RPA vs API: when each one is worth it

RPA vs API: when each one is worth it

· by Panayot Dimkov

RPA or API? The short answer: if the system has an API, integrate through the API, because it is faster, more stable and cheaper to maintain. RPA is worth it when there is no API, when enabling one costs more than the problem, as a temporary bridge or to validate a process before investing in the integration.

RPA is software that uses your applications the way a person would: it opens the window, clicks, types, copies and pastes. It does not know what an invoice is or what a field means; it knows that on that screen, in that position, something has to be typed.

The comparison that explains everything: a programming interface is a door the vendor opened on purpose; RPA comes in through the window. Both get you inside. Only one was designed for you to use.

1RPA vs API: the difference that matters

API integrationRPA
How it gets in Through a channel designed for it, with terms of use Through the same screen a person uses
What breaks it A version change, usually announced in advance A button moving, a new pop-up, or the system running slowly
Speed Milliseconds per operation Seconds: it has to wait for the screen to load
Start-up cost High if it must be developed; low if a ready-made connector exists Low: it is built on what is already there
Cost over three years Stable Rising

That last row decides almost every case. RPA is cheap to build and expensive to maintain; integration is the opposite. If the process is going to last three years, the arithmetic flips.

The rule I use: if an API exists, use the API. RPA is not an equivalent technical alternative, it is what you do when there is nothing else. Treating it as a legitimate design option leads to building castles on screens.

2When RPA is the right answer

01
The system has no API, full stop
Old management applications, niche sector-specific programs, government portals. There is no door, so to integrate a system without an API the window is the only way in.
02
It has one, but it costs more than the problem
Some vendors price the integration module like a project. If the API licence exceeds the expected saving, the calculation inverts.
03
It is a temporary bridge
You are replacing the system within a year and need to survive until then. A robot with a known expiry date is a reasonable decision.
04
You want to validate before investing
Building the robot in two weeks proves whether automating the process helps, and that evidence justifies the real integration. It is an expensive prototype, but a prototype.

3When it is an expensive patch

  • When there is an API and nobody asked. It happens more than you would think: the robot gets built because the team knows the tool, without checking whether the vendor offers integration.
  • When the volume is high. A robot taking fifteen seconds per record does not process fifty thousand records overnight. The arithmetic does not work.
  • When the application changes often. A cloud system that updates monthly is RPA's enemy: every release can move something.
  • When it is used to avoid fixing the process. Putting a robot to copy between two systems that should be connected institutionalises the problem.
That last point is the most common and the most expensive. RPA has a dangerous quality: it makes a broken process work well enough that nobody fixes it. The cost does not appear on the invoice, it appears five years later.

4Attended and unattended RPA

This is the practical distinction that most affects cost and complexity, and it is worth understanding before asking for a quote. In Power Automate, for instance, attended RPA is licensed per user and unattended per bot; the figures are in Power Automate pricing for businesses.

👤
Attended
The robot runs on a person's machine, who launches it when needed and watches what it does. Simpler, cheaper and a good fit for occasional tasks belonging to one person.
🌙
Unattended
It runs by itself, on a dedicated machine, on a schedule or a trigger. This is what people picture when they hear RPA, and it is also what needs different licences, a server and monitoring.
🔐
The credentials problem
An unattended robot needs to log into systems as some user. That user has to exist, have bounded permissions and not be a real person's account. It sounds obvious and it almost never happens.
🖥️
The dedicated machine
An unattended robot cannot share a desktop: if someone moves the mouse or locks the session, it breaks. That computer is infrastructure, not a spare laptop.

5How to make RPA less fragile

If it has to be done, six decisions separate a robot that lasts two years from one that fails in the first week.

01
Identify by attribute, not position
Find the field by its internal name rather than its screen coordinates. It survives layout changes, which are the number one cause of breakage.
02
Wait on conditions
Wait for an element to appear, not a fixed five seconds. A robot that assumes timings fails the day the network is slow.
03
Fail out loud
If something does not add up, stop and report with a screenshot. The dangerous thing is not a robot that halts, it is one that keeps entering wrong data.
04
Idempotent retries
If it stops halfway, it should be relaunchable without duplicating what is done. That is the difference between resuming and checking record by record.
05
A log of what it did
What it processed, when and with what result. Without that trail, a silent failure can go weeks undetected.
06
A plan for when it dies
What happens if the robot is down for two days. If the answer is "nothing, it piles up", that process needed an integration, not a robot.

6What AI changes

Classic RPA is blind: it follows instructions on a screen it does not comprehend. Adding vision and language models changes two concrete things.

The first is tolerance to change: an agent that interprets the screen can find the "Save" button even if it has moved, whereas a rigid selector fails. The second is the ability to decide: reading an unstructured document and acting on it, something that previously required a rule for every case. That is the territory of intelligent document processing.

That said, the essential point should not be lost: this improves the robustness of screen-based access, but it is still coming in through the window. If there is a door, the door is still better.

ToolsIntegrationAutomation

Frequently asked questions

What exactly is RPA?

Software that operates other applications through their graphical interface, just as a person would: opening windows, clicking and typing. It does not integrate with the system internally, it drives it from outside. That is its great advantage — it works with any application — and its great weakness, because it depends on the screen not changing.

How do I know whether my system has an API?

By asking the vendor or your integrator, and searching for "API" or "web services" in their documentation. It is the first check in any project and takes one phone call. Ask what it costs to enable, too, because sometimes it exists but sits behind a paid module.

Is RPA reliable?

It is exactly as reliable as the application it drives is stable. With an internal system that does not get updated, it can run for years untouched. With a cloud application that changes monthly, expect periodic maintenance. Reliability does not depend on the robot but on its environment.

How long does it take compared with an integration?

A simple robot goes live in days or a few weeks; an integration depends on whether the API exists and how well it is documented. That initial gap is real and it is what makes RPA attractive. The mistake is comparing only the start: over three years the relationship usually flips because of maintenance.

Can I combine both?

That is common and advisable. One process can read data through APIs from the systems that offer them and use RPA only for the leg involving the legacy program that does not. That way the fragile part is confined to the minimum instead of building the whole flow on screens.

In short

If an API exists, use the API. RPA is the answer when there is no door: systems with no possible integration, modules costing more than the problem, or bridges with an expiry date. Outside those cases it is usually a patch paid for in maintenance.

And when it has to be done, do it properly: identify by attribute, wait on conditions, fail out loud and have a plan for the day it breaks. That is how we approach AI process automation.

Got a system that refuses to integrate?

Tell us which one and what you need to move. We check first whether an API is available and, if not, design the robot with the fragile part kept to a minimum.

Request a process analysis