Financial firms should limit what AI agents can do with client funds and require human sign-off on irreversible transfers, Global Finance & Technology Network (GFTN) Chairman Ravi Menon said.
Menon outlined five principles for AI agent governance at the HSBC Singapore Wealth Investment Summit.
They spell ALPHA: alignment, limits, prudence, humility and accountability.
He urged firms to explain the purpose of a task to an agent so it does not hit a target in a way that defeats the point of the task.
Menon compared an AI agent’s authority to that of an external fund manager.
A family office can let a manager trade without approving every transaction, while setting rules on what the manager can buy, how large a position they can take and how far the portfolio can fall before they must stop and call the client.
AI agents should earn more freedom through testing
Menon suggested letting agents make recommendations while people still make the decisions.
Firms can compare the results before letting agents act on their own, starting with low-risk tasks where staff can undo errors.
He also called for independent checks and ongoing monitoring because a vendor update can change how an agent behaves.
For major risk, credit and investment decisions, Menon urged firms to use more than one underlying AI model.
Agents that run on the same model can share the same blind spots.
He also urged firms to report errors, client challenges and instances where staff overrule an agent to leadership so the firm can learn from them.
“You can never delegate moral, legal, or fiduciary ownership of the outcome,”
Menon said.
Featured image: Edited by Fintech News Singapore, based on image by swampfire via Magnific


