Français
Deutsch
From tool to agent: a paradigm shift
For decades, IT systems were passive executors: give them an instruction, they process it, they stop. Agentic AI reverses this model entirely. An AI agent doesn't simply execute a task — it perceives its environment, plans a sequence of actions, uses tools, self-corrects, and pursues a goal without human intervention at every step.
For a Business Analyst or IT Project Manager in Switzerland, this isn't an abstract evolution. It concretely redefines how entire processes — onboarding, reporting, data validation, client communications — can be orchestrated autonomously.
What exactly is an AI agent?
An AI agent is a system capable of perceiving (reading emails, Jira tickets, APIs), reasoning (decomposing a complex goal into sub-tasks), acting (calling tools, writing code, sending messages), memorising (retaining context between steps) and self-evaluating (verifying its own results and retrying).
Frameworks like LangGraph, AutoGen, CrewAI and the native capabilities of Claude 3.5 Sonnet already allow deploying such agents in production. The ecosystem is evolving at unprecedented speed.
The three levels of agentivity
Level 1 — Reactive agent
The agent responds to a trigger (an incoming email, a created ticket) and executes a predefined sequence. Example: upon receiving a support ticket, the agent extracts context, queries the knowledge base, and drafts an initial response for human validation.
Level 2 — Planning agent
The agent receives a high-level goal ("prepare Friday's sprint report") and decomposes sub-tasks itself: extract Jira data, calculate velocity, generate charts, write the executive summary, post to the right Slack channel.
Level 3 — Collaborative multi-agent
Multiple specialised agents work in parallel and coordinate. An "orchestrator" agent drives "specialist" agents (analysis, writing, verification, sending). This is where productivity gains become exponential.
Concrete use cases in IT project management
"A sprint tracking agent can, every Friday, automatically extract Jira data, calculate team velocity, identify recurring blockers, generate a narrative report and post it to the team Slack channel — with no human intervention."
| Process | Agentivity level | Estimated gain |
|---|---|---|
| Sprint report generation | Level 2 | 4–6h / week |
| Support ticket triage | Level 1 | 2–3h / day |
| New member onboarding | Level 2 | 1 day / hire |
| Competitive intelligence | Level 2 | 3–5h / week |
Challenges nobody talks about
The reliability of external tools is critical — an agent is only as reliable as the APIs it calls. Infinite loops must be contained with proper guardrails (max iteration count, token budgets, timeouts). In Switzerland's regulated sectors, auditability of every agent action is non-negotiable.
How to get started in 2025
My recommendation: start with a Level 1 agent on a non-critical process. Measure the real gains. Iterate. The agile approach applies perfectly to AI agent deployment. Available tools — Claude via API, MCP servers, Gumloop, LangGraph — allow building first prototypes in just a few days.