AI Agents for Business: When They Make Sense—and When a Workflow Is Enough

An agent can go further and, for example:

• retrieve information from multiple data sources
• call APIs or other software
• evaluate intermediate results
• choose its next step based on the result so far
• pursue a task across multiple steps
• request human approval when needed

Autonomy does not mean the agent acts without limits.

Goals, available tools, permissions, and boundaries are still defined by people.

That is an important distinction from the common market narrative of a fully autonomous AI system.

What does agentic AI mean?

The terms AI agent, AI Agent, and agentic AI are not used consistently across the market.

For this article, we therefore use a practical distinction based on Anthropic’s framework:

Agentic systems is the umbrella term.

Within that, we distinguish between:

• workflows with predefined execution paths
• agents that dynamically determine their path to a solution within a defined framework

For businesses, that distinction is more useful than debating terminology.

Anthropic distinguishes precisely between these two models:

In a workflow, models and tools are orchestrated through predefined paths.

With an agent, the model dynamically decides which steps and tools it needs for the task.

That does not mean an agent is technically superior.

Quite the opposite.

Agents often require more model calls, more tool calls, and more control mechanisms. That can increase cost and runtime. At the same time, errors can propagate across multiple steps.

If a fixed workflow can perform the same task reliably, additional autonomy is often unnecessary.

A good example is combining multiple forms of automation.

Not every request needs to go directly to a powerful agent.

An upstream workflow can first identify and handle simple cases. Only complex or ambiguous cases are then passed to an agent.

That means AI is used where its flexibility is actually needed.

We would therefore answer at least six questions:

• Does the task occur often enough?
• Is the desired outcome clearly defined?
• Are the required data and systems accessible?
• Can good and bad outcomes be evaluated?
• Is it clear which actions the agent is allowed to perform?
• Do we know what should happen when something goes wrong?

The harder these questions are to answer, the harder it becomes to build reliable automation.

A good candidateThe process occurs regularly, requires context or unstructured information, and still has a clearly definable goal.

A poor candidateThe process changes constantly, has no measurable outcome, or depends on decisions whose rules are unclear even within the company.

A warning signNo one can explain how the process actually works today. In that case, clarify the process first before asking AI to take it over.

Blanket claims such as “An AI agent costs €2,500” or “AI saves companies 70 percent” are therefore not very useful without specific assumptions.

It is more useful to break the cost down into five layers.

One frequently underestimated factor is ongoing model and tool usage.

An agent can perform multiple intermediate steps for a single task.

Each additional step can generate more inputs, outputs, or tool calls.

More autonomy can therefore mean not only more capabilities, but also higher ongoing usage costs.

Operations matter even more.

Monitoring, evaluation, and error handling do not disappear after launch.

Especially for business-critical processes, the cost calculation should include not only development but also ongoing oversight.

A risk-based approach makes more sense.

The NIST AI Risk Management Framework considers risk as a combination of likelihood and potential impact.

For companies, that leads to a practical question: “What happens if the agent is wrong?”

The greater the potential harm, the stronger the controls and approval requirements should be.

Human approval is particularly useful when:

• an action has external consequences
• money is moved
• contracts or binding offers are involved
• personal or sensitive data is processed
• a decision cannot easily be reversed
• the case falls outside tested scenarios
• the agent iterates repeatedly without making progress
• the result later needs a traceable explanation

When the cost of errors is low, more can be automated.

Data protection and the EU AI Act

For German companies, data protection and regulation should be considered from the design stage.

Before implementation, clarify at least:

• Which personal data does the system process?
• Which data leaves the company’s own infrastructure?
• Which external model providers and platforms receive access?
• Which tools is the agent allowed to call?
• Which actions are logged?
• Who can change permissions?
• Which decisions require human review?

The EU AI Act does not require human oversight as a blanket rule for every business AI agent.

Article 14 applies to high-risk AI systems.

Whether a specific use case falls into that category needs to be assessed on a case-by-case basis, with appropriate specialist and, where necessary, legal review.

This article does not constitute legal advice.

An off-the-shelf platform can make sense when standard integrations and existing agent features already cover the process well.

A custom implementation becomes more attractive when proprietary systems, specific permissions, or unique process logic need to be integrated.

The decision is rarely purely technical.

What matters is which parts of the system are strategically important to the company.

That includes, for example:

• process logic
• prompts and instructions
• tool definitions
• data sources
• permissions
• evaluations
• logs

A sensible architecture should therefore consider before implementation which parts may need to be replaced or extended later.

We would be more likely to advise against an agent when:

• a simple workflow solves the same task reliably
• the decision is entirely rule-based
• only a few cases occur each year
• errors could have very serious or irreversible consequences
• the data foundation is unclear or unreliable
• there is no way to evaluate outcomes
• no one is responsible for ongoing operations
• the process itself has not yet been defined reliably

Anthropic itself recommends starting with the simplest possible solutions and adding complexity only when it produces measurably better results.

That is a good principle for businesses as well.

For more complex initiatives, an AI agency can help evaluate the process, automation approach, integrations, and control mechanisms together before technical implementation begins.

We would start in five steps:

• 1. Select the process
• 2. Document the current process
• 3. Compare a workflow with an agent
• 4. Build a small pilot with clear boundaries
• 5. Evaluate the results before expanding

For the pilot, define in advance:

• What counts as a correct result?
• Which errors are acceptable?
• Which actions may the agent perform independently?
• When must a human take over?
• Which data and systems may the agent use?
• How is every relevant action logged?

Only once these fundamentals work should the agent’s scope of action be expanded.

That reduces the risk of building a technically impressive agent that fails to create reliable value in day-to-day business.