Artificial intelligence has moved from experiment to everyday tool in many companies. Two current reports put the same question centre stage: do AI tools actually solve existing problems – or do they create new ones elsewhere?
In the t3n podcast MeisterPrompter, the interim conclusion is partly sobering. It cites hallucinations, i.e. invented but plausibly worded answers from language models, the high resource requirements of such systems, and productivity gains that cannot be clearly quantified. This is not a rejection of AI, but an indication that the benefits vary greatly depending on the use case.
At the same time, a case from media production: an AI agent is said to have deleted the raw footage of an entire season of the series "The Challenge Sverige". The broadcast of the second season was planned for November; now the footage has to be edited again. The wording in the reporting is cautious – the course of events has not been confirmed in all details. Nevertheless, the incident serves as an instructive example.
Autonomy is not an end in itself
An AI agent is a system that not only generates text but independently executes actions: moving files, starting scripts, calling interfaces. This is precisely where the risk shifts. A badly worded paragraph in a generated text is annoying and will be caught in editing. An irreversible action on a production system is something else entirely.
The decisive difference lies in reversibility. As long as an action can be undone, a mistake is a loss of time. As soon as it cannot, the mistake becomes damage. This distinction should be at the start of any consideration of how much autonomy an AI tool is given in one's own operation.
Hallucinations are a process problem
That language models invent content is technically known and has not yet been fully solved. For day-to-day project work, this means: every point at which a model supplies facts needs a review instance. In development, this affects generated code that compiles but contains wrong assumptions about an interface. In editorial work, it affects product details, figures or legal notices.
In practice, this does not mean manually rebuilding every AI output twice. It means moving checks to where they can run automatically: tests, linting, schema validation, the four-eyes principle for publications. Where no such check exists, using AI at that point is simply riskier than it appears at first glance.
Productivity must become measurable
The reference to unclear productivity gains deserves particular attention. AI tools quickly feel fast because the first output arrives within seconds. Whether the entire process – including correction, coordination and rework – is actually shorter only becomes clear from an honest look at the complete workflow.
Simple but consistently collected metrics are helpful for this:
- Throughput time of a typical process before and after introducing the tool
- Share of AI results that can be adopted without rework
- Errors that are only noticed late in the process
- Effort for operation, licences and training
Without such figures, the assessment remains a matter of taste – and thus susceptible to expectations set from outside.
Backups as a basic requirement, not an option
The case from series production is a reminder of something self-evident that is often not fully implemented in practice: a backup is only a backup if it is located separately from the production system and restoration has been regularly rehearsed. As soon as an automated system has write access to data, this applies doubly.
Graduated permissions are just as important. An agent that is supposed to analyse content generally only needs read access. An agent that is supposed to manage files does not need access to the archive. This separation costs time during setup and saves a project in an emergency.
A pragmatic framework
A simple sequence has proven effective for getting started:
- Name the use case and describe the problem that is to be solved – do not think from the tool outwards.
- Check whether the results can be verified and who bears the responsibility.
- Limit access rights to the necessary minimum.
- Define a fallback: backup, versioning, manual route.
- After a few weeks, measure whether the expected effect occurs.
This framework does not prevent mistakes, but it limits their reach. And it makes the discussion about AI in the company more objective, because it is guided by concrete workflows rather than general promises.
Classification for web and software projects
For web and software projects, this means: AI belongs where results can be verified and errors can be corrected – for example with drafts, research or code suggestions followed by testing. The more far-reaching and irreversible an automated action is, the more strictly rights, approvals and backups must be regulated.
Whoever defines these guardrails before the rollout uses the benefits without increasing the project risk in an uncontrolled way.