The finance function is getting AI colleagues. Here is how to stay in control

For a decade, finance automation meant rules: a bot that copied figures from one screen to another until the screen changed. The new generation of AI does something different. Agents can read an invoice they have never seen, decide which ledger it belongs to, notice that the supplier's bank details changed last week and ask someone before paying. That shift, from automating keystrokes to delegating tasks, changes how a finance function should be designed.

From tools to colleagues

An AI agent is software that is given a goal, a set of tools and permissions, and the ability to plan the steps in between. In finance, that means an agent can be told to reconcile the bank accounts for the month, and it will pull the statements, match transactions, investigate the differences it understands and escalate the ones it does not.

The more interesting step is when several agents work together. One prepares accruals, another reconciles intercompany balances, a third drafts the management commentary from the results, and a coordinating agent tracks what is done and what is waiting for a human. The structure starts to resemble a team: roles, handovers, a reviewer and a manager. That is why the most useful way to think about agents is not as a tool purchase but as an organisational design question.

What agents are good at today, and what they are not

Agents perform best on work that is high-volume, rules-heavy and well documented, where the data is digital and the cost of a mistake is caught by a review step. In most finance teams that describes a large share of the week:

• Matching and reconciliation: bank, intercompany, supplier statements and subledgers
• Document handling: invoices, receipts, contracts and purchase orders read, coded and filed
• Preparation: first drafts of journals, schedules, variance explanations and board-pack commentary
• Follow-up: chasing approvals, missing documents and overdue balances through email or chat

They are weaker where judgement is the work itself: estimates and provisions, revenue recognition on unusual contracts, tax positions, anything that will be defended to an auditor, investor or tax authority. Agents can prepare the analysis for those decisions. They should not make them.

The control problem nobody budgets for

Gartner has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. In finance, the third reason is the one that bites. An agent that can post a journal or release a payment is, in control terms, a new employee with system access, and it needs the same things a new employee needs: a defined role, limited permissions, supervision and a record of what it did.

The controls that matter are not exotic. They are the ones auditors already test, applied to a new kind of actor:

• Segregation of duties: the agent that prepares a payment cannot approve it
• Least-privilege access: each agent sees only the systems and data its role requires
• A complete audit trail: every action logged with its inputs, its reasoning and the person who approved it
• Materiality thresholds: above a set amount or outside a set pattern, a human decides
• Periodic testing: sample the agents' work the way an internal audit would sample a team's

Designed this way, an agent-run process can be more controlled than the manual process it replaces, because every step is recorded and nothing depends on someone remembering to tick a box.

What changes for the people

The fear is that agents replace the finance team. The more likely outcome is that they replace the part of the job nobody wanted. When preparation is done by agents, the team's work moves to review, exceptions and analysis: why margins moved, which customers are slow to pay and why, what the forecast implies for cash. Smaller companies gain the most, because they can run a finance function with the discipline of a much larger one without hiring for every task.

That shift needs planning. Reviewers need to understand what the agent did and why, which means the work must be explainable. Job descriptions change from doing to supervising. And someone senior has to own the design, because a badly designed agent team fails in the same ways a badly designed human team does: unclear roles, duplicated work and things that fall between the gaps.

Where to start

The companies making progress share a pattern. They pick one process with clear data and a measurable cost, typically reconciliations, accounts payable or the monthly reporting pack. They map it as it really runs, including the workarounds. They decide which steps an agent prepares, which a person approves and what evidence the reviewer needs. Then they run the agent alongside the existing process for a cycle or two and compare results before switching over.

Three questions are worth asking before any investment:

• Which decisions in this process require judgement, and who will own them?
• What would an auditor need to see to rely on the output?
• How will we know, every month, that the agents are still getting it right?

The bottom line

AI agents will become a normal part of how finance work gets done, in companies of every size. The advantage will go to businesses that treat them as members of the team, with roles, permissions and supervision, rather than as software switched on and left alone. Getting that design right is a finance, controls and technology problem at the same time, which is exactly why it should not be left to any one of those disciplines alone.

Sources

• Gartner, press release, June 2025: prediction that over 40% of agentic AI projects will be cancelled by end of 2027
• Gartner, finance research on AI adoption in the finance function, 2025–2026