A full inbox is the first agenda item in many companies — and often the least productive one. A current report describes an obvious counter-model: the prompt engineer Susanne Renate Schneider has set up an AI agent that pre-sorts her emails every morning. She shares her instructions for this in a podcast format.

What is interesting is less the individual case than the pattern behind it. In this context, an agent is nothing other than a language model that not only responds once, but repeatedly performs a defined task — here: reviewing incoming messages, classifying them and putting them in an order. The logic for this is not contained in a classically programmed rule set, but in an instruction in natural language.

Why this is relevant for companies

Many office tasks have exactly this profile: they are recurring, rule-based enough to be describable, but too fuzzy for rigid filter rules. Whether a request is urgent, whether it belongs in sales or in support, whether it needs a response at all — such decisions depend on the content, not on the sender or subject line.

Until now, automating such tasks was a project: clarify requirements, develop, test, roll out. Prompt-driven approaches shift this threshold. Anyone who can describe their own way of working precisely arrives at a first working result without writing a single line of code. This makes such solutions a sensible entry point for organisations that want to try out AI-supported process optimisation on a small scale first.

The real work lies in the description

The catch is the same as with any automation undertaking: a task can only be delegated if it is clear how it is to be done. Anyone who wants to tell an agent what is important must first decide themselves what is important. This clarification is precisely the part that many teams have never explicitly undertaken — it resides in the minds of individual people.

In practice, this means: before the prompt comes a brief stocktake. Which categories actually exist? By what criteria is prioritisation done? What happens with borderline cases? Answering these questions has value even if no agent is created in the end.

Thinking through limits

An agent that reads an inbox gains access to partly sensitive content — customer data, HR topics, contract drafts. For private use, this is a personal trade-off; in a corporate context, it is a question of data protection, confidentiality and internal policy. Before such a setup grows beyond the individual workstation, the following points should be clarified:

  • Data flow: Which content leaves your own system, and where does it go?
  • Permissions: May the agent only read and sort, or also delete, forward, reply?
  • Traceability: Can it be reconstructed in case of doubt why a message was classified the way it was?
  • Fault tolerance: What does a misclassification cost — a minute of searching or a missed deadline?

The last question determines the sensible area of use. Language models work probabilistically, so they make assessments and not deterministic decisions. For pre-sorting, summarisation and prioritisation suggestions, this is uncritical as long as a human skims the results. For steps with external impact — automatic replies, deletions, forwarding to third parties — the human should retain approval.

A realistic entry point

Anyone who wants to test the principle in their own company is best served by a narrow scope. A single workstation, a clearly defined task, a manageable period. Afterwards, it can be checked whether the sorting stands up to one's own assessment and how much reworking remains necessary.

  1. Select a task that arises daily and whose criteria are describable.
  2. Record the criteria in writing before the first prompt is created.
  3. Run in parallel for two to three weeks and check the hits on a sample basis.
  4. Only then decide whether a broader rollout or a more stable technical implementation is worthwhile.

At the latest at rollout, the purely prompt-driven approach usually ends. What works as a personal aid requires, in multi-user operation, connection to existing systems, permission management, logging and a unit that intervenes in the event of misconduct.

What this means for projects

For web and software projects, this above all shifts the starting point: a prototype is often created today in hours instead of weeks, and by the specialist departments themselves. Project work then no longer begins with the question of whether an idea works, but with the question of how to bring it into the existing system landscape safely, traceably and maintainably. Anyone who plans this transition from the outset saves themselves the later rebuild of a grown isolated solution.

Sources