RPA or AI agents? The difference and when to choose which

RPA stands for robotic process automation: software robots that operate a user interface the way a person does. A bot opens a program, clicks buttons, copies a value from one screen and pastes it into another. Well-known names are UiPath, Blue Prism and the desktop flows in Microsoft Power Automate.

The bot follows rules that were set in advance, often by recording an employee's actions and then refining them. The bot knows where a field is and what goes in it, but not what it means. That's its strength, and also its weakness.

What RPA is good at

• Work that goes exactly the same way every time
• High volumes, including at night and at weekends
• Systems without an API, such as an older ERP or the web portal of a party you have no integration with
• Results without having to change the underlying system

Where RPA gets stuck

• A screen changes. A software update, a button that moves, a field with a new name. The bot looks for something that's no longer there and stops, or fills in the wrong field.
• The input is different. An invoice in a different layout, an order in the body of an email, a customer who fills in the form differently.
• A judgement call is needed. Is this a complaint or a question? Can this exception go through? A rule can only handle that if someone thought of the situation in advance.

That's why RPA needs a lot of maintenance. Every change to a screen is a change to the bot. There's also a short definition in our glossary.

An AI agent understands what's in front of it. The agent reads an email, a PDF or a free-text question, pulls out what it needs and decides, within fixed limits, what the next step is. The brain is a language model. The hands are connections to your systems.

That makes an agent strong where RPA is weak:

• Unstructured input. Every supplier sends a different invoice, every customer writes differently. An agent reads the meaning, not the position of a field.
• Variation and exceptions. An agent can deal with a situation nobody wrote down as a rule in advance, or recognize that it's in doubt and ask a person.
• Judgement. Summarizing, categorizing, drafting a reply, making a proposal.

On the other hand, an agent is less predictable than a rule. An agent can get things wrong, and sound convincing while doing it. That's why we build agents with fixed limits, restricted permissions and a person who approves where it matters. That's called human in the loop.

An example from our own work. At an enterprise client, 6,000 customer questions come in each month. At first, 30 percent of them had a prepared answer; now it's 80 percent, ready before the ticket opens. An employee always checks and sends. You don't build that with fixed rules: every question is worded slightly differently.

How an agent works is explained in what is an AI agent.

Between RPA and AI sits a third technique that often gets forgotten in this discussion: the workflow. A workflow tool such as n8n, Make or Zapier connects systems through their APIs, the fixed entrance through which software talks to other software. Something happens, such as a new order or a completed form, and then fixed steps follow.

The difference from RPA: a workflow doesn't operate a screen, it talks directly to the system. A button that moves or a screen that looks different after an update makes no difference. As long as the system has an API, a workflow is more stable and lighter to maintain than a bot.

A workflow can also contain AI steps: read this PDF and return the order number. And we usually build an agent inside a workflow, so there are fixed steps, checks and an approval around it.

A practical rule of thumb: if the system has an API, you don't need a bot for that part. So for existing bots, check whether the software has an API by now. How to connect per system is explained in connecting AI to your systems. Which workflow tool fits is in n8n, Make or Zapier.

By giving each technique the part it's good at. An example: invoice processing at a company with an older finance package without an API. The purchasing system does have an API.

1. 01Workflow. An invoice arrives in the accounts payable inbox. The workflow picks up the PDF and starts the process.
2. 02AI. An AI step reads the invoice: supplier, amounts, VAT and purchase order number. Even from a supplier with a layout it has never seen.
3. 03Workflow. Through the purchasing system's API, the workflow checks whether the purchase order exists and whether the amount is right. Mismatches go to a separate list.
4. 04RPA. A bot enters the booking as a draft in the old finance package, field by field, the way an employee would.
5. 05Human. Accounts payable approves with one click, or corrects it first.

The bot doesn't have to read, the AI doesn't have to click, and the person only looks at what matters. If the finance package gets an API or is replaced, you only replace step 4.

Already running RPA?

Don't throw it away. Look at where your bots get stuck or need a lot of maintenance. Often that's exactly where the input is unstructured: an email that's slightly different, a PDF with a different layout. Put AI in front of the bot there, turning the input into fixed fields. The bot then does what it always did.

