concrete controls must you demonstrate?

1. Appoint a single named owner per workflow — one accountable party for each AI-assisted decision, even where review is distributed across several people.
2. Separate recommendation rights from execution rights — explicitly record in the register which actions each agent may only suggest and which it may execute.
3. Test authority before every model or tool change — tie a review moment to every modification, so that the agent's authority is re-validated before it remains in production.
4. Suspend deployment until the register is complete — do not put an agent into production if its workflow entry remains incomplete, because you cannot yet demonstrate its authority or accountability.
5. Define system access authority with override and shutdown procedures — agents with system access must have clearly defined authority, named ownership, override arrangements and tested intervention or shutdown procedures.

How does the register serve due diligence and board oversight?

A complete and current register gives investment committees, limited partner questionnaires and transaction due diligence a concrete evidence object. The register translates governance intent into diligence evidence. At board level, it operationalises the governance elements that oversight frameworks require: delegated authority, accountability, escalation, intervention and shutdown. The register also supports the audit trail and explainability that regulatory or internal review may demand. Responsibility for the decision itself remains with management and the board; the register simply makes that responsibility demonstrable.

Tooling can help maintain the register and flag when entries become stale or incomplete. What tooling cannot do is determine which decisions should be delegated to an agent in the first place, or judge whether a particular review was adequate for a particular decision. That remains your professional judgement, informed by your firm's risk appetite, the decision's materiality and the applicable legal and contractual retention periods. Legal and privacy advice must determine which rules govern the evidence you retain.