RPA, Intelligent Automation or AI Agent: Which Automation for Which Process

RPA automates fixed rules on structured data, intelligent automation adds AI for documents and classification, and an AI agent plans its own route to the result. Three process characteristics decide which level fits: the stability of the rules, the structure of the inputs and how open the solution path is. This article provides the criteria, a calculation method and a checklist.
What is the difference between RPA and AI?
RPA follows predefined rules, while AI makes assessments based on learned patterns. Robotic process automation is software that imitates a person's actions on the user interface of existing applications: reading fields, transferring values, filing documents. The sequence is fully specified in advance. If an amount sits in the expected field and falls below the threshold, the bot posts it to the designated account. It does not recognize deviations. It stops or moves the case to an exception queue.
Information systems research describes RPA as a solution for high-volume, clearly rule-based tasks that bridge systems without an interface. A bot is quick to set up and fully traceable. When a screen layout or a file format changes, however, it stops, so large RPA estates carry maintenance effort that the original business case often omits.
AI, by contrast, reads an invoice whose layout it has never seen or recognizes a cancellation in a customer email, with an accuracy below 100 percent. AI replaces human judgment in narrowly defined steps; RPA replaces human keystrokes.
What is intelligent automation or hyperautomation?
Intelligent automation combines rule-based automation with AI components in an orchestrated workflow. A typical chain: an AI model reads incoming documents, a rule checks plausibility, a bot transfers the values into the ERP system, and a workflow routes uncertain cases to a person. The standards body IEEE published a guide with shared terminology for this field as early as 2017 and uses the term intelligent process automation.
Hyperautomation describes the ambition to automate as many business processes as possible with a coordinated toolkit, from process mining through RPA to AI and low-code. It is a program more than a technology, and it needs orchestration that defines which component handles which step and where a person decides. Measuring the process first with process mining reveals variants and exception rates before any technology is chosen.
How does an AI agent differ from RPA?
An AI agent receives a goal and determines the route itself, whereas an RPA bot executes a predefined route. An agent built on a language model breaks the task into steps, selects tools such as search or database queries, evaluates interim results and corrects its plan. This allows it to handle cases whose sequence varies from one case to the next: a supplier inquiry with follow-up questions, a complaint with incomplete details, research across several systems.
The price of this flexibility: an agent can solve the same case differently on two occasions, so repeatability has to be created through tests and approval levels. Agentic AI and RPA therefore differ in planning (open versus fixed), in inputs (language and context versus fields) and in assurance effort (monitoring of outcomes versus functional testing of the sequence).
Reliability determines the degree of autonomy
Research shows clearly where agents stand. A benchmark from Carnegie Mellon University simulates a small software company with realistic office tasks. The best system tested completed around 30 percent of the tasks fully autonomously. A second benchmark tests agents in customer interactions for an online retailer and an airline. Leading models solved fewer than half of the tasks. The probability of solving the same task correctly in each of eight runs was below 25 percent in the retail scenario.
The rule that follows: the more expensive an error, the tighter the guardrails, with agents preparing decisions and a person approving them. Only when the measured error rate in your own process has stayed below the agreed threshold for several weeks do you extend autonomy. The levels between approving every case and pure oversight are covered in the article on human in the loop.
Three process characteristics determine the right level
The choice of automation follows from the process. Three characteristics are enough for a robust first allocation.
Stability of the rules. If every step can be written down as an if-then rule, and those rules change less than twice a year, RPA or a conventional interface is the first choice. If a step requires the kind of judgment experienced staff make from experience, it needs AI.
Structure of the inputs. If the data comes from fields or fixed forms, RPA is sufficient. If it arrives as PDF, scan or free text, the process needs an AI component for recognition. This is the classic entry point to intelligent automation, described in detail in the article on document processing with AI.
Openness of the solution path. If the sequence is fixed in advance, an agent is oversized. If the order of steps depends on the content of the case, with follow-up questions and changing systems, the agent is the right level.
Invoice intake shows why the answer is usually a combination: recognition (AI), plausibility checks (rules), posting (RPA or interface) and resolving disputed cases (person or agent).
When does each type of automation pay off?
Automation pays off when the annual net benefit covers the one-off costs within 12 to 24 months and error costs remain contained. The calculation is the same for all three levels; only the weights differ.
