Robotic process automation (RPA) consists of configuring software — the so-called "bots" — to carry out repetitive digital tasks by mimicking a person's interaction with applications: clicking, reading screens, copying data between systems, filling in forms or reconciling records. Unlike an API integration, an RPA bot operates at the presentation layer, that is, on the user interface itself, which allows it to connect legacy applications that expose no programmatic interface. That is its great strength and, at the same time, the source of its main fragility.
What RPA is and what it is not
RPA is not artificial intelligence in its own right. In its classic form it is deterministic, rule-based automation: the bot does exactly what it has been programmed to do, without learning or handling ambiguous decisions. It works superbly for high-volume, structured processes with clear rules, and it fails when unstructured content must be interpreted or judgement calls must be made. For this reason it is worth distinguishing three levels:
- Attended RPA. The bot runs on the employee's workstation and assists them in real time; human supervision is required.
- Unattended RPA. The bot runs autonomously on servers on a schedule, without human intervention — ideal for overnight batch processes.
- Intelligent automation (AI + RPA). Combines bots with computer vision and natural language processing to handle unstructured inputs, such as reading a PDF invoice (IDP, intelligent document processing).
Use cases with proven return
The processes best suited to RPA share a common pattern: high volume, stable rules, multiple applications involved and low tolerance for human error. Some concrete examples by department:
| Department | Automatable process |
|---|---|
| Finance | Bank reconciliation, supplier invoice posting, accounting closes |
| Human Resources | Employee onboarding and offboarding, payroll management |
| Customer service | Ticket creation, CRM data updates, standardized responses |
| Supply chain | Order processing, shipment tracking, stock updates |
| Compliance | Data extraction for regulatory reports, KYC verification |
How to select the right process
The single strongest predictor of RPA project success is not the platform chosen but the selection of the process. A strong candidate meets several criteria: it is repetitive and high-volume, follows clear and stable rules, works with structured, digital data, has few exceptions and runs on stable interfaces that do not change every week. Before automating it is worth applying process mining to discover how the process is actually carried out (not how the manual says it should be) and, in many cases, to redesign it first: automating an inefficient process merely produces inefficiency at greater speed.
Implementation methodology
- Discovery and prioritization. Identify candidates and score them by automation effort versus expected benefit.
- Analysis and design (PDD). Document the as-is process and the target to-be process in a process definition document.
- Bot development. Build the workflow in an RPA platform, modularizing reusable components.
- Testing. Validate the happy path and, especially, exception handling and boundary data.
- Deployment and orchestration. Publish the bot, schedule it, and manage work queues and credentials in a secure vault.
- Monitoring and maintenance. Track success rates, execution times and failures; keep the bot aligned with changes in the underlying applications.
How the bot locates screen elements
The technical core of an RPA bot is its ability to reliably identify the elements it interacts with: a text field, a button, a table cell. There are three strategies, ordered from least to most robust. Coordinate-based location (clicking pixel X,Y) is the most fragile: any change in resolution or layout breaks it. Image recognition, which searches for a visual pattern on the screen, is slightly more resilient but fails with theme or font changes. The preferred option is object selector anchoring (the internal identifier of the control in the DOM or application accessibility tree), which remains stable even if the element changes position. Designing bots with resilient selectors rather than coordinates is the technical decision that most reduces long-term maintenance cost.
When the process involves reading unstructured documents, intelligent document processing (IDP) comes into play, combining OCR (optical character recognition) with language models to extract fields from invoices, delivery notes or contracts. Unlike classic OCR, IDP does not depend on rigid templates: it learns to locate the total amount or the VAT number even when each supplier presents the invoice in a different layout. This is the component that transforms a deterministic bot into automation capable of handling the variability of the real world.
How to calculate return on investment correctly
The ROI calculation for an RPA project must go beyond hours saved. An honest formula includes, on the benefit side, the time freed up valued at real cost, the reduction in errors and rework, and the improvement in compliance; and on the cost side, not only licences and initial development, but also ongoing maintenance, which is structurally high in RPA because every update to the underlying applications may require adjustments to the bot.
| Metric | What it indicates |
|---|---|
| Bot success rate | Percentage of executions completed without human intervention |
| Exception rate | Percentage of cases escalated to a person; if persistently high, the process was a poor choice |
| Average execution time | Compares bot speed against the manual process |
| Maintenance cost | Technical hours spent repairing and adjusting the bot |
A persistently high exception rate is the most important warning sign: it indicates that the process had too much variability to automate by rules alone and that an AI layer or a redesign should have come first. Tracking this metric from day one prevents teams from keeping bots alive that generate more supervisory work than they save.
Governance: the centre of excellence (CoE)
RPA programmes that scale beyond the first few bots typically rely on a centre of excellence — a structure that centralizes development standards, credential management, version control and metrics. Without this governance, uncontrolled bot proliferation (bot sprawl) sets in: dozens of ad hoc automations with no documentation that nobody knows how to maintain, which break in a cascade when an application is updated. The CoE also manages the security risk, because a bot handles credentials that often carry elevated permissions.
RPA, intelligent automation and hyperautomation
Classic RPA has evolved towards intelligent automation, which adds cognitive capabilities for handling unstructured inputs (emails, PDFs, images). When this approach is combined with process mining, workflow management and orchestration at the scale of the entire organization, it is called hyperautomation. The practical boundary is clear: RPA moves and types data according to rules; the AI layer decides and interprets when a rule is not sufficient. Both complement each other — they do not compete.
Common mistakes
The first is automating a broken process without optimizing it first. The second is selecting processes with too many exceptions, where the bot spends more time escalating cases to humans than resolving them. The third is fragility when the interface changes: a bot that locates buttons by screen coordinates breaks as soon as the layout changes; anchoring to stable selectors is essential. The fourth is neglecting credential security, storing them in the workflow itself instead of in a secure vault. The fifth is measuring ROI incorrectly by counting only hours saved and ignoring maintenance cost, which in RPA is significant and ongoing.
Frequently asked questions
Does RPA replace people? In practice, it redistributes work: the bot takes on the repetitive task and the person handles exceptions, judgement calls and customer relationships. The realistic objective is to free up time from low-value tasks, not to eliminate jobs.
When is an API integration better than RPA? If the applications expose stable APIs, direct integration is more robust and maintainable. RPA shines precisely when there is no API and the only way to connect systems is through the user interface of a legacy application.
How long does it take to see a return? Simple, high-volume processes typically pay back within a few months, but the calculation must include ongoing maintenance. A bot is not an asset you install and forget: it requires monitoring and adjustment every time the underlying applications change.
Do I need AI to do RPA? Not for classic rule-based cases. AI is only necessary when the process involves interpreting unstructured content or making ambiguous decisions; at that point you are moving into intelligent automation.
Conclusion
RPA is neither magic nor intelligence: it is a tireless, literal and fragile digital worker that delivers outstanding results when assigned the right process and surrounded by proper governance. The decisive factor is not the platform chosen but two prior disciplines: selecting high-volume processes with stable rules, and redesigning them before automating in order not to perpetuate inefficiencies. Programmes that fail usually do so because they automated the wrong things or allowed bots to proliferate without control or maintenance. At Summum we support the discovery and prioritization of processes, build bots with maintainable components and set up the governance structure that prevents bot sprawl, so that automation continues to deliver value long after the first deployment.