Process Automation with AI Agents: From Rigid Rules to Clear Goals

Why Does AI-Powered Process Automation Need New Models?

AI process automation refers to the automation of business processes using Artificial Intelligence (AI). Unlike rules-based automation, it also handles steps that cannot be fully defined by rules – for example, because requests are phrased freely or context needs to be interpreted.

A holiday request is quick to model: submit the request, check the duration, obtain approval, update the calendar. Then come the exceptions. Leave during school holidays requires additional approval, employees abroad are subject to different rules, and a request spanning the year-end affects two leave balances. Four steps become a rule set with dozens of branches.

More complex processes such as order-to-cash work in a similar way, but involve greater dependencies and higher costs. AI agents promise to handle variations more flexibly. This agent-based process automation – or Agentic Process Automation – independently pursues a defined goal through multiple steps. However, this only works if organisations describe processes differently for AI agents and establish the foundations they rely on.

Why Do Rule-Based Processes Become Difficult to Manage Over Time?

Rule-based processes become difficult to manage because every exception requires an additional rule. These workflows are often modelled using Business Process Model and Notation (BPMN), a standardised notation that enables organisations to represent business processes graphically. Its strength lies in traceability: every step is documented and every decision can be reviewed.

Over time, however, typical limitations emerge. Every changed rule requires the model to be updated and retested. Exceptions multiply the number of variants until hardly anyone retains an overview of all process paths. If a case enters a variant not covered by the model, the process stops and requires manual intervention to continue.

In addition, a process knows its rules, but not their purpose. The requirement to approve leave of five days or more aims to ensure that a team remains operational. However, the process requires approval even when the team is in a quiet period. What a person could assess in seconds, the model must represent through an additional rule.

The effort becomes most apparent during system migrations. Many companies have developed bespoke process logic in SAP ECC 6.0 over many years. Standard maintenance ends at the end of 2027, with a paid extension available until the end of 2030. When moving to SAP S/4HANA, organisations often cannot transfer established bespoke logic unchanged and must rebuild it with significant effort.

This does not argue against rule-based automation. For stable processes with clear rules, it remains the more cost-effective and reliable solution. Processes become challenging when their rule set consists primarily of exceptions. This is where AI-based process automation is worth considering.

What Changes When AI Agents Manage Processes Through Goals Rather Than Rules?

A goal-based process is a workflow in which organisations define the desired outcome, the available context and the boundaries, while an AI agent determines the individual steps within these boundaries. Rather than modelling every path in advance, organisations describe what they want to achieve in the end. Goal-based processes therefore form the core of Agentic Process Automation.

Using a leave request as an example, the objective might be: “Ensure that the leave request is appropriately reviewed, approved and recorded.” The agent then independently checks the remaining leave entitlement in the HR system, compares the requested period with project deadlines, obtains approval from the line manager or contacts the designated cover if they are absent, and documents its decision.

This shifts the central question in process design. It is no longer “How do I make the process reliable through rules?” but “What context does the agent need to achieve the goal reliably?” A robust goal description includes several components:

• Objective and success criterion: How can you tell that the process has been completed correctly?
• Context and data sources: What information does the agent need, and which systems hold it?
• Tools: Which actions can the agent perform, such as API calls or existing automations?
• Boundaries and handovers: What is the agent not authorised to decide, and when does a person take over?
• Documentation: How does the agent record why it made a particular decision?

Fixed rules remain in place. Anything that can be clearly defined, such as statutory holiday entitlement or an approval threshold, remains stored in the system as a fixed rule. The agent handles the elements where the process cannot be determined in advance.

A realistic expectation is essential: AI agents based on language models do not learn independently during operation. If an agent makes an incorrect decision, people adjust the objective, instructions or context and test the change before it takes effect.

Why Does AI Process Automation Often Remain Stuck in the Pilot Stage?

Many AI pilots for process automation stall because the transition from a test environment to live operations fails due to processes, integration and data. According to the MIT NANDA initiative’s report, “The GenAI Divide: State of AI in Business 2025”, 95% of the organisations examined achieved no measurable financial return from their generative AI initiatives. The authors attribute this primarily to insufficient integration into business operations. The figure is based on interviews, surveys and an analysis of published projects, so it cannot be applied universally. However, the overall direction aligns with practical experience, where three main causes emerge:

• First, the pilot works, but the underlying process does not.
  An AI agent handles a clearly defined use case in the test environment. In production, it encounters the full range of variations, for which the process was never designed with an agent in mind.
• Secondly, systems are only connected in pairs.
  With these point-to-point integrations, there is no central layer through which an agent can reliably access data and actions. Each new connection becomes a standalone project.
• Thirdly, data is fragmented.
  Without context, the AI agent makes incorrect decisions or invents information. In an environment where it acts autonomously, the consequences are particularly significant.

This creates a clear sequence. Effective process automation with AI agents requires a digital core – CRM, ERP and cloud systems with standardised processes. This is supported by an integration layer comprising interfaces, middleware and identity management, followed by harmonised data and clear rules for data quality, security and access rights.

What Happens When Every Platform Brings Its Own AI Agents?

When every platform in a system landscape comes with its own AI agents, it creates siloed solutions that each cover one process step but not an end-to-end process. Almost every software provider now offers agents for its own platform. Companies use them in the CRM system, the ERP system and the office environment, often independently of one another.

The development echoes the uncontrolled proliferation of integrations seen in previous years – a growing number of interfaces that no one has a complete overview of. The consequences are similar for agents. They access different data sources and therefore each see only part of the context. Security, governance and monitoring are managed separately for each system. Dependence on individual providers increases.

As long as agents handle individual tasks, this may be sufficient. For a process that spans multiple systems, organisations need an overarching layer that enables them to manage the orchestration, governance, security and entire lifecycle of their agents – from development through to retirement. This requires answering key questions early on: Which agents exist within the organisation, who is responsible for them, what are they permitted to access, and how do they hand over tasks to one another?

What Does Agentic Process Automation Mean for Process Owners?

Process owners will define objectives and provide context rather than maintain rules. Their role will shift from modelling every branch to describing outcomes, success criteria and boundaries.

This includes testing an AI agent’s behaviour with real cases, particularly the edge cases around which the legacy rule set has evolved. It also involves clarifying jointly with IT which data and interfaces the agent requires. The transformation is less a technical than an organisational challenge. Teams that have worked with rules for years need to understand why a process is changing and what impact this will have. Change management must therefore form part of the planning from the outset.

Conclusion: Rules Where They Add Value and Goals Where They Are Missing

AI agents do not replace traditional process models. Clearly defined elements remain within the rules framework. Where variations, context and judgement determine the workflow, an agent takes over with a clear objective. The combination of both makes processes more flexible without compromising traceability.

The sequence is crucial. Process automation with AI agents needs connected systems, reliable data and shared governance before it delivers value in live operations. A good starting point is a process currently burdened by exceptions and supported by readily accessible data.

Would you like to identify which of your processes are suitable for goal-based workflows with AI agents? Salesfive experts support you with process analysis, integration and data architecture, as well as governance for agents across your system landscape. Get in touch!