AI agents can now move money. Most finance functions have not yet decided who answers for them when they get it wrong.
Picture an ordinary Tuesday in a finance team. An AI agent reads the accounts payable inbox, matches invoices to purchase orders, checks due dates and settles the bills that fit its rules. Nobody clicks approve.
By lunchtime, forty suppliers have been paid. One of them changed its bank details last week in an email that looked exactly like the real thing.
Who signs off on that payment? Not the agent. It cannot be dismissed, fined or called before an audit committee.
Not the software vendor, whose contract will almost certainly say so. In most organisations the honest answer is that nobody does, and that is the problem.
The first agent payments are already live
This is no longer a thought experiment. In March 2026, Banco Santander and Mastercard completed what they described as Europe’s first live end-to-end payment executed by an AI agent within a regulated banking framework. BBVA with Visa and Nordea with Mastercard have since announced similar transactions.
The International Monetary Fund, in an April 2026 note on agentic AI and payments, said the technology “could shift transactions from human-initiated instructions to agent-mediated decisions”.
Look at how those pilots were designed. Santander’s agent could only pay within limits and permissions the account holder had set in advance. BBVA’s transaction ran on tokenised card credentials, real-time fraud monitoring and the strong customer authentication that European rules already require.
The banks did not trust the agent. They trusted the boundaries around it. Most corporate finance teams adopting agents for payables, expenses and procurement have not yet built boundaries of the same quality.
Why the old controls do not fit
Finance controls were designed for people with job titles. A delegation of authority matrix says a finance manager can approve payments up to a set limit, a director up to a higher one and the board above that. Segregation of duties stops the person who sets up a supplier from also paying it. The four-eyes principle puts a second person between a decision and the bank.
An agent breaks each of these assumptions quietly. It may set up the supplier, match the invoice and release the payment in a single run, collapsing three duties into one. It does not appear in the delegation matrix because nobody thought to add software to a document written for humans. Its “second pair of eyes” is often another automated check written by the same vendor.
Agents also fail differently from people. A tired clerk makes random mistakes. An agent makes the same mistake at speed and at scale.
It can be manipulated by a convincing email, a doctored invoice or a planted instruction, and it will act on that manipulation with complete confidence. As the World Economic Forum noted in July 2026, “the difference between helpful automation and harmful activity may come down to whether institutions can understand intent, authority and context before money moves.”
Regulators are already clear on one point
The direction of travel is not ambiguous. In March 2026, the UK Competition and Markets Authority published guidance on using AI agents. It is written for consumer-facing businesses, but its central principle applies to any finance function: a business is responsible for what its AI agent does, as it would be for an employee, even when a third party designed or supplied the agent.
In Europe, the EU AI Act requires human oversight, record-keeping and governance for high-risk AI systems. Most payables agents will not fall into that category, but its expectations of oversight and traceable records are a sensible benchmark for any agent that moves money.
You can delegate a task to an agent. You cannot delegate the accountability. When an agent pays the wrong supplier, the loss, the explanation and the reputational damage all land on the organisation, and in practice on the finance leader who let it run.
Five questions to answer before an agent touches money
As a chartered accountant with an audit background, I would ask any finance team the same five questions before an agent is given authority to make payments.
First, who owns the agent’s decisions? Not who maintains the software, but which named person answers for each payment it makes. If the answer is a team or a vendor, there is no owner.
Second, what exactly is its mandate? Write the agent into your delegation of authority like any employee: which payment types, which suppliers, what value limits and what it must never do, such as changing bank details or setting up new payees.
Third, what evidence does it leave? Every payment should leave a trail showing what the agent saw, which rule it applied and why it decided to pay. If you cannot reconstruct the decision afterwards, you cannot defend it to an auditor.
Fourth, where must a human still decide? Define the triggers in advance. Typical candidates are new suppliers, changed bank details, payments above a threshold and anything the agent flags as unusual. Those decisions should reach a person with the time and the authority to say no.
Fifth, who tests it independently? The team that configured the agent should not be the only team checking it. Internal audit or an independent reviewer should test it the way they test any other key control, including deliberately feeding it a fraudulent invoice to see what happens.
Speed is not the same as control
AI agents will make finance faster, cheaper and in many ways more accurate. I am not arguing for keeping them away from money. I am arguing that authority given to software must be matched by accountability held by people, and that the finance function is best placed to design that match. It already does this for people every day through approval limits, reconciliations and sign-off.
The organisations that get this right will not be the ones with the most agents. They will be the ones that can answer, for every payment an agent makes, a simple question that boards, auditors and regulators will increasingly ask: who signed off?
About the author
Ademola Isimeme Odewade is the Founder and CEO of K-DEMS Consulting Limited in London, which helps CEOs and finance leaders adopt AI without losing control. A Chartered Accountant (ACA) with an audit background at PwC and KPMG, he is the author of Human-Led, AI-Powered Finance and The AI-Powered Future of Work, and hosts The King Dems Podcast and Diary of a CFO.
