Process Automation

Robotic process automation promised to eliminate manual work by teaching software to click buttons like a human. Five years into enterprise RPA programmes, many organisations have discovered that automating a bad process just produces bad outcomes faster. The maintenance cost of screen-scraping bots, the brittleness when a UI changes, and the scaling ceiling are all well understood. What comes after RPA, and where does process automation actually create lasting value?

What is process automation?

Process automation is the use of technology to execute business processes with minimal human intervention. It ranges from simple task automation, sending an email when a condition is met, to complex end-to-end workflow orchestration involving multiple systems, decision points, and exception handling. In financial services, the processes that matter most are those that cross organisational boundaries: a customer onboarding process that spans sales, compliance, operations, and IT, or a claims process that involves the customer, the insurer, the repairer, and the regulator.

RPA, the dominant automation technology of the 2015-2022 era, works by recording and replaying user interface interactions. A bot logs into a system, navigates to a screen, copies data, pastes it into another system, and clicks a button. This is effective for stable, high-volume, rule-based tasks. It fails when the underlying system changes its interface, when the process requires judgement, or when exceptions occur that the bot's script does not anticipate. Maintenance costs for mature RPA programmes commonly reach 30 to 40 per cent of the initial development cost annually.

The highest-value automation targets are not the highest-volume tasks but the tasks with the most handoffs. A process that passes through six people across three departments, with each handoff introducing a day of latency and a risk of error, benefits more from automation than a high-volume task performed by a single person. Cycle time reduction, not FTE reduction, is the metric that predicts lasting automation value.

The landscape

The shift from RPA to AI-native automation is underway. Instead of recording UI interactions, modern automation platforms use APIs to connect systems directly, machine learning to handle variability, and orchestration engines to manage the end-to-end workflow. This approach is more robust (no screen-scraping), more adaptable (the model can handle variations), and more expensive to build initially but cheaper to maintain.

Process mining, the analysis of event logs from enterprise systems to discover how processes actually execute, has matured significantly. Tools like Celonis and Microsoft Process Mining reveal the gap between the process as designed and the process as performed. For financial services, this reveals bottlenecks, rework loops, and compliance gaps that are invisible to process owners who see only their part of the workflow. Process mining should precede automation: automate the process as it should work, not the process as it currently works.

Regulatory processes are among the most automation-friendly in financial services. Regulatory reporting, compliance monitoring, and audit evidence gathering follow predictable patterns, involve structured data, and have clear correctness criteria. Automating these processes reduces both cost and compliance risk.

How AI changes this

AI enables automation of processes that require judgement, not just rule-following. An email from a customer requesting a change to their policy can be interpreted by an LLM, classified by request type, and routed to the appropriate workflow, whether that is a simple endorsement that can be processed automatically or a complex request that requires human attention. The AI handles the interpretation; the automation handles the execution.

Document intelligence is the input layer for many automated processes. A process that begins with a document, a claim form, a loan application, a regulatory filing, cannot be automated until the document is converted to structured data. The combination of document intelligence and process automation is more valuable than either capability alone.

Exception handling is where AI transforms automation economics. Traditional automation stops at exceptions, escalating to a human handler. AI can handle a proportion of exceptions by interpreting the unusual case and determining whether it falls within policy or requires genuine human judgement. Increasing the exception-handling capability of an automated process from zero to 50 per cent can double the throughput without adding staff.

Process monitoring and self-optimisation is emerging. AI systems that observe automated process performance, identify bottlenecks, suggest process improvements, and adapt routing rules based on outcome data create a feedback loop that continuously improves the process. This moves automation from a static implementation to a dynamic system that improves over time.

What to know before you start

Automate the right process, not just any process. The right process has high volume, clear rules, multiple handoffs, and measurable outcomes. If you cannot define when the process succeeds and when it fails in measurable terms, you cannot measure the impact of automation. Start with a process that has a clear cycle time metric and a known pain point.

API-first integration is non-negotiable for durable automation. If your automation relies on screen-scraping or UI manipulation, it will break when the underlying system is updated. If the systems you need to connect do not have APIs, the investment in building APIs, or adopting middleware that provides API access, is a prerequisite for automation, not an optional enhancement.

Measure cycle time, not headcount reduction. Automation that reduces the time from customer request to resolution from five days to five hours is unambiguously valuable, regardless of whether it changes headcount. Automation programmes that are positioned as cost-cutting exercises face organisational resistance that automation programmes positioned as service improvement exercises do not.

Start with process mining. Understand how the process actually works before automating it. The gap between the documented process and the actual process is always larger than process owners believe. Mining reveals the rework loops, the informal workarounds, and the exception patterns that must be addressed in the automation design. Automating the documented process without mining the actual process guarantees that you are automating the wrong thing.