What Finance AI Agents Actually Do
The term "AI agent" is used loosely, but in a finance context it describes software that can receive a goal, break it into sub-tasks, execute those sub-tasks using tools and data sources, and return a result, all without human intervention at each step. This distinguishes agents from simpler AI assistants that require a human to approve each action.
In practice, the finance tasks where agents are being deployed in 2026 fall into four categories. Understanding which category you are dealing with matters, because the governance requirements are very different across them.
- Routing and triage: Spend approval routing (sending invoices above a threshold to the right approver), variance flagging (identifying P&L lines that deviate from budget and notifying the relevant owner), and exception identification in reconciliations. These are low-risk, high-volume tasks where the agent is essentially a sophisticated rules engine with natural language capability.
- Processing and extraction: Invoice processing (reading unstructured invoices and posting them to the correct accounts), bank statement reconciliation, and expense categorisation. The agent is doing cognitive work that previously required a human to read and interpret documents. Accuracy rates in production environments are now typically 85 to 95 percent, which is good but not infallible.
- Analysis and forecasting: Cash flow forecasting using historical patterns and forward order data, variance analysis with automated commentary, and rolling budget reforecast. These tasks involve the agent generating outputs that inform decisions; they do not execute decisions.
- Execution and action: Initiating payment runs, updating ERP records, submitting regulatory returns, or adjusting credit limits. These are the high-risk applications where the agent is not just producing an output but taking an action with real financial consequences.
Most finance AI agents in production in April 2026 are operating in the first two categories. The third category is emerging. The fourth is where the serious governance questions live, and it is where vendors tend to be most optimistic and CFOs should be most sceptical.
Separating Genuine Gains from Vendor Hype
The productivity case for AI agents in finance is real, but it is narrower than the marketing materials suggest. Here is an honest assessment based on what firms are actually reporting, rather than what vendors are projecting.
The honest counterpoint is this: most of the headline productivity numbers come from firms with genuinely chaotic manual processes that were badly automated before. If your invoice processing was already working well in a structured AP tool, the marginal gain from adding an AI layer is smaller. The agent is not magic; it is pattern recognition over structured data. If your data is poor, your master data is inconsistent, or your chart of accounts is a mess, the agent will produce poor outputs, and it will do so faster than a human would.
The Gartner Finance Automation Report for 2026 estimated that only 23 percent of finance AI implementations in 2025 delivered the productivity benefits initially projected, with the most common failure modes being data quality problems, insufficient change management, and governance gaps that forced firms to add manual review steps that eliminated the efficiency gain.
Governance and Control Risks
The governance risks of autonomous finance agents are not theoretical. They are specific, well-defined, and need to be addressed before deployment rather than after an incident. There are three primary risk categories.
Authorisation Limits and Delegation
Any agent that can take action must have a clearly defined authorisation limit. If the agent can initiate payments, what is the maximum value it can initiate without human approval? If it can update credit limits, what is the maximum movement it can make? These are not technical questions; they are policy questions that the CFO must answer and that must be encoded into the agent's operating parameters. Most finance teams deploying agents in 2025-2026 have found that defining these limits forces a useful clarification of the existing human authorisation policy, which is often undocumented or inconsistently applied.
Audit Trails and Explainability
Every action an AI agent takes must generate a complete audit trail that is readable by a human auditor. This means: what input did the agent receive, what decision logic did it apply, what action did it take, and what was the outcome? For regulatory purposes, and for your own internal control framework, you must be able to reconstruct the decision chain for any agent-initiated transaction. Many early agent deployments failed this test because the agent's reasoning was opaque and the audit log was incomplete. Your external auditor will ask about this in your next audit cycle if you have agents in production.
