Finance executives are under pressure to turn artificial intelligence from a collection of experiments into a measurable operating result. Accounts payable is emerging as a practical answer because it combines high transaction volume, structured records, explicit rules, and costly exceptions.
Research from Basware and FT Longitude reports that general AI projects achieved 67% return on investment, while autonomous agents averaged 80%. The distinction is important: a generative tool produces an output for a person to interpret, whereas an agent can complete a bounded workflow within approval thresholds.
The headline does not mean autonomy creates returns automatically. Sixty-one percent of finance leaders said their organisations introduced custom agents primarily to test capability rather than solve a business problem. Among teams with weak returns, 71% had acted under pressure without clear direction, compared with 13% of strong performers.
Invoices Offer a Controlled Path to Action
Seventy-two percent of finance leaders see accounts payable as the obvious starting point for agents. Invoices arrive, data is extracted and cleaned, compliance rules are checked, duplicates or fraud are identified, and a payment is prepared. Each stage has evidence, ownership, and an observable outcome.
That structure makes value easier to measure. Teams can compare processing time, exception rates, duplicate payments, overpayments, fraud detection, and cost per invoice before and after deployment. An agent that merely drafts a summary is less consequential than one that removes a recurring manual handoff.
Data quality remains decisive. Basware says its systems use more than two billion processed invoices, providing context for distinguishing an unusual but legitimate transaction from an error. A company’s own supplier records, purchase orders, approval rules, and payment history must be equally dependable. Autonomy amplifies whatever foundation it receives.
Accounts payable is therefore a proving ground, not because it is trivial, but because it offers scale with clear controls and accountability.
Buy the Common Process, Build the Difference
Procurement choices vary by function. Thirty-two percent of finance leaders prefer agents embedded in existing accounts-payable software, while 20% favour internal development. In financial planning and analysis, the balance changes: 35% build in-house and 29% choose embedded systems.
The pattern suggests a useful rule. A standard process shared by thousands of organisations is a strong candidate for a vendor product with established integrations and training data. A workflow that represents a company’s distinctive strategy may justify internal engineering.
Buying does not outsource responsibility. Finance leaders still need to understand permissions, data retention, error handling, audit trails, and how the system behaves when documents conflict. They should also know whether workflows and records remain portable if the provider changes.
The word “agent” covers everything from a scripted automation to a model making contextual decisions. Buyers need to evaluate the actual autonomy and risk rather than the label.
Governance Lets Autonomy Expand Safely
Forty-six percent of finance leaders will not deploy an agent without clear governance. The strongest adopters treat that caution as an enabler. They are more likely to use agents for complex compliance checks because they have established limits and review mechanisms.
One practical model is to manage an agent like a junior colleague. It begins with low-risk work, operates under thresholds, shows its evidence, and earns broader authority through stable performance. Humans remain responsible for high-value payments, novel suppliers, or ambiguous exceptions.
Employee impact should also be assessed beyond headcount. Removing PDF extraction and repetitive comparison may let finance staff focus on liquidity, supplier issues, controls, and faster closes. Hidden verification work, however, can erase the claimed saving if people must recheck every result.
Accounts payable gives agentic AI something many enterprise programmes lack: a specific job and a visible scorecard. Its lesson is wider than finance. Returns emerge when an agent is embedded in a real workflow, supplied with trustworthy data, and granted only the autonomy that evidence can support.
