Martin Weitzel

But what should I actually do with AI agents?

AIAI AgentsAutomationWorkflowsEnterprise

“But what should I actually do with AI agents?”

I get this question every time I post about OpenClaw, Claude, or Codex. Because it keeps coming up, it deserves a clear answer.

AI agents automate workflows: (chains of) tasks that knowledge workers have previously done manually in the company.

The principle isn’t new. Companies have been automating since the dawn of digitization. What’s new is where the boundary lies.

Previously, automation was mostly for highly standardized processes: always the exact same sequence, a finite number of case types, high volumes.

AI agents are shifting this boundary:

  • You no longer need thousands of process transactions for automation to be worthwhile. Even a few incidents a month are enough.
  • The sequence no longer needs to be completely strictly defined. Agents decide autonomously which step makes sense next (without pre-conceived if-then-else chains).

I see two variations for this:

Variant 1: Fixed Workflows with AI Action Nodes (e.g., via n8n)

You define processes (visually). Some actions are handled via classical software code, others are AI-supported by calling LLM interfaces.

  • Pros: Easily documentable, versionable, and controllable.
  • Cons: You have to define the processes in detail. If a process run doesn’t fit the mold, the automation fails.

This is the standard in companies today.

Variant 2: Autonomous Agents (e.g., OpenClaw, Claude, Codex)

Agents decide autonomously how to achieve outcomes. To do this, they either work with APIs/MCP or directly on the PC, operating browsers, calling tools, literally seeing “what the human sees.” Approaches can be taught to the AI via Skills (Markdown files) in natural language.

Examples: Mapping tender texts against a portfolio or master data maintenance with crosschecks in different systems (even where there isn’t an API at all).

  • Pros: This is genuine delegating, like an assistant at the desk.
  • Cons: Harder to control. Prompt injection can manipulate them, decision paths are harder to trace.

To be fair: Humans have similar problems (phishing, lack of auditability).

I have experimented a lot myself with OpenClaw, for example: In theory, it works exactly how you imagine “intelligent agents.” In practice, I was often busier fixing faulty runs than extracting value. This is still in its infancy.

Important point (regardless of the variant): Before you even think about agents, you need clarity about the process: How are decisions made? Which systems are looked into? How do you spot problems? If all of this is simply in someone’s head or a gut feeling, no agent in the world will help. In that case, you should first revisit and document the processes.

My current recommendation for corporate workflows: Start with Variant 1. Clearly define workflows and only make individual action nodes AI-supported. Autonomous agents come after that, once data, roles, guardrails, and monitoring are well established. And once the agent harnesses are truly enterprise-ready.