What reading unstructured input delivers

Work where AI reads what comes in and prepares it in your own systems saves clients a lot of time:

• Van Berkel Professionals. More than 100 requests a day arrived by email and form and were retyped by hand. Now the flow is read, filtered and placed in their own systems, and a person approves every request. That saves 192 hours a month.
• EasyToys. Product information arrives in supplier PDFs and emails. AI reads them, writes the product copy in six languages and prepares everything in the ERP. That saves 132 hours and € 6,600 a month.

In both cases the input is different every time. Fixed rules on fixed screens won't get you far there.

It rarely goes wrong on the technology. It goes wrong on the choice, or on what happens after go-live.

• Choosing by vendor instead of by step. Whoever sells RPA sees bots everywhere. Whoever sells AI sees agents everywhere. Start with the process and choose per step.
• Automating a messy process. If three colleagues do it three different ways, you automate the chaos. Agree on the process first.
• Putting RPA on unstructured input. A bot that reads invoices from ten suppliers gets stuck on the eleventh layout. That calls for an AI step.
• Using an agent where a rule will do. If every step is fixed, a workflow is cheaper and more predictable than AI doing the same thing.
• Clicking where there's an API. A bot on the screen while the system offers an integration is needlessly fragile.
• No owner and no monitoring. A bot goes quiet after a software update, an agent after a change to an integration. Without an owner, nobody notices until a customer calls.
• Too many permissions. A bot or agent running under an account with access to everything can also get everything wrong. Give it only the permissions it needs for that one task.
• No person in the right place. Not at every step, but where a mistake costs money or affects a customer.

With all three you pay for building, for licences or usage, and for maintenance. The difference is where the maintenance comes from. So don't only calculate with the price of the software, but with what it costs to keep it running for a few years.

Costs

• RPA. Licences, building and above all maintenance. Every change to a screen, every software update and every new exception means work on the bot.
• Workflow. A licence or hosting, building and maintenance. As long as the integrations stay the same, maintenance is light.
• AI agent. Building, AI usage per task, hosting and monitoring. On top of that, maintaining instructions and examples, and a new model now and then.

For all three, include the exceptions that stay with a person. An automation that passes many cases on to an employee delivers less than the business case promises. How to do that calculation is in building an AI business case. What an agent costs is on what does an AI agent cost.

Risks

• With RPA: silent failures after a screen change, a bot filling in the wrong field, and knowledge about the bot sitting with one person.
• With an AI agent: mistakes that sound convincing, hidden instructions in emails (prompt injection), personal data passing through an AI service and costs that rise when an agent gets stuck in a loop.
• With all three: too many permissions, no logging and nobody who owns it.

All three can be managed if you arrange it upfront: restricted permissions, logging, monitoring and a person who approves where it matters. More in using AI safely.

Process first, technology later. That's the first principle of our APO method, and it's exactly what this choice needs.

1. 01We map the process. Together with the people who do the work, in measurable steps. Whatever already runs, including an existing bot, simply goes on the map.
2. 02We give each step three numbers. How long does it take, with how many people, and how often a month? That tells you what's at stake.
3. 03We choose per step. Does it stay human work? Do we automate it with fixed rules, through a workflow or, if there's no other way, on the screen? Or does AI do it, with a person who checks?
4. 04We do the maths. What does it deliver, what does it cost to build, and when has it paid for itself?

In the method, choosing AI comes in two forms: AI assists (the person decides, AI does the groundwork) or human approves (AI does the work, a person checks). Fully AI is the exception. How that works is in human in the loop.

Whatever already runs stably and works well, we leave alone. We don't build an agent where a rule will do, or a bot where an integration is possible. We build workflows and AI agents in n8n ourselves, hosted in the EU.

This is how it goes:

• AI Quick Scan. One day on site with you. A business case and a roadmap within a week. € 6,000.
• AI Partnership. A dedicated team that builds, connects and keeps things running. From € 4,500 per month, for a minimum of twelve months. We only build within the partnership.

All amounts exclude VAT.