A workday with AI agents: how a Mittelstand company works with Claude Code and Cowork

Forget the futurist essays. This is what a perfectly normal Tuesday looks like in a Mittelstand company that works agentically, with Claude Code in engineering and Cowork in the business teams. Four scenes, all composed from real projects, none of them science fiction.

I (Sebastian) have been building workflows like these with Mittelstand companies for two years. The most surprising part is never the technology. It is how quickly the question shifts: from "what can the AI do?" to "what do I delegate today?".

08:40, engineering: the ticket is already running

The developer does not start her day with the ticket, but with the review of the ticket the agent prepared overnight. Claude Code has narrowed down the bug, written a fix and run the tests. She checks the plan, corrects one assumption about the database migration and approves. The routine is delegated, the decision stays with her. While the fix runs through, she puts two more agents to work in parallel: one writes tests for a legacy module, one documents the payment flow.

The point of the scene: the bottleneck is no longer typing speed, it is the quality of briefs and reviews.

09:30, marketing: 20 files become one report

The marketing manager drags the campaign folder into Cowork and describes the goal: a monthly overview as a table plus a one-page management summary, tone factual. The agent reads the exports, builds the table, flags two outliers and saves both, cleanly named, into the folder. The manager checks the outliers (one is real, one is a tracking artefact) and sends the report.

The point of the scene: this is not chat, this is delegation. Briefing, checking, approving are the new habits, and they are learnable.

11:00, controlling: the reconciliation nobody misses

The controller has Cowork reconcile the month's incoming invoices against the order list. Three discrepancies, neatly listed with sources. This used to be half a day of Excel digging; now it is half an hour of checking the three cases. The decision about what happens with the discrepancies is hers, not the agent's. For anything touching payment approval or personnel, the rule stays absolute: a human gate, no exceptions. Why that must remain is in the human-in-the-loop guide.

15:00, management: the decision memo

The managing director needs a memo on three vendor offers for Thursday. Cowork reads the PDFs, builds the comparison table, lists open questions. He sharpens two criteria and has the table re-sorted. The memo is done in an hour instead of Friday night.

What really changes (and what does not)

What changes: work shifts from doing to commissioning and checking. Whoever briefs well is suddenly more productive than whoever types fast. Routine work loses its share of the day, judgement gains.

What stays: responsibility, review and the ground rules. An agent is like a very fast new employee in their first week: astonishingly capable, but nothing goes out unchecked. And data policy stays a leadership topic: company plans (Claude for Work, API) instead of private accounts, because there inputs are not used for model training by default (as of June 2026).

It is not self-running: the gap between "tools available" and "tools used" is real and well documented, see our analysis of the employee usage gap. The difference between the four scenes above and an expensive licence graveyard is training plus ground rules, not the tool.

How to start without overreaching

Three steps that have proven themselves:

1. One pilot team per world. A dev team with Claude Code, a business team with Cowork. Small, supervised, with real tasks.

2. Ground rules before scaling. What may be delegated, what needs approval, what is off-limits. Plus the tier and data-policy decision.

3. Training as a programme, not an event. Briefing, checking, approving is a competence. It is not there one workshop day later; after eight supervised weeks it is. The entry point is in our courses, our way of working under how we work.

FAQ

Is this only for tech companies? No. Three of the four scenes above are business teams. The prerequisite is not tech affinity but clean task descriptions and review discipline.

How long until this workday is reality? Pilot teams work like this after four to eight weeks. The broad organisation takes longer, because ground rules, training and trust have to grow.

What about agent mistakes? They happen, which is why review steps and approvals are built in. The honest benchmark is not perfection, it is the error rate of the previous manual work.

Claude Code or Cowork, which first? Both in parallel, one pilot team each. They reinforce each other: the principle is identical, the audiences complement each other.

Read more

• What is Claude Code? The AI agent in the terminal, simply explained
• Claude Cowork explained: the AI agent for everyone who does not code
• What is agentic AI? The executive crash course
• Human-in-the-loop: how much autonomy should an agent have
• First AI agent: from use case to production
• The McKinsey gap: employees use more AI than bosses think
• Rolling out Claude Code in the team: training, configuration, governance

Sources

• Anthropic, product pages for Claude Code and Claude Cowork, claude.com / anthropic.com (as of June 2026)
• Sentient Dynamics workshop aggregate (DACH Mittelstand clients, 2025-2026); scenes are composed from real projects and anonymised

How Sentient Dynamics can help

We build exactly this workday with you: pilot teams for Claude Code and Cowork, ground rules and data policy, training as a programme and a measurement system that proves the effect instead of claiming it.

Book a demo