.svg)
61% of finance leaders admit their organisations rolled out AI agents largely as experiments to test capabilities rather than to solve business problems.
That figure, from a Basware and FT Longitude joint report, is the most honest summary of where most finance functions are with agentic AI in 2026. Interest is high. Pilots are everywhere. Production deployments that deliver measurable ROI are significantly rarer.
The gap between experimentation and operation is not a technology problem. The technology works. The same report found that autonomous agents delivered an average ROI of 80%, compared to 67% for general AI projects. IDC's research puts the average return at 2.3x within 13 months for organisations that move to production deployment.
The question is not whether agentic AI in finance delivers ROI. It is why so many finance teams cannot get there.
The pattern is familiar to most finance leaders: a compelling vendor demo, an internal champion, a carefully scoped pilot. The pilot runs. Results look promising. The next step, scaling to production, stalls.
The stall happens for consistent reasons. The pilot ran on clean data that was prepared specifically for it. The production environment has messier data. The pilot covered one process in one entity. Production requires handling multiple entities, currencies, and approval hierarchies. The pilot team was hand-selected and motivated. The production rollout requires embedding a new way of working across a finance function with existing habits and reasonable scepticism.
Bain's 2025 Technology Report is direct about what causes pilots not to scale: "Fragmented workflows, insufficient integration, and misalignment between AI capabilities and business processes. Many early ambitions, such as 30 to 50% efficiency improvements, haven't materialised due to orchestration gaps." The agent works. The surrounding architecture does not support it at scale.
40% of agentic AI projects fail due to inadequate infrastructure foundations, according to a cross-industry analysis by Landbase. In finance specifically, the foundation requirements are well-understood: data quality, ERP integration, governance frameworks, and the process design that tells the agent what it should do and when to escalate.
The second most common reason finance teams stay in pilot mode is governance concerns. 46% of finance leaders will not consider deploying an agent without clear governance in place. That caution is rational, particularly in regulated environments where audit trails and accountability matter.
The paradox is that the finance teams with the most sophisticated governance are the ones moving fastest to production, not the ones staying in pilots. As Anssi Ruokonen, Head of Data and AI at Basware, put it: "These leaders are significantly more likely to use agents for complex tasks like compliance checks (50%) compared to their less confident peers (6%). They use governance to scale, not to stop."
Governance is not a prerequisite that must be perfect before deployment begins. It is a framework that is built, tested, and refined alongside the deployment. Waiting for perfect governance before moving to production is a form of expensive caution that produces neither safety nor results.
The defining characteristic of agentic AI is not that it processes faster. It is that it initiates. Traditional automation waits for a human to trigger a process and then executes it. Agentic AI monitors conditions, recognises when action is required, and takes that action without waiting for a prompt.
In accounts payable, the practical difference is significant. An automated AP system processes an invoice when someone uploads it. An agentic AP system monitors the invoice inbox, ingests new invoices as they arrive, extracts the data, matches against the purchase order and delivery note, routes for approval according to the approval rules, sends reminders when approvals are pending, and flags exceptions to the relevant person, all without anyone initiating any step.
The finance team is not processing invoices. The agent is processing invoices. The finance team is managing exceptions and reviewing what the agent has flagged.
Kevin Kamau, Director of Product Management for Data and AI at Basware, describes AP as a "proving ground" for agentic AI because it "combines scale, control, and accountability in a way few other finance processes can." The tasks that agentic AI handles reliably in production AP deployments include:
The most effective agentic finance operations are not those that minimise human involvement. They are those that concentrate human involvement on the decisions that require it.
The agent processes the routine. The finance team handles the genuine exceptions: a supplier dispute that requires relationship judgement, a payment that involves contractual interpretation, a compliance question that requires legal input. The quality of the team's work on those decisions improves because they are not spending 80% of their time on data entry.
82% of midsize companies that have adopted agentic AI report that it improved their operational efficiency and workforce productivity, according to Citizens Bank's 2026 survey. The improvement is not from reducing headcount. It is from redirecting the same headcount toward work that requires human capability.
Beyond the infrastructure and governance issues above, three specific failure points account for the majority of stalled agentic AI deployments in finance:
The data preparation gap. Pilots are run with data that has been cleaned specifically for the pilot. Production environments use the actual data, which includes duplicate vendor records, inconsistent supplier names, missing fields, and historical entries that were never corrected. The agent that performed beautifully in the pilot struggles with the entropy of real production data.
The scope creep problem. A pilot that worked for one invoice type is expanded to cover all invoice types before the underlying process is designed for the full scope. The agent encounters edge cases it was not configured to handle, generates exceptions that the team does not know how to resolve, and trust in the system erodes before it has the opportunity to demonstrate its full capability.
The integration assumption. Many pilots are run on a subset of data that was manually exported from the ERP. Production requires a live, bidirectional integration that keeps the agent's view of vendor records, purchase orders, and approval hierarchies current in real time. When that integration does not exist or is shallower than required, the production deployment produces a different quality of output than the pilot.
48% of organisations cite data governance as their primary agentic AI implementation challenge, according to KPMG's 2025 analysis. The specific data quality requirements for finance agents are well-defined:
The data readiness review is the most valuable investment before production deployment. Two to three weeks spent on data assessment and clean-up before go-live prevents months of exception management after it.
The finance professionals most resistant to agentic AI adoption are typically the most experienced. Their resistance is not irrational. They have built their professional value around the detailed knowledge of the AP process that they have accumulated over years. An agent that handles the process automatically appears to devalue that knowledge.
