Why Many AI Bots Disappoint
Many companies introduced chatbots in 2023 and 2024 and found that they hardly took any work off anyone's desk. The reason is simple: they are not agents.
- Chatbot: "Here is the link to the invoice." (passive)
- Agent: "I created the invoice, posted it in the ERP and sent it to the customer." (active)
Mid-sized companies lag behind with real agents. According to McKinsey's "The state of AI in 2026", 40 % of companies with more than US$ 1 billion in revenue scale AI agents in at least one function; among smaller organisations the figