Process automation
RPA or AI agent
Two words that often get mixed up. RPA is a robot that carries out fixed steps. An AI agent is a model that decides itself which steps are needed. The difference determines what it costs, what can go wrong and who is responsible. This is how you choose.
Fixed steps or own choice
Anthropic, one of the makers of large language models, draws the line sharply. A workflow is a system where a model and tools run along predefined paths. An agent is a system where the model directs its own process and tool use. RPA belongs to the first group, without a model: all steps are fixed.
When RPA is enough
If the work is the same every time and the rules fit on paper, a robot is the cheapest and most reliable. Lacity and Willcocks showed at Telefónica O2 that this kind of process, with high volumes and fixed rules, gives the largest gain. A robot does what you say, and nothing else. In administration that is an advantage.
When you need an agent
When it is not fixed in advance how many steps are needed, or when every case is different. An email with a question that first has to be understood. A file where three sources are weighed against each other. Anthropic advises agents only for open problems where the steps cannot be predicted, and otherwise the simplest system that works.
What the difference costs
- A robot is predictable and cheap per transaction, but breaks at a screen change.
- An agent can handle exceptions, but costs more per case and can make a mistake that sounds convincing.
- So an agent always comes with someone who checks, certainly at the start. A robot comes with someone who handles the exceptions.
Often both
In practice the work sits in between. An agent reads the email and determines what it is, a fixed flow processes it afterwards. That way you get the model's understanding where needed, and the predictability of fixed rules where possible. That is also how our AI employees are built: the model prepares, a person approves, and fixed rules execute.
Frequently asked questions
Is an AI agent not simply better?
Not for work that is fixed. An agent makes choices, and every choice is a chance of a mistake. For fixed rules a system that makes no choices is better.
Can an agent operate my systems?
Yes, through integrations or through the screen. But give it only the permissions the task needs, and do not let it send or post anything without a person having seen it, certainly at the start.
What is intelligent process automation?
The combination: fixed rules with a model added for the steps that need judgement, such as reading a document. IEEE fixed the terms in a standard in 2017, because vendors called everything RPA.
Sources
Where every figure on this page comes from.
- Anthropic: Building effective agents · Het verschil tussen workflows en agents, en wanneer je welke kiest
- Lacity en Willcocks: Robotic Process Automation at Telefónica O2 · MIS Quarterly Executive, maart 2016
- IEEE 2755-2017: Guide for Terms and Concepts in Intelligent Process Automation · Gepubliceerd 28 september 2017
Our approach
From advice to management, with one partner
The same rhythm for every project.
Advice
We start with a conversation, not with code. First clarity on what you need and what you do not.
Build
We build and integrate to measure, with technology that fits you. You own your data and your code.
Manage
We stay involved: monitoring, adjusting and growing with your business.
Not sure whether you need a robot or an agent?
Describe the work in a thirty-minute conversation, and we tell you what fits and why.
Call 085 083 5775A 30-minute intake