Segregation of Duties in Automated Workflows
Traditional segregation of duties requires that no single person can both initiate and approve a transaction. In an automated workflow, the equivalent principle requires that the agent cannot both propose and execute a transaction without at least one independent control checkpoint. This is straightforward in theory but requires careful design in practice. An agent that identifies a duplicate invoice and also cancels it has effectively collapsed the initiation and approval steps into a single automated process. Whether that is acceptable depends on the value and risk of the transactions involved.
"The question to ask about any AI agent in finance is not whether it is accurate enough to trust. It is what happens when it is wrong, and whether you will know quickly enough to stop the damage. Agents fail at scale; the errors are fast and systematic, not slow and random."
Building a Controls Framework for AI-Assisted Spend
A practical controls framework for AI agents in finance has four components. These are not theoretical governance principles; they are specific design requirements that must be implemented before go-live.
- Pre-action controls: Define what the agent is permitted to do, in what circumstances, and up to what value or impact threshold. Document these permissions formally, version-control them, and require CFO sign-off on any changes. This is the policy layer.
- In-action monitoring: Real-time logging of every action the agent takes, with anomaly detection that flags patterns inconsistent with the defined permissions. If the agent initiates 47 payments in a 10-minute window when the average is three, that needs to trigger an alert to a human. This is the detection layer.
- Post-action reconciliation: Every agent-initiated action must be reconciled against expected outcomes at a defined frequency (daily at minimum for payment-related actions). Unexplained variances must require human investigation. This is the review layer.
- Kill switch and rollback: You must be able to suspend the agent instantly, and you must have a defined procedure for reversing or escalating agent-initiated actions that are subsequently found to be incorrect. For payment-initiated actions, the window for reversal is short; this process must be pre-planned, not improvised.
Implementation Sequence for a 20-50 Person Finance Team
If you are running a finance team of 20 to 50 people and considering deploying AI agents, the following sequence reduces risk and maximises the probability of genuine efficiency gains. It is not the fastest path to deployment; it is the path most likely to result in a deployment that still works six months later.
The Regulatory Context: FCA AI Discussion Paper
The FCA published its AI Discussion Paper in 2025, setting out its expectations for firms using AI in regulated activities. While the paper is directed primarily at regulated financial services firms rather than their finance functions, the principles it articulates are relevant to how CFOs should think about AI governance more broadly.
The FCA's core concern is explainability and accountability: when an AI system makes a decision that affects customers or the firm's financial position, there must be a human who is accountable for that decision and who can explain the reasoning. This principle applies directly to the use of AI agents in finance: if an agent initiates a payment incorrectly or produces a materially wrong cash forecast, the CFO remains accountable. The technology does not dilute personal accountability.
For fintech and financial services firms specifically, using AI agents in processes that touch on regulatory reporting, customer money, or prudential compliance carries additional risk. The FCA has indicated that it will treat AI-related control failures with the same seriousness as traditional control failures. An audit trail that says "the agent did it" is not an acceptable response to an FCA supervisory request.
Key Takeaways
- Finance AI agents are most productive in routing, triage, and processing tasks. Execution-layer tasks (initiating payments, updating records) require much more rigorous governance and should be approached with caution.
- Data quality is the binding constraint. If your ERP, AP system, or master data is inconsistent, agent accuracy will be poor regardless of the quality of the model. Fix the data before you deploy the agent.
- Design the governance framework before deployment, not after. Authorisation limits, audit trails, reconciliation cadence, and kill switch procedures must all be defined and documented as preconditions for go-live.
- Segregation of duties must be preserved in automated workflows. The agent must not be able to both propose and execute a transaction without an independent control checkpoint.
- Run every new agent in shadow mode (read-only, producing outputs but taking no actions) for at least eight weeks before granting it execution permissions. Compare accuracy rigorously against human baseline.
- The CFO remains personally accountable for AI-assisted processes. The FCA treats AI control failures the same as traditional control failures. "The agent did it" is not an acceptable explanation.
- Budget realistically: a well-governed AI agent deployment for a 20-50 person finance team takes six to nine months end-to-end and requires significant change management investment.
