Finance workflows combine several kinds of work. Moving an approved value between systems is different from interpreting an email about a disputed charge. Checking an approval threshold is different again from deciding whether to make a commercial exception.
Those differences should shape the automation. RPA, AI agents, conventional software and human review each have a useful role. The investment decision is how to combine them into a process that completes the work accurately, economically and within the company's controls.
Curia's managed approach starts with that process. It redesigns, builds and runs the agreed workflow while the finance team retains approval authority, exceptions and judgement.
Where RPA works well
Robotic process automation follows specified steps to interact with software, often through the same interfaces a person uses. It can be useful for predictable, repetitive work, especially where an application has no suitable integration interface.
Examples include downloading a standard report, entering an approved journal or transferring validated records between systems. When the rules are clear and the environment is stable, a predefined sequence can be easier to test and control than a system that chooses its next action dynamically.
UI automation does, however, depend on the application environment. Microsoft's Power Automate documentation identifies application updates, interface changes, permissions and display settings among the factors that can affect selectors used to locate controls. It also describes ways to improve resilience. Maintenance is part of operating this technology. Microsoft, Troubleshoot unattended desktop flow execution failures
AI agents that operate through interfaces can encounter similar dependencies. Adding a language model does not remove the need to manage integrations, access and application changes.
Where agents can add value
An AI agent can interpret context and select among permitted actions as a case develops. That is useful when inputs arrive in different formats or the next step depends on information that must first be gathered.
In finance, candidate activities include interpreting correspondence about a short payment, locating evidence across documents, identifying the likely reason for a mismatch and preparing an explanation for an approver. Each capability must be tested against the organisation's actual material and rules.
There is credible evidence that AI assistance can improve work involving varied language and cases. A study published in the Quarterly Journal of Economics in 2025 examined 5,172 customer support agents and found a 15% average increase in issues resolved per hour following access to an AI assistant. Results varied substantially by worker. This was a customer support study, so the percentage should not be used as a finance automation forecast. Brynjolfsson, Li and Raymond, Generative AI at Work
The study supports testing AI within a defined operational task and measuring the result. It also illustrates that useful assistance can create value while a person remains responsible for the interaction.
The market already combines agents, robots and people
The distinction between RPA and agents is useful for designing a workflow, but it no longer neatly separates vendors. UiPath's Maestro documentation describes orchestration across AI agents, robots and people, including governance and human involvement. A contemporary evaluation should examine the capabilities of the actual platform and service being offered. UiPath, What you get with Maestro
For a finance team, the following allocation is a practical starting point:
| Work to be done | Suitable starting approach | What to verify |
|---|---|---|
| Calculate a variance or check an approval limit | Conventional code and explicit rules | Correct inputs, policy version and boundary cases |
| Transfer approved records through a stable integration | API or existing system integration | Permissions, reconciliation and duplicate prevention |
| Perform repeatable steps in a legacy application | RPA | Interface stability, failure detection and recovery |
| Interpret correspondence or assemble a case explanation | AI assistance or a bounded agent | Source accuracy, relevant context and correct escalation |
| Approve a sensitive action or exercise commercial judgement | An authorised person | Clear evidence, decision rights and recorded approval |
These are design recommendations, rather than fixed restrictions. The appropriate combination depends on the consequences of error and the performance demonstrated in testing.
More autonomy creates a larger testing obligation
Autonomy should be granted deliberately. In the Bank of England and FCA's 2024 survey of UK financial services firms, 55% of reported AI use cases involved some automated decision-making, while 2% were fully autonomous. These figures cover AI broadly in regulated financial services; they are not a survey of corporate finance agents. They nevertheless provide a useful example of adoption with differentiated levels of control. Bank of England and FCA, Artificial intelligence in UK financial services—2024
For an accounts receivable dispute, a bounded agent might collect the invoice, delivery evidence and customer correspondence, then prepare a proposed resolution. Approval to issue a credit can remain with a person. The permitted actions and financial thresholds should be explicit.
A similar principle applies to complexity. Anthropic's engineering guidance recommends starting with the simplest effective solution and adding agentic behaviour where the flexibility justifies the extra cost and latency. Anthropic, Building effective agents
For finance, the practical test is whether each autonomous step removes worthwhile work while preserving a control the business can understand and verify.
Test the completed workflow, including its failure paths
A sequence of individually impressive results can still produce disappointing overall reliability.
For illustration, assume a workflow has ten steps, each with an independent 99% probability of success, and every step must succeed. Without retries or recovery, the probability of complete success is 0.99 to the power of 10: approximately 90.4%. At 99.9% per step, the same simplified calculation gives approximately 99.0%.
Real failures can be correlated, and well-designed workflows include validation and recovery. This calculation is not a measured agent failure rate. It demonstrates why testing isolated tasks is insufficient.
An evaluation should include the final system records, required approvals and evidence trail. It should also establish what happens when a source document is missing, two sources conflict, a connected system is unavailable or an action is attempted twice.
For sensitive updates, recovery deserves particular attention. If a connection fails after an action may have succeeded, retrying blindly could create a duplicate. The workflow needs a way to check the actual state before continuing.
Give the business a scorecard it can use
The strongest measure is the cost of a correctly completed case, including the human effort needed to review, correct and recover it. Track that alongside elapsed time, the proportion of cases escalated, control failures and evidence completeness.
An escalation rate should be interpreted carefully. Routing an ambiguous case to a person can be the correct outcome. A low escalation rate becomes useful only when the workflow also demonstrates that it recognises situations outside its authority or capability.
Compare results by case type. Easy cases can dominate an average and conceal poor performance on the exceptions that consume most of the team's time. This analysis helps determine where further automation is valuable and where human involvement remains appropriate.
Where Curia fits
An organisation with an established automation team may be able to design and operate this combination itself. A team buying a platform should explicitly allocate responsibility for process design, integration, controls, testing and maintenance.
Curia provides a managed route through that work. It assesses the opportunity using the client's economics, redesigns the process, builds and tests the agreed workflow in the client's environment, then operates it as a service. The finance team keeps the decisions and sign-off that belong to it.
This is particularly relevant when the bottleneck spans several systems or departments and nobody has the capacity to lead the complete change. Curia's proposition brings the process work and ongoing operation into the same engagement.
Bring Curia a process that repeatedly stalls between systems or people. We will assess which steps can be automated, where judgement should remain and what evidence is needed to justify the investment. Discuss a finance workflow with Curia