Trust is the infrastructure autonomous finance runs on

AI agents steered US$14.2 billion of shoppers' money over a single November weekend. Corporate finance will hand over its payment runs once every agent decision survives the audit that follows it.

On the last Friday of November 2025, while much of America slept off Thanksgiving, agents went shopping. Traffic from AI agents to US retail sites ran 805% above the previous Black Friday, and by the close of the holiday weekend agents had steered US$14.2 billion in online sales globally.

Picture this: in the midst of that flood, agents read reviews for near-identical air fryers, chose the cheaper one, filled in a saved card number, and clicked buy. The retailer shipped the box. At no point did anyone speak to a human. And that freedom has a source. eCommerce has been building the machinery to absorb mistakes for thirty years. A wrong purchase meets a returns policy, a disputed charge meets a chargeback, and the card networks wrote the liability rules long before software learned to shop.

Corporate finance offers an agent none of that shelter. In our recent study Building An AI-Ready Finance Function, we commissioned Forrester Consulting to survey 1,279 finance leaders globally and found that just 11% of core finance workflows run without a human in the loop. Finance leaders want autonomy, and until a team can account for what an agent did after the fact, 11% is where the number stays.

The price of being wrong in finance

The World Economic Forum lists trust, infrastructure, and data as the three biggest blockers to agentic AI adoption. Finance feels the trust one first because finance tolerates the least. A hallucinated sentence in a press release costs one edit and a raised eyebrow. A hallucinated payment costs cash, and in a regulated business it can cost a licence.

KPMG has priced the reward for solving this at US$3 trillion in global productivity gains, which for the average Fortune 1000 company works out to a 5% lift in EBITDA. That figure explains why 74% of the leaders in our study ranked integrating AI among their top priorities for the coming year. And it explains why the same leaders, in the same breath, ranked strengthening financial controls, audit readiness, and regulatory compliance just as high. They want the EBITDA lift, and they want to pass the audit that comes with it.

Every agent decision needs a paper trail

Spreadsheets are wonderfully predictable. Feed them the same numbers twice and you'll get the same answer twice, an auditor checks the formula once and trusts every result it produces. AI agents play by different rules. Give one the same prompt on Tuesday and on Thursday and it might take two different actions, because it reasons its way to each decision fresh. So the audit has to grow. It now covers every choice the agent makes, and someone has to explain each choice when the auditor asks.

As the Forrester study shows, most finance teams aren't set up for that yet. 44% of organisations have consolidated their finance platforms, yet 84% still need manual steps to finish their workflows, and 53% say their people have little experience running AI-enabled processes. And consolidation solves a different problem anyway. It puts every balance on one screen, which is visibility. Explaining why an agent moved the money is a second job.

That second job, explainability, is the test every autonomous workflow will face. An agent can pay the right supplier on the right day, and if nobody can show the auditor how it got there, the workflow fails.

Know Your Customer, meet Know Your Agent

Banks in the 1960s moved money for people they'd never checked. Someone could open an account under a made-up name, wire the cash to a bank in another country, and nobody in that chain had to write down who they were or where the money came from. Needless to say, problems arose. So in 1970, the US Congress passed the Bank Secrecy Act, and it set a rule every bank still works to. Before you move someone's money, you need to know who they are, what they're allowed to do, and who answers for it when it goes wrong – Know Your Customer (KYC).

Agents drop finance back at that same starting line. A team can now hand a payment run to software that opens its own sessions, calls its own APIs, and acts without anyone typing a command, and the old three questions apply again to a piece of code instead of a person.

The answer taking shape is Know Your Agent, and it asks four things: what the agent is, what it's allowed to do, who is the human or institution accountable for what it does, and who’s watching after approval.

Safeguards have to sit across the whole stack

Governance that lives in the model alone guards one room of a large house. In our study, 65% of organisations name fragmented finance data as the top barrier to scaling AI, and the logic follows from the last section. Nobody can supervise what they cannot see. An agent that touches payments, treasury, and reconciliation needs oversight in each of those layers, and regulation is moving the same direction. The EU AI Act and ISO 42001 both point toward documented, auditable AI governance, and they will reach finance teams whether or not those teams have read a page of either.

That layered coverage already exists on our platform. Administrators set spend limits per card and per employee, freeze a card the moment something looks wrong, and watch each transaction land as it happens. Every safeguard writes its own record, so the audit trail builds while the work runs.

The same discipline runs on the payments side too. Our fraud engine scores every transaction while it's still in flight, returns an allow, challenge, or block decision before authorisation completes, and attaches a two-level reason code that turns the model's logic into an explanation a merchant's risk team can act on. Explainability is a habit that we’ve already built into our stack.

What changes once trust is built in

Building that kind of oversight in-house is a big ask, and most finance teams have no intention of trying. In the Forrester study, 43% expect their AI capability to arrive through vendor platforms, with their own people running orchestration, integration, and control. That split makes sense. The vendor carries the licences, the certifications, and the record of every decision, and the finance team decides where an agent gets to act.

Once explanations come as standard, the interesting work starts. Agents will price risk transaction by transaction, catch fraud inside the authorisation window, and move liquidity between entities at machine speed, each decision leaving a paper trail a board can read the next morning.

Somewhere in the Black Friday numbers of 2028 sits a supplier payment run that an agent approved on a Sunday night, that a reason code explains in one line, and that an auditor signs off without picking up the phone. The whole point of this work is to make that payment run boring. The 11% climbs the day a CFO can answer for a machine's decision with the same confidence they'd back a junior analyst's, and the teams that get there first have already started testing and will be automating while everyone else is still drafting the policy.

Source: Unless stated, all statistics referenced in this article come from "Building An AI-Ready Finance Function", a commissioned study conducted by Forrester Consulting on behalf of Airwallex, June 2026.