Solution
Business agents
An agent isn’t an assistant you talk to: it’s a program that completes a task, triggered by an event rather than by a question. The hard part isn’t making it work once, it’s making it dependable every day.
The gains from the agents
The work happens overnight
Event-triggered work no longer waits on somebody being available. By Monday morning it’s already done.
The guardrails are written, not hoped for
What the agent may do is declared. It cannot send an external email or change an accounting entry on its own judgement.
Prompts are code
Versioned, reviewed and tested like the rest. A wording change that breaks a case shows up before production rather than after.
In practice
- Event triggers: mail received, file dropped, deadline reached, entry created
- Multi-step chains with error recovery, rather than a script that stops at the first incident
- Human checkpoints placed where a mistake is expensive, and nowhere else
- A declared scope of action: the permitted tools, the amount limits, the possible recipients
- Prompts and rules versioned in the repository, with test cases that run on every change
- A full record of what was done, readable by someone who was not there
Systems involved
- Mail, calendars and shared folders
- ERP, CRM and accounting
- Code repositories and integration pipelines
- Internal databases
- External services over APIs, with usage limits respected
Service lineArtificial intelligence →
The chained processes
The same work, against each sector’s own constraints. Every card opens the full sector.
Banking and insurance
Processing chains with a checkpoint
The repetitive steps of a case chained at portfolio volume, with a handler approving at the points where the decision commits and throughput measured.
Energy and utilities
Chaining operational tasks
Readings, checks and notifications triggered by the state of the system, with an operator deciding on deviations.
Healthcare and life sciences
Chaining administrative tasks
File preparation, reminders and transmissions chained together, with human approval before anything concerning a patient goes out.
Logistics and supply chain
Partner integration
One connector per exchange format, tested and monitored, rather than one unmaintained script per partner.
Manufacturing
Chains triggered by shop-floor state
Parts ordering, maintenance orders and documentation updated from the same signals, with a supervisor approving.
How it runs
One task
One precisely described task, with its edge cases and what must stay human. A vague scope is the main cause of failure.
Supervised
The agent proposes and doesn’t execute. Its proposals are compared to the real work until the gap becomes boring.
Measured autonomy
The safest cases move to direct execution, the rest keep coming back. The boundary moves with the numbers.
Operation
Alerts, log, cost per run and a stop procedure. An agent you can’t stop cleanly isn’t finished.
To go deeper
The EU AI Act now applies: four questions for a Swiss company
A Swiss address does not settle whether the EU AI Act applies. Start with the company’s role, the EU connection, the effect on people and the evidence needed before production.
Agent memory: one word, several mechanisms
Karpathy's LLM Wiki pattern, Anthropic's memory tool, context editing and managed stores do different things that the word memory conflates. What to write down, what never to, and what changes under Swiss law.
An AI agent in practice: its work, and where it stops
Bank reconciliation in an accounting practice, incoming mail in a municipality, quotations in an SME, research monitoring in an institute: realistic AI agents, with their guardrails.
Test the fit: Business agents
Describe the context, constraints and decision you need to make. The first conversation qualifies scope, boundaries and the next useful step.
Describe the situation