Digift sets controls for AI agents in tokenised finance

Digift has set out a control framework for institutions weighing whether AI agents should be allowed to move money in tokenised and decentralised finance portfolios. The guidance focuses on limiting an agent's authority to monitoring and tightly defined tasks before granting any broader discretion.

The framework emerged from a closed-door discussion in Singapore on investment management, risk, compliance and on-chain execution. It centres on the distinction between identifying a portfolio issue, recommending an action and carrying out a transaction, with separate checks at each stage.

At the heart of the approach is the view that not all AI involvement carries the same level of authority. An institution might first use a tool to detect that exposure to a single asset has risen above an agreed threshold, then let it suggest a rebalance, and only later consider whether it should be allowed to execute that rebalance within pre-set limits.

That sequence matters because each step shifts more decision-making from human teams to software. Institutions need to define clearly which decisions stay with people and which actions, if any, can be delegated to an agent.

Defined roles

The paper argues that a practical starting point is tasks with a narrow purpose and a reviewable output. These include monitoring collateral, reviewing a protocol and spotting portfolio concentration risks, all of which can support investment teams without handing over full execution authority.

By keeping the remit narrow at first, firms can assess whether an agent is producing useful analysis before extending its permissions. The model is presented as a staged path from assistance to delegation, rather than a single leap to automated portfolio management.

The paper also argues that data alone is not enough for investment decisions. If an agent recommends shifting assets from one vault to another, the destination cannot be judged on yield figures or wallet history alone.

Instead, the assessment needs context on collateral, liquidity and any dependencies on other protocols. Off-chain market developments also need to be considered because they shape the portfolio impact of a proposed rebalance.

Compliance checks

The same need for context applies to compliance screening. A collateral deposit, a vault subscription and a direct transfer may involve the same blockchain address, but they can represent different economic activities and require different controls.

Firms therefore need to understand the purpose of a proposed transaction before deciding whether it should proceed. Ongoing monitoring after execution also remains necessary, rather than treating screening as a one-off approval step.

Digift also draws a firm line between a model's recommendation and permission to act on it. In its view, any transaction proposed by an AI system should face independent authorisation checks before assets move.

Those checks would test whether the destination is approved, whether the resulting exposure remains within investment limits and whether the action itself is authorised under policy. A separate policy engine, rather than the model making the recommendation, would apply those predefined rules.

Restricted wallet permissions are another part of the suggested control structure. By limiting what an agent can do on-chain, institutions can reduce the potential damage if a model makes an error or the agent is compromised.

Audit trails and the option for human intervention are also described as core safeguards. Together, those measures are intended to create a process that lets institutions oversee an agent's actions and halt unexpected behaviour.

Ownership questions

Beyond operational controls, the paper raises legal and governance questions about ownership and accountability. Before an agent is allowed to move capital into a vault, an institution needs to understand what rights it holds once the assets enter a smart contract and how those rights are recorded.

That issue becomes more complex when a receipt token represents an interest in the underlying assets. In those cases, the legal basis of that token and the enforceability of related claims need to be established before delegation is considered.

Possible structures discussed included legal wrappers and separately managed accounts, though both would require further assessment. Digift added that investment, legal, compliance and custody teams need to work together on those questions even when an agent is technically able to execute the transaction.

The broader message is that institutions exploring AI agents in tokenised finance should separate analysis, recommendation and execution into distinct responsibilities, and show that controls work before wider authority is granted. Ownership and accountability also need to be settled before capital moves.