AI Agents in Finance: What 75% of Finance Leaders Are Piloting, and What It Means

What the AI Agents in Finance Survey Found

The research, published by OvationCXM on 29 September 2026, surveyed 520 US finance leaders, including CFOs, finance heads, treasurers, controllers and finance managers, at companies ranging from under $5m to over $1bn in annual revenue. The headline findings:

• 75% actively use AI agents, or are piloting or testing them, in at least one finance or banking workflow.
• 68% expect their AI agents to make transactions within the next 12 months.
• 50% are willing to delegate transaction-level tasks to an agent, provided a person approves them.
• 54% would find an internal or third-party workaround if their bank could not support their agents.
• 15% would move banking business to a provider that could.

The top workflows included transaction tracking, payment approvals and payment or transfer initiation, with 28% expecting to use agents for ACH payments and 23% for wire transfers. Those are the US equivalents of BACS and CHAPS, which gives a sense of how close to real money movement these plans already are.

Two caveats are worth keeping in mind. This is a US sample, and UK adoption patterns, payment systems and regulation differ. And "using or piloting in at least one workflow" is a broad definition that includes small tests. The direction of travel is clear, but the figure is not evidence that three quarters of finance teams have agents in production.

From Reading to Transacting: The Big Shift

The most important number in the survey is not the 75%. It is the 68% who expect agents to transact within a year.

Most AI use in finance so far has been read-only. AI summarises, drafts, reconciles and flags, and a person acts on the result. If the AI gets something wrong, a reviewer catches it before it matters. An agent that initiates a payment, approves an invoice or updates bank details is a different kind of system. It acts, and errors become real money.

That does not make transacting agents a bad idea. Much of the payments process is highly structured, repetitive and already rule-based, which is exactly what automation does well. But it changes the questions a finance leader needs to answer. Who is accountable for a payment an agent initiated? How is segregation of duties maintained when the same agent prepares and submits? How do you prevent an agent being manipulated by a fraudulent email asking it to change supplier details?

The survey's own finding gives a sensible starting point. Half of respondents would delegate transaction-level tasks only with human approval. That model, where the agent prepares and a person approves, captures most of the time saving while keeping the control where it belongs. For UK teams, it also fits naturally with existing approval workflows for BACS, Faster Payments and CHAPS.

The Workaround Risk Finance Leaders Should Take Seriously

The finding that deserves more attention than it got is that 54% would find a workaround if their bank could not support their agents.

Workarounds are where finance risk tends to hide. An agent connected to a bank portal through a browser, a third-party tool bolted on without proper review, or a script that a well-meaning analyst built over a weekend can all work perfectly until the day they do not. They often sit outside the controls, audit trails and access reviews that cover your core systems.

This is the same pattern we saw with spreadsheets and macros, and more recently with staff using personal AI accounts for work. The demand is real, so blocking it rarely works. The answer is to provide an approved route quickly enough that people do not build their own.

In practice, that means three conversations. With your bank, about what agent access and approval APIs they support or plan to. With IT, about which agent platforms are approved and how their permissions are managed. And with your own team, about what is and is not acceptable while those approved routes are being set up. A short written policy is far better than a silent gap.

How to Run a First AI Agent Pilot in Finance

If your team is not yet in the 75%, the right first pilot is narrow, low risk and measurable. We use the same structure with clients:

Pick a read-and-prepare task, not a transact task. Good candidates are payment run preparation, where the agent assembles the proposed run and supporting evidence for approval, supplier statement reconciliation, or month-end close tracking. Each saves real time and keeps a person in control of the outcome.

Define success before you start. Hours saved per month, error rate against a manual baseline, and how often a person had to correct the output. Without a baseline, every pilot looks like a success and none can be defended.

Run it in parallel first. For the first month, let the agent work alongside the existing process and compare results. Only switch over once the numbers support it.

Set the controls from day one: a named owner, minimum access, human approval for anything that moves money or changes records, and a log of what the agent did.

Decide at 90 days. Scale it, adjust it or stop it, based on the evidence. Stopping a pilot that did not deliver is a good outcome, not a failure, because it saves you from scaling the wrong thing. Our 90-day AI roadmap for finance sets out the full sequence.

What This Means for UK Finance Teams

The survey is a useful signal even for UK teams. Agents are becoming normal in finance faster than most planning cycles assumed, and the move from reading to transacting is coming within the next year for many organisations.

The teams that benefit will not be the ones that adopt fastest. They will be the ones that build the skills and controls first, so that when agents are ready to act, the organisation is ready to let them. That means people who can set a clear goal for an agent, judge whether its output is right, and design the approval points that keep it safe.

Those skills do not depend on any particular vendor, and they are useful today, long before an agent touches a payment. If you are starting from scratch, begin with your team's everyday AI use, then move to a single contained agent pilot, then widen from there. Our guide to always-on AI agents covers the main platforms, and our AI agents for finance guide covers the use cases in more depth.

One last point. Surveys like this tend to reflect the most enthusiastic adopters, and they can create pressure to move faster than your controls allow. A team that pilots one agent well this year will be in a far stronger position than one that rushes three into production and has to unwind them.

Frequently Asked Questions