Automation was long regarded as a developer topic: anyone who wanted to digitalise a recurring process needed a ticket, a budget and a place on the roadmap. A current example from the t3n editorial team points to a shift. There, a team created an AI agent in everyday language – that is, without classic programming, solely via descriptions of what should happen. The result: today the agent creates more than 80 percent of the advertising orders.
The report does not provide more details. Nevertheless, the case is revealing because it hits a type of process that practically every mid-sized company knows: structured, rule-based data entries that take a lot of time and require little specialist decision-making.
What is meant by vibe coding
Vibe coding refers to the creation of working software through natural-language instructions to an AI model, instead of by writing code. The person describes the desired behaviour, checks the result and refines the description until the flow is right.
An AI agent, in turn, is more than a chatbot: it executes work steps – reads inputs, applies rules, writes data into a system. The combination of both is the real lever. Not the model that formulates texts, but the model that processes a defined process repeatably.
Why order creation is a good entry-level case
The fact that the creation of advertising orders in particular was automated fits a pattern found in many companies. Such processes typically have characteristics that favour automation:
- They run frequently and according to largely identical rules.
- The input data is formulated in an unstructured way, but is limited in content.
- The target system expects a fixed data format.
- Errors are noticed quickly and can be corrected.
The last point is decisive. An agent that takes over 80 percent of the cases does not replace the human – it relieves them. The remaining cases are precisely those that require experience and follow-up questions. This division of labour is more realistic than the claim of handing over a process completely.
The specialist department as author of the process
The remarkable part of the example is not the technology, but the distribution of roles. The agent was built by a team that executes the process daily – not by an external development department.
This solves a classic problem: in automation, a lot of time is usually spent translating implicit specialist knowledge into a specification. Those who describe the process themselves skip this translation step. Special cases, exceptions and unwritten rules end up directly in the description because the person knows them.
At the same time, a new demand arises for specialist departments. Anyone who builds an agent must be able to think through their own process cleanly: What inputs are there? What decision follows from them? When is a case not automatable? This clarification is process work, not IT work – and it pays off even if no agent is created in the end.
Where the limits lie
No general all-clear can be derived from a single successful case. Several points remain to be clarified regardless of the tool:
- Data access: An agent that creates orders needs write permissions in a productive system. Permissions, logging and rollbacks are part of this from the outset.
- Quality control: A rate of over 80 percent says nothing about how the remaining cases are recognised. A reliable criterion is needed for when a human takes over.
- Operation and maintenance: Processes change, and so do interfaces. Whoever built the agent must also be able to maintain it – or someone else must be able to take it over.
- Documentation: Natural-language instructions are readable, but not automatically traceable. Without versioning, the tool quickly becomes a black box.
- Data protection: Depending on the content processed, personal data is involved. Which models run where is a governance question.
From individual case to pattern
Anyone who wants to transfer the example is best advised to look not for the largest but for the clearest process. Flows that a team already handles with an internal checklist are well suited – because this checklist is already half the description of the agent.
It also makes sense to calibrate expectations early. An agent that reliably handles part of the cases and hands over the rest cleanly is productive. An agent that handles all cases half-well creates rework and quickly loses the trust of users.
For web and software projects, this shifts the division of tasks: specialist departments can increasingly automate simple, clearly defined flows themselves, while the technical side remains responsible for interfaces, permissions, testability and operation. Anyone operating systems such as TYPO3 or connected specialist applications should therefore check whether clean interfaces and traceable permissions are in place – they determine whether such agents can connect at all. The greatest gain here lies less in the tool than in the discipline of describing one's own process precisely beforehand.