RPA and AI: what each one does, and when you need both

The short version

• RPA repeats fixed, rule-based steps across software and makes no judgements of its own.
• AI models read unstructured input such as emails and documents, and their output can be wrong.
• An RPA bot tends to fail visibly when a screen changes, while an AI model can fail quietly with a plausible wrong answer.
• RPA and AI usually combine as a model that reads and classifies, a validation check, and a bot that enters the result.
• Agentic AI chooses its next step at run time, which suits variable work and is overkill for a process you can write down as rules.

What is the difference between RPA and AI?

RPA follows rules and AI makes judgements. Robotic process automation is software that repeats fixed steps across other software: open this screen, copy this field, paste it there. Artificial intelligence, in this context, usually means a model that reads unstructured input and produces its best estimate of what that input means.

That difference decides how each one fails. Given the same input and the same screens, an RPA bot does the same thing every time, so you can test it like any other program. A model may return a different answer to the same input, and a convincing answer can still be wrong. You test a model by measuring it on samples, not by checking that a script passes.

RPA and AI are not really competitors, because they answer different questions. RPA answers 'how do we do this step without a person?' and AI answers 'what does this email, form or document say?'. Many real processes contain both questions.

How do RPA and AI compare side by side?

What does a bot in RPA actually do?

An RPA bot operates other applications through the same interfaces a person uses. It logs in, clicks, types, reads fields and moves files. The usual reason to run one is that the target system has no usable API.

RPA bots come in two forms. Attended bots run on a person's desktop and work alongside that person. Unattended bots run by themselves, usually from a schedule or a queue. Microsoft prices them differently: attended desktop flows sit in a per-user plan and unattended capacity is sold per bot, with one bot executing one unattended desktop flow run at a time.

None of this involves intelligence. A bot does not know what an invoice is. It knows that a value sits in a certain field on a certain screen. On a stable process that is a strength, because a bot will not improvise.

How is AI used in RPA?

AI is used in RPA at the points where a bot would otherwise need a person to read something. A model reads the incoming email, invoice or form, extracts the fields and classifies the request. The bot then enters those values into the system of record using its fixed steps.

RPA with AI works best when there is a check between the two. Model output should pass through validation before a bot acts on it: required fields present, totals that add up, a supplier that exists. Anything that fails goes to a person. Without that check, a bot will enter a wrong value as faithfully as a right one.

Other common uses follow the same pattern. A model can route a support request, summarise a case for the person handling an exception, or suggest a category that a bot then applies. In each case the model supplies an interpretation, and deterministic code decides what happens next.

Does AI fix the problems with RPA?

AI fixes some RPA problems and adds others. It helps with input variation, because a model can often read a new invoice layout where a fixed template would fail. It does not stop a bot breaking when an application changes its screens.

Upkeep is the long-standing complaint. A Pega poll released in 2019, covering 509 decision makers at businesses using RPA or robotic desktop automation, all with at least $100 million in revenue, was headlined as finding RPA effective but hard to deploy and maintain. Treat that as a signal, not a measurement. Pega sells a competing approach, the firms are large and the poll is old.

The mechanism is easy to see. A bot depends on things it does not control: screen layouts, field names, login flows and other teams' release schedules. Adding a model adds more dependencies: model versions, prompts and changes in the input. You now have two components to monitor, and they fail in different ways. That can still be worth it, but it is not a simplification.

What are the risks of letting AI drive an RPA bot?

The main risk is that the bot's permissions become the model's permissions. A bot typically holds a service account with write access to real systems. Once model output decides what the bot does, anything that can influence the model can influence those actions. That includes the documents it reads, because text inside an email or PDF can carry instructions. This is known as prompt injection.

The controls are ordinary engineering. Give the bot the narrowest account that works. Treat model output as untrusted input and validate it. Let the model choose from a fixed set of actions instead of composing commands. Require a person's approval for anything hard to reverse, such as a payment or a deletion. Log the input, the model's output and the action taken, so a wrong entry can be traced later.

What is the difference between RPA and agentic AI?

RPA vs agentic AI comes down to who decides the next step. In RPA the developer fixes the path when the bot is built. In agentic AI a model chooses the path at run time, picking which tool to call based on what it has seen so far.

An agent suits work where the steps really do vary, such as an enquiry that might need an order lookup, a refund check or an escalation. It is a poor fit for a process you can write down. You would pay in run-time cost, test effort and audit complexity for flexibility you never use.

RPA and agents also stack. An agent can call an RPA bot as a tool when the target system has no API, so the model decides and the bot performs a known action. That gives the agent a narrow, predictable way to act. Vendors are blurring the labels too: the title of UiPath's pricing page reads 'Scalable Agentic Automation Solutions'. Judge a product by what decides the next step at run time, not by what it is called.

When should you use RPA, AI, both or neither?

Use RPA for stable, rule-based steps with no API. Use AI where the hard part is reading or judgement. Use both when a process has each kind of step. Use neither when a direct integration, or removing the step, would do the job.

Check first whether the step needs to exist. Automating a redundant approval gives you a faster redundant approval.

Which option fits which process?

• •Use an API or plain code if the target system offers one. It is less fragile than a bot driving a screen.
• •Use RPA if the steps are fixed, the input is structured, the application rarely changes and there is no usable API.
• •Use an AI model if the hard part is reading or classifying unstructured input, and a mistake can be caught before it matters.
• •Use both if a model's reading feeds a fixed sequence of entries in a system without an API.
• •Use an agent only if the sequence of steps cannot be known in advance.
• •Do nothing if the process runs a few times a month. The upkeep is likely to cost more than the time saved.

What do RPA and AI cost?

Published entry prices look small, and they are not the main cost. The ongoing spend is in upkeep: fixing bots when applications change, checking model accuracy as inputs change, and paying people to handle exceptions.

Here are two vendor examples, both read in November 2026. Microsoft lists Power Automate Premium, with attended desktop flows, at $15.00 per user per month, paid yearly, and says actual prices are confirmed at checkout. UiPath's pricing page shows $25 for its Basic plan but does not say whether that is per user or state the billing term. Basic is limited to 5 Basic, 5 Plus and 1 Pro User, and 2 robots, and users and robots may require additional licence purchases. UiPath also says an updated plan offering is coming soon, so recheck the page before you rely on this.

Those prices do not tell you what a process costs to run. Where a model is involved, model usage is a further cost that scales with volume.

Payback figures need the same care. Deloitte's 2021/22 Global Intelligent Automation survey of 479 executives from 35 countries gives pilot payback figures of 16 and 22 months. Those are pilots, not finished programmes, and company sizes are not stated, so they are not a forecast for your firm. We have no payback results of our own to quote.

How many businesses combine RPA and AI?

We do not know, and we have not found a reliable figure that measures the combination. The figures we can cite measure AI use in general and say nothing about RPA.

Eurostat reported that 8% of EU enterprises used AI technologies in 2023, in a survey covering enterprises with at least 10 employees and self-employed persons. The US Census Bureau's Business Trends and Outlook Survey, using data collected from December 2025 to May 2026, reports that firms with at least 20 employees are the biggest AI users. It asks whether a firm used AI in any business function in the past two weeks, which is a low bar.

These figures do not tell a smaller firm that it is behind, and they do not tell it what to do. The process in front of you is better evidence.