A worked example with round assumptions shows the logic. Invoice intake with 40,000 cases a year, six minutes per case and a fully loaded cost of 60 euros per hour ties up 240,000 euros a year. If an intelligent automation solution automates 70 percent of cases, the gross benefit is 168,000 euros. If a person reviews every fifth automated case in one minute, that costs 5,600 euros. With running costs of 40,000 euros, a net benefit of just over 120,000 euros remains before error costs.
An agent achieves higher rates on open cases than RPA, but requires more review and produces errors that are harder to detect. Calculate each level with measured values from a prototype, never with vendor claims. How to set out benefits and costs robustly is covered in the article on the AI business case.
Checklist: which level fits your process?
Answer the following questions for each process step. The first “no” determines the minimum level.
- Can the step be fully written down as a rule that changes less than twice a year? Then RPA or an interface is sufficient.
- Are the inputs available as fields or tables? If not, the step needs intelligent automation.
- Is the sequence of work steps fixed before the case begins? If not, an agent is a candidate.
- Is the cost of an error lower than the cost of a human review? If not, a person remains in the approval loop.
- Is a system interface available? If so, prefer it to a bot on the user interface, because it survives screen changes.
- Has an operational owner been named who handles exceptions? If not, the process is not ready for any level.
Combination is the norm in operations
Mature automation estates consist of layers. Interfaces move data between systems, bots work where legacy systems offer no interface, AI components recognize documents, and agents take on the open cases, calling bots as tools. An orchestration layer decides which case takes which route and logs every decision.
Three design rules follow. First, every agent receives only the tools and permissions its task requires. Second, stable bots stay in service and become tools of the overarching orchestration. Third, every layer has a business owner who knows its exception rate and error rate.
For agents, an approval rule is added: which actions may the agent carry out itself, and which may it only propose? The overall route from prototype to operations is set out in the article From pilot to scale.
How Niehues Consulting supports the choice
Niehues Consulting combines AI management with delivery management and supports organizations from process analysis through to stable operations. The foundation is the Niehues Delivery Model, with three phases and fixed review questions. Three of them matter most when choosing the level of automation:
- What is the baseline, and how will we measure success? Without measured case volumes and handling times, every business case remains an estimate.
- Does the prototype prove its value in the real process with real data? Only a test with real cases shows whether RPA is enough or an AI component is required.
- Who owns operations from day one? Bots and agents need a business owner who handles exceptions and tracks the metrics.
The result is an allocation per process step, a business case based on real values and an implementation with named roles. Marin Niehues has led initiatives ranging from VC-backed scale-ups to DAX 40 corporations in 20+ countries and brings this delivery perspective to every engagement. The starting point is a free 30-minute initial consultation via the contact form.
Frequently asked questions
What is RPA in simple terms?
RPA is software that carries out repetitive actions on the user interface of existing programs in the same way a person would. The bot reads fields and transfers values according to fixed rules. It suits high-volume processes with structured data and stable rules.
Is RPA a form of artificial intelligence?
RPA on its own is not artificial intelligence, because it executes predefined rules without learning or making assessments. It becomes part of an AI solution when it is combined with components such as document recognition or language models. That combination is called intelligent automation.
Will AI agents replace RPA?
AI agents replace RPA only where the solution path is open. For stable, rule-based processes, RPA remains cheaper and more reliable because it treats every case identically. In mature estates, agents call bots as tools.
What is the difference between agentic AI and RPA?
Agentic AI plans its own route to a goal, while RPA executes a predefined route. Agents process language and context and correct their plan, whereas RPA works with fields and fixed rules. In return, RPA is fully repeatable, while agents must be safeguarded through tests and approvals.
How long does it take to implement automation?
A single RPA process can be in production within a few weeks, and an intelligent automation solution with document recognition within two to four months. Agents additionally need a phase of tightly limited autonomy in which error rates are measured. Data access and exception handling take up the largest share of the time.
Who should own automation in an organization?
Business ownership lies with the process owner in the relevant function, and technical ownership with a central automation or AI unit. The process owner knows the exceptions and metrics, while the central unit ensures standards and security. Without a named process owner, every bot remains a project without operations.
How do you measure the success of automation?
Four metrics against the baseline show success: automation rate, lead time, error rate after review and cost per case. The exception rate, meaning the share of cases returned to people, adds a further signal. If it rises over several weeks, the process is unstable.
Let’s talk about your initiative.
In 30 minutes we clarify where your organization stands and which step will pay off most now.
Request a consultation30 minutes · free · no obligation



