Finance is not short of AI. The Cambridge Centre for Alternative Finance found that 81% of financial services firms are using AI at some level, but only 40% have reached the advanced stages it calls ‘scaling’ or ‘transforming’. (CCAF, 2026 Global AI in Financial Services Report)
Agents sit at the sharp end of that wedge. Gartner says 57% of finance teams are implementing agentic AI, or plan to do so. Even more interesting, it predicts that more than 40% of those projects will be canceled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls. (Gartner)
Caution matters more in finance than almost anywhere else, and so the practical question is what should an AI agent actually be allowed to do.
Reading an invoice? That carries one level of risk. Suggesting an account code? That's a level up. Changing financial data or initiating an action that affects money? Now we're into the territory of authority normally reserved only for trusted, long-serving colleagues.
So picture an agent that reads the invoice, weighs what it found, picks a route, decides on the fly whether this one needs a human or not, and keeps going. All of that sits inside approval workflows somebody set up in advance, and a person still owns the decisions that carry real consequence. We call that approval intelligence.
AI agents in finance are software systems that interpret information, choose a sequence of steps and perform actions toward a goal — at a moment when only 40% of financial services firms have reached an advanced "scaling" or "transforming" level of AI use. In accounts payable, an agent typically reads invoice data, checks the supplier, compares the invoice with a purchase order, identifies a discrepancy, sends the exception to the appropriate person, and continues the process once the issue is resolved. The recommended approach is to expand an agent's authority only after its recommendations have been measured against human decisions, not before.
The important difference is the ability to choose what happens next.
That capability is arriving inside the software finance teams already run. Gartner expects 40% of enterprise applications to include task-specific AI agents by the end of 2026, up from less than 5% in 2025. (Gartner)
The distinction matters because acting carries a different weight from answering. A wrong summary wastes a minute. A wrong action in accounts payable moves money.
These technologies often work together.
| Technology | What it does | Finance example |
|---|---|---|
| Rules-based automation | Follows predefined instructions | Route invoices over $10,000 to the CFO |
| Generative AI | Produces or interprets information | Summarize the reason for an invoice exception |
| AI agent | Chooses and performs several steps toward a goal | Review an invoice, identify a problem and send it to the appropriate reviewer |
OCR, coding suggestions and conversational assistants can form part of an agentic process. An agent goes further by using information to select and perform subsequent actions without requiring a new instruction at each stage.
That difference matters when assigning permissions. Software that only recommends an account code presents a different risk from software that can change transaction data.
Finance contains large volumes of repetitive work alongside decisions that still need judgement.
Accounts payable shows the opportunity clearly. Ardent Partners reports an average invoice-processing time of 8.2 days and an average cost of $9.84 per invoice, while the best-performing 20% of AP teams process invoices 79% faster than everyone else. The average invoice exception rate sits at 18.4%, and Ardent describes exceptions as the single biggest reason those benchmarks are not lower. (Ardent Partners)
AP leaders already know what they want from the technology. In the same research, 72% named smarter systems that can route invoices automatically and flag anomalies as essential to their next stage of performance.
Between receiving an invoice and making a payment, a finance team may need to:
Much of this work already suits AP automation. AI agents become useful where the subsequent action depends on context rather than a fixed rule.
The same idea applies elsewhere in finance.
| Finance area | Possible role for an AI agent |
|---|---|
| Accounts payable | Read invoices, compare records, suggest coding and route exceptions |
| Procurement | Check requests against policy and direct them into the appropriate approval process |
| Close and reconciliation | Identify unmatched items and prepare exceptions for review |
| FP&A | Gather approved data, prepare analysis and run defined scenarios |
| Collections | Prioritize accounts and prepare follow-up actions under an agreed policy |
The useful question is whether an agent can remove work while remaining inside the finance team's rules.
The clearest examples of this authority ladder show up in accounts payable — a use case explored in more depth in AI in accounts payable: what it does and where it struggles — where reading, coding, matching and routing work is repetitive enough to automate but variable enough to still need judgement.
AI can read supplier documents, identify fields and suggest accounting codes.
