
Hey {{first_name | there}} ,
I have been reading through the programme for this week's Singapore FinTech Festival, taking place from 18 to 20 November. There is plenty in the impressive speaker line-up I want to hear. I'll be following from London, with one question in mind.
When an AI agent acts on our behalf, how do we know it has stayed within the authority we actually gave it?
Imagine asking an agent to book a flight for no more than £800.
It finds one for £780. By checkout, the fare has risen to £815. The agent books anyway, reasoning that an extra £15 is preferable to losing the seat.
It may have made the right economic decision. But it broke a clear instruction. Would “around £800” change that? What if the higher fare included baggage?
The uncomfortable point is that a good result and an authorised action are not the same thing.
The missing layer between intention and transaction

Gif by aetv on Giphy
On Thursday, a Singapore FinTech Festival panel, “Agentic Payments: Are Our Rails Ready?”, brings together voices from Mastercard, Google and Ant International, moderated by the Monetary Authority of Singapore. Its agenda asks who is accountable when agents pay on behalf of people and businesses.
That is exactly the question the industry needs to take seriously.
Agents could discover a product, negotiate terms and initiate payment without the customer touching a checkout screen.
But removing friction also removes moments when a human might notice that something has changed.
A payment credential may establish that a system can initiate a transaction. It does not, by itself, tell us whether the particular purchase remained within the customer's instructions. Authentication, consent and delegated authority are related, but they are not interchangeable.
The FCA's 2026 emerging-technology horizon scan imagines a “proxy economy”, in which AI representatives increasingly manage consumers' financial affairs. It cautions that people may approve permission requests as casually as cookie notices.
That is not a reassuring model of informed delegation.
Why are so many brands shifting ad spend to CTV?
People spend more time streaming than they do with cable, social media or video games, and advertisers know it. In fact, according to Advertiser Perceptions and Premion's 2026 CTV/OTT Advertiser Survey, 70% of advertisers are raising CTV budgets in 2026, with much of that budget diverted from social advertising.
For marketers who want a presence in streaming, Roku is the obvious place to go, with scale that speaks for itself. With more than 100 million households representing 47% of all streaming time in the US, Roku easily has more reach than every competitor.
With Roku Ads Manager, you can reach viewers across more than 100 premium apps in a single buy. Combine your customer data with Roku's insights to reach only the households you want, then measure and optimize to your heart's content.
Want to give it a try? Unleash your first campaign in just four steps: (1) Create your account, (2) define your objective and target audience, (3) upload creative, and (4) launch.
First-time spenders get a $5,000 match on the first $5,000 spent with code BEEHIIV5K. Terms apply.
Permission has to survive contact with reality
Working on BuildMoat, I keep returning to the idea of a behavioural envelope: the conditions within which an agent may operate.
A spending cap is one condition. Others could include approved counterparties, permitted accounts, transaction types, geographic restrictions, expiry times and triggers for seeking fresh approval.
Consider an agent managing a customer's overdue repayment. It may be allowed to suggest a plan within predefined terms, but not waive a balance, disclose sensitive information or commit the lender to an exception. An agent helping a treasury team might move funds between specified accounts, but should not infer permission to send money to a newly supplied beneficiary.
These are different levels of authority, even when the agent's stated objective is perfectly legitimate.
The difficult cases are plausible exceptions: a higher price, an unfamiliar beneficiary, a second agent acting for the first.
If an agent can reinterpret its own limits whenever circumstances become inconvenient, those limits are merely suggestions.
Human oversight must arrive before the point of no return
The answer cannot be to ask a person to approve every action. That would defeat much of the value of autonomous systems.
But neither is “human in the loop” sufficient as a label.
In its October guidance on AI good practice, the FCA emphasises that meaningful oversight requires people with the authority, information, capability and escalation routes to intervene. It also points to agent inventories, clear ownership, continuous monitoring and auditable records.
The timing matters as much as the existence of a reviewer.
If a human approved a £780 purchase, the approval should not silently carry over to £815. If a reviewer can stop an agent only after funds have settled, the organisation has a remediation process, not a preventive control.
For financial institutions, this makes the design question concrete: which decisions can be made autonomously, which require renewed authority, and which must be refused?
Boundaries should reflect consequences, not model confidence.
Evidence of success is not evidence of permission

Gif by nbc on Giphy
Another Singapore event caught my attention: the 19 November launch of the AI Spotlight Observatory, developed in the context of the FCA's AI Lab with industry collaborators. Its purpose is to help institutions compare real-world AI use cases and the evidence behind them.
That is valuable. But I hope the broader conversation about evidence goes one step further.
When a transaction is challenged, can the institution reconstruct the instruction, the version of the limits in force, the relevant approvals, any changes and the final action? Can it show not just that an agent produced an acceptable outcome, but that it was entitled to take the steps it took?
Logs cannot undo a harmful payment, but can establish what happened.
Across jurisdictions, legal responsibilities will differ. The operational challenge remains remarkably consistent: someone must define authority, someone must enforce it, and someone must be able to explain what happened.
What I hope Singapore pushes forward

Giphy
I expect plenty of demonstrations of agents doing impressive things this week. They deserve attention. What interests me more is whether the industry is beginning to define the infrastructure for permitted action.
Can a customer withdraw permission before an irreversible step? Can a bank distinguish a valid credential from a valid mandate? Can a senior manager see which agents are operating, under whose authority, and with what limits?
If we cannot answer those questions, faster agentic payments may simply mean faster unauthorised decisions.
The most important capability of an AI agent may not be knowing how to act. It may be knowing when the next decision is no longer its to make.
Who is accountable when AI agents act?
Every week, The Accountability Record examines the decisions, regulatory developments and governance challenges shaping AI in financial services.
Subscribe to The Accountability Record for a sharper view of what accountable leaders need to know.
Disclosure: I am the founder of BuildMoat, which is developing controls for AI-agent actions and delegated authority.

