Robotic process automation (RPA) has nothing to do with physical robots. It is software that operates other software the way a person would: logging into applications, reading fields on screen, copying values, clicking buttons, filling forms, opening spreadsheets and sending emails. The “robot” is a script running on a desktop or server, following a set of recorded and configured steps. Popular platforms include UiPath, Automation Anywhere, Blue Prism and Microsoft Power Automate, though the concepts are the same across tools.

Why RPA became popular in BPOs

Much back-office work exists because systems do not talk to each other. An associate reads a claim from one system, checks a policy in a second, and updates a third. The proper fix is integration through APIs, but the systems may be old, owned by the client, or expensive to change. RPA works on the user interface, so it needs no changes to the underlying systems. A bot can be built in weeks rather than the months an integration project might take, and it can be pointed at exactly the screens a person already uses.

That is RPA’s strength and its weakness in one sentence. Because it depends on the screen, it breaks when the screen changes.

Attended and unattended bots

Unattended bots run on their own, usually on servers, processing queues of work end to end: reconciling bank statements overnight, generating daily reports, updating records from a file. No human triggers each run.

Attended bots run on an agent’s desktop and are triggered by the agent. During a call, an agent might click one button to have a bot pull the customer’s details from four systems into a single view, or to complete a ten-screen address change while the agent keeps talking. Attended automation reduces handle time and errors without removing the human from the interaction.

Choosing what to automate

Most failed RPA programmes failed at selection. They automated something complex because it was painful, and the bot was as unreliable as the process. A good candidate tends to have these characteristics:

  • Rule-based: decisions can be written as clear if-then rules, without judgement.
  • Structured, digital inputs: data arrives in consistent fields or files, not handwritten forms or free-text emails (unless combined with document-reading tools, discussed below).
  • Stable: the screens, systems and rules are not expected to change for months.
  • High volume and repetitive: enough transactions to repay the build and maintenance effort.
  • Low exception rate: most cases follow the standard path.

A quick sizing helps compare candidates. Take a process with 6,000 transactions a month at 4 minutes each, a workload of 400 hours. If 70% of transactions follow the standard path and can be processed straight through by the bot, 280 hours a month are freed. At about 128 productive hours per associate per month (160 hours at 80% productivity), that is a little over 2 FTEs. Against that sit the costs: licences, development, infrastructure, and above all ongoing maintenance and monitoring. A process that saves two FTEs but breaks every month may not be worth it; one that saves two FTEs and runs untouched for a year usually is.

Benefits are not only about headcount. Automation often matters most where errors are costly, where work must happen out of hours, or where turnaround time is a competitive issue. Some of the best RPA cases are justified on accuracy and speed rather than FTE savings.

Where bots break

Bots fail in predictable ways. A client upgrades its CRM and a button moves; the bot clicks the wrong thing or stops. A password expires. A field that was always populated arrives empty. A pop-up message appears that never appeared during testing. A slow application takes longer to load than the bot waits.

Good RPA design plans for this. Bots should check that each screen is the expected one before acting, handle exceptions by routing the case to a human queue rather than guessing, log every action for audit, and alert a support team when error rates rise. Change management with the client matters too: application changes need to be communicated to the automation team before release, not discovered when the bot fails at 2 am.

There is also a governance angle. Bots usually run with their own credentials and can access sensitive data at scale. Access should be limited to what each bot needs, credentials stored securely, and bot activity included in audit trails, the same as for human users.

Beyond rules: intelligent automation

Pure RPA cannot read a scanned invoice or understand an email. In recent years it has been paired with other technologies: optical character recognition and intelligent document processing to extract data from documents, machine-learning classifiers to route emails, and now large language models to interpret free text and draft responses. This combination, often called intelligent automation or hyperautomation, extends automation into work that was previously considered unsuitable.

These additions change the risk profile. A rules-based bot is either right or it stops; a model that reads documents can be confidently wrong. Confidence thresholds, human review of low-confidence cases and sampled quality checks on automated output become essential, and the exception queue becomes a permanent part of the process design.

Fix the process before automating it

Automating a bad process produces a fast bad process. Before building a bot, it is worth asking whether a step is needed at all, whether two steps can be merged, and whether the input can be standardised at source. Sometimes a simpler form, a validation rule or a small change in the client’s system removes the work entirely. RPA is a strong tool for stable, rule-based work that cannot be fixed at the system level. It is a poor substitute for process redesign.

Written by the RP Strativa Team. Figures in the examples are illustrative.

Leave a Reply

Trending