The reframing that works in practice: the agent handles the volume that was consuming their time. Their knowledge is now available for the work that the agent cannot do: exception investigation, supplier relationship management, process improvement, and the strategic financial analysis that manual processing left no time for.
The finance teams that get this right involve their most experienced members in the design of the agent's decision rules. Their knowledge of the edge cases, the supplier quirks, the exceptions that occur regularly, becomes the configuration that makes the agent work correctly in production. Their institutional knowledge is not replaced. It is codified.
ROI measurement needs to be structured across time because the returns compound. The metrics that are meaningful at each stage differ.
At 90 days:
At 6 months:
At 12 months:
IDC's research is specific about the divergence in returns: "Frontier firms leading in AI adoption achieve returns of 2.84x on their investments, compared to just 0.84x for laggards." The difference is not that frontier firms deployed better technology. It is that they deployed earlier, accumulated more transaction history in their agents, and their agents have been improving on that data for longer.
Gartner's projection that 33% of enterprise software will include agentic AI by 2028 creates a specific competitive dynamic: the businesses that are production-ready today will have 2 to 3 years of agent improvement built into their performance by the time the majority of the market reaches the same point. That lead does not close easily.
The most immediate change is visibility. Agentic AP creates a real-time view of every invoice in the process: where it is, what its status is, what action is required, and when it is due. That visibility replaces the patchwork of spreadsheets and email chains that most manual AP environments use to track invoice status, and it does so from the first week of operation.
The second change is exception quality. When the agent handles routine processing, the exceptions that reach the finance team are genuine ones: price mismatches that require supplier contact, delivery discrepancies that need warehouse verification, invoices from new suppliers that warrant additional scrutiny. The team stops spending time on routine checks that the agent should handle, and starts spending it on the issues that actually require their judgement.
At 12 months, the finance teams running agentic AP consistently describe the same shift: the month-end close has changed character. Where the close previously involved a significant data assembly exercise, reconciling invoices, matching payments, and preparing reports, the agent has been doing this work continuously throughout the month. The close is a review and confirmation, not a construction project.
The second change at 12 months is the quality of supplier relationships. On-time payment rates above 95%, achieved through automated payment scheduling rather than manual runs, produce measurably better supplier terms over time. Suppliers who are consistently paid on time extend better pricing, more flexible terms, and priority allocation in constrained supply situations.
The distinction between an AI-native platform and a platform with AI features is the difference between agentic AI that works from day one and agentic AI that requires months of training before it reaches production accuracy.
An AI-native platform was built with autonomous decision-making as the design goal. Every component, from data extraction to three-way matching to approval routing, was designed to operate with minimal human intervention. The exceptions are the edges, not the normal path.
A platform where AI features were added to a rules-based system operates differently. The rules handle the normal path. The AI features handle specific augmentation tasks. The system still requires human intervention for the cases the rules did not anticipate, which is a higher proportion of cases than an AI-native system would generate.
Dost's AP and AR platform processes the first invoice it receives with the same autonomous logic it applies to the ten-thousandth. There is no training period in which the team manually processes invoices while the system learns. There is no template library that needs to be populated before new supplier formats can be handled. The agent is operating from day one.
The straight-through processing rate improves over time as the agent builds supplier-specific baselines, but the starting point is not zero. Most Dost deployments achieve 70 to 80% straight-through processing in the first month, reaching 85 to 90% by month six as supplier patterns are established.
For a well-scoped agentic AP deployment, the path from pilot to production is typically 8 to 12 weeks, assuming the data readiness work is done properly upfront. The most time-consuming element is usually the ERP integration and the vendor master clean-up, not the agent configuration itself. Finance teams that try to skip the data preparation step consistently take longer in total, because data quality issues surface as exceptions during production rather than being resolved before go-live. The 6-month ROI timeline cited in enterprise AP research assumes a 2 to 3 month implementation period followed by 3 to 4 months of ramping performance.
For platforms designed for mid-market finance teams, the ongoing operation of an agentic AP workflow requires no technical resource. The finance team configures and maintains the approval rules, cost centre allocations, and exception routing through a self-service interface. The ERP integration is maintained by the vendor. The agent's learning from transaction data is automatic. Technical resource from IT is typically required only for the initial integration setup, which in a native connector deployment takes two to four weeks. After go-live, IT involvement in day-to-day operations should be zero.
The governance framework for agentic finance AI covers four areas. First, decision boundaries: a clear definition of what decisions the agent makes autonomously and at what threshold it escalates to a human. Second, audit trail: every action the agent takes is logged with a timestamp, the data it relied on, and the rule it applied. Third, exception review: a defined process for reviewing the exceptions the agent surfaces, with clear ownership and resolution timelines. Fourth, performance monitoring: regular review of the straight-through processing rate, exception rate, and accuracy metrics, with a process for adjusting the agent's configuration when performance deviates from expectations. This framework does not prevent agentic deployment. It makes it sustainable and auditable, which is what regulated finance environments require.
Agentic AI for finance is not a future aspiration. Finance teams that have moved to production are reporting 80% average ROI, 2.3x returns within 13 months, and the compounding performance advantage that comes from an agent that improves with every transaction it processes.
The teams still in pilot mode are not waiting for the technology to mature. They are waiting for their internal conditions to be right: clean data, integrated systems, governance frameworks, and the organisational confidence to move from test to operation.
Those conditions are achievable, and the path to them is well-understood. The finance teams that resolve them fastest are the ones building the performance advantage that compounds over the 2 to 3 year horizon that separates frontier firms from laggards.
The cost of staying in pilot mode is not zero. It is the difference between 2.84x returns and 0.84x returns, measured against the same investment, over the same period.
