Zum Inhalt springen
Agent Hub
← Knowledge hub

What is an AI agent?

AI agent? We explain what's actually behind it.

An honest primer: what an AI agent is, how it works, what it already does well — and where 'autonomous AI' is still marketing talk. Reading time: 9 minutes.

What an AI agent actually is

<p id="definition">An <strong>AI agent</strong> is a program built on a large language model (LLM) like Claude, GPT-4 or Mistral that completes <em>tasks on its own</em>: it understands a request, plans steps, calls external tools, checks results, and responds with a finished outcome — not just text.</p><p>Plainly put: a classic chatbot talks, an agent <strong>acts</strong>.</p><p>Where AI agents typically show up today:</p><ul><li>Customer support that actually cancels your order instead of quoting the FAQ.</li><li>Sales advice 24/7 that books a meeting directly in the calendar.</li><li>Back-office workflows: classify mails, extract data, write to the right system.</li></ul>
01features
agent-vs-chatbot

Agent vs. chatbot — the difference at a glance.

  • Classic chatbot

    Follows a decision tree or rule-based script. Replies to recognized phrases with prepared text. Can't execute anything.

    • Script
    • FAQ
    • Static

  • AI agent

    Understands free language, plans in real time, calls APIs, writes to databases, books meetings. Learns from conversations.

    • LLM-based
    • Tool use
    • Adaptive
02architecture

How an agent is technically built.

Three building blocks: a language model as the brain, tools to act with, memory for context and state.

Your request
HTTP · Widget · API
Language model
Understand · plan
Specialist agent
Persona · skills
Router
Intent detection
Memory
Context · history
tools · skills
  • RAG · knowledge
  • CRM API
  • Calendar
  • Email
03features
applications

What agents already do well.

Not everything promised in 2026 actually works in production. These use cases run reliably.

  • Customer support

    Auto-handle standard inquiries, with clean escalation to humans when things get complex.

    • RAG
    • Escalation

  • Sales advice

    Pre-sales questions, lead qualification, meeting booking — 24/7.

    • BANT
    • Calendar

  • Onboarding

    Walk new customers step by step through complex setups.

    • Tutorials
    • Proactive

  • Internal knowledge

    Years of docs queryable by question — for technicians, HR, compliance.

    • Search
    • Audit logs

  • Compliance checks

    Validate contracts, requests or inputs against internal rules.

    • Rules
    • Flagging

  • Operations

    Sort mail, extract data, write to the ERP.

    • Extract
    • Routing

Where the limits are — and why 'autonomous agent' is marketing speak today.

<p id="grenzen">AI agents are not magic. Anyone promising that hasn't shipped one in production. From two years of practice:</p><ul><li><strong>Hallucinations</strong> remain the main problem. Even good models invent plausible-sounding nonsense. Antidotes: <a href="/glossary/rag">RAG</a>, source enforcement, confidence thresholds, clean escalation.</li><li><strong>Complex multi-step reasoning</strong> works sometimes, fails sometimes. Multi-hop is not reliable yet.</li><li><strong>Long conversations</strong> lose context. Even 1M-token windows get confused after 50+ messages.</li><li><strong>"Autonomous" is dangerous</strong>. We build agents with explicit action permissions — critical actions need user confirmation or human approval.</li></ul><p>Our default: the agent solves 80% of requests alone, the remaining 20% get <em>clean handoff</em> to a human — with full conversation context.</p>
04faq

Frequent questions about AI agents.

Sounds like something you need?

Let's talk for 30 minutes. Free, with concrete suggestions for your use case.

Book a callRead the glossary