The next step may still need a person. An uncertain supplier, an unusual tax treatment or unexpected coding can be routed for review rather than accepted automatically.
An agent can also use extracted information to decide whether an invoice is ready for the next stage or needs attention first.
Matching combines structured rules with contextual decisions.
A system can compare: supplier, purchase order, amount, quantity, receipt information, invoice number.
An AI agent can help classify discrepancies and send them to the person responsible for resolving them. Matching at volume is already routine: ApprovalMax customers matched more than 2.4 million bills to purchase orders in 2025. (ApprovalMax 2025 Year in Review)
The person making the commercial decision still works inside the company's invoice approval process.
AI can identify unusual patterns across large transaction sets, including duplicates, unexpected amounts or changes to supplier information.
This is useful because weak controls remain a major source of fraud exposure. The Association of Certified Fraud Examiners found that more than half of occupational fraud cases in its study involved either a lack of internal controls or an override of existing controls. (ACFE, Report to the Nations)
AI can flag activity that deserves attention. The subsequent action should follow the company's existing audit and fraud controls.
Approval routing can depend on amount, supplier, department, entity and delegated authority.
Many of those conditions belong in fixed workflow rules. AI becomes useful when information needs interpretation before the correct route is obvious or when an exception needs classification.
The volumes involved make the point. In 2025, 17.1 million bills and 3.3 million purchase orders passed through ApprovalMax workflows, of which 614,000 were approved automatically under rules the customer had set in advance. Automatic approval at that scale works because the authority was defined first, not because the software decided it was safe.
The organization's approval policy remains the single source of authority.
Historical behavior alone should not become policy by default simply because an AI system has learned that one person usually approves a particular transaction.
A useful way to assess finance agents is to group them by the authority they receive.
The agent can inspect approved data sources and produce information.
Example: identify invoices with unusual values.
The finance team defines which information the agent can access and records its activity.
The agent proposes an action. A person reviews the recommendation.
Example: suggest an account code or recommend which manager should review an exception.
Human corrections should be measured. A recommendation system that people frequently change should not receive more authority without an explicit decision.
The agent can prepare or perform an action once a person explicitly authorizes it.
Example: prepare a transaction and continue the workflow once approval has been given.
At this level, segregation of duties, permissions and approval history become particularly important.
The agent performs specified actions independently within a narrow scope. Anything outside that scope returns to a person.
Example: process a defined group of low-risk transactions within established limits while sending exceptions for review.
This requires strict permissions, thresholds, monitoring, a named owner and a reliable way to stop or restrict the agent.
Good performance at one level does not automatically justify moving to the next.
The National Institute of Standards and Technology's AI Risk Management Framework provides a useful basis for this.
NIST recommends defining roles and responsibilities for human and AI participants, documenting the intended scope of an AI system, setting out human oversight, measuring performance and monitoring systems after deployment. It also includes mechanisms for human override, incident response and decommissioning. (NIST AI Risk Management Framework)
Those principles translate well into finance.
Before an agent starts work, finance should know:
This creates a much clearer boundary than a general instruction to "keep a human in the loop."
Measurement is where most organizations are weakest. The CCAF found that 55% of financial services respondents struggle to measure the value of an AI deployment, rising to 76% at large institutions. An agent whose performance nobody can measure is an agent whose authority nobody can justify extending.
An explanation can help someone understand why an AI system recommended an action.
Finance also needs evidence that can be checked independently.
For an AI-assisted transaction, that record may include:
A generated explanation can add context. The underlying transaction record remains the evidence.
This is why the approval history matters. A finance team should be able to trace the decision even when the AI-generated explanation is set aside.
That fits naturally with ApprovalMax's audit and fraud controls, where approval decisions, comments, delegations and other actions form part of the audit record.
An agent can follow its instructions correctly and produce a poor outcome when its input is incomplete, stale or incorrect.
Finance also contains values and calculations where reproducibility matters. A probabilistic model should not become the source of record for an accounting figure.
NIST's AI framework specifically calls for organizations to document data suitability, system limitations, expected uses and the conditions under which people oversee AI output.
A sensible division of work is:
When something goes wrong, that structure also makes it easier to identify whether the problem came from the source data, a workflow rule, the AI output or the subsequent decision.
Approval intelligence is our term for using AI within a governed decision process where authority, responsibility and evidence are defined before the system acts.
It is a practical model for finance teams deciding where agents belong.
Specify what the agent can read, recommend, change and execute. Include system permissions and financial limits.
An agent reading invoice data needs fewer safeguards than one permitted to alter supplier information or initiate a financial action.
Someone remains responsible for its scope, permissions, performance and review.
Low-confidence outputs, policy exceptions, unusual supplier changes and threshold breaches need an assigned destination.
AI and human actions should form part of the same record.
These principles fit the way multi-level approval workflows already operate: rules define the route, responsibilities are separated and each decision is recorded.
As agents perform more preparatory work, that framework gives finance a way to decide where their authority ends.
Most finance teams already have the infrastructure this framework depends on: an approval workflow that defines who signs off on what, and an audit trail that records every step. ApprovalMax applies that same structure to AI-assisted work — routing invoices, exceptions and purchase orders through the approval rules a finance team has already set, and logging every AI recommendation, human decision and change to the transaction in one place. That means an AI agent can be introduced at Level 1 or Level 2 authority immediately, with the audit trail and escalation routes already in place, rather than needing custom permission logic built around it.
Choose repetitive work with reliable source data and an obvious point for human review. Invoice classification or exception routing gives a team more control during early testing than autonomous payment execution.
Document: accessible systems, readable data, permitted changes, financial limits, approval requirements, escalation conditions.
NIST recommends specifying intended scope and human oversight as part of AI risk management rather than waiting for problems after deployment.
Include: missing documents, duplicates, new suppliers, changed bank information, unusual amounts, low-confidence outputs, policy exceptions, unavailable approvers.
Normal transactions tell you very little about how a system behaves when judgement is required.
Track: recommendations accepted, recommendations changed, recommendations rejected, exceptions sent to people, errors found after an AI action, review time.
NIST similarly calls for AI systems to be tested before deployment and monitored while in use.
Start with access to information or recommendations where that suits the task.
Additional permissions should follow evidence from the agent's measured performance.
Ask:
The answers tell you far more than the presence of an "AI agent" label.
AI is well suited to preparation, checking, classification, routing and follow-up — the exoskeleton framing explored in "The exoskeleton, not the robot": what AI's impact is on jobs in finance, where AI extends what a finance team can do rather than replacing the judgement calls that still need a name attached.
The direction of travel is not in doubt. Gartner expects at least 15% of day-to-day work decisions to be made autonomously through agentic AI by 2028, up from 0% in 2024. (Gartner) The question for finance is which 15%.
Approval assigns authority and accountability. For many material financial decisions, that remains a reason to keep a named person involved.
Examples include:
The boundary can move as systems improve and individual use cases prove reliable. Finance teams still need to decide deliberately which authority they are handing over.
Approval intelligence gives that decision a structure: define the authority, set the controls and measure what happens before expanding the agent's role.
Sources
AI agents in finance are software systems that can interpret information, plan steps and perform actions toward a financial or operational goal with limited human direction.
Conventional automation follows predefined instructions. An AI agent can use context to select between possible subsequent actions and continue working toward a goal.
OCR reads information from documents. It can provide data to an AI agent, but document extraction by itself does not perform the broader decision-making associated with an agent.
Software can be given that authority. Whether it should receive it depends on the transaction, company policy, the system's measured performance and the controls surrounding the action.
In accounts payable, an AI agent reads invoice data, checks it against the supplier record and purchase order, classifies any discrepancy and routes the exception to the right person. Ardent Partners puts the average invoice exception rate at 18.4%, which is where most of the manual effort in AP still sits.
The controls depend on their authority. Common requirements include restricted system access, financial limits, approval rules, named ownership, activity records, exception handling, monitoring and a way to restrict or stop the system.
Gartner predicts that more than 40% will be canceled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls. A bounded use case, defined responsibilities and measured performance give finance teams a clearer basis for deciding whether an agent should receive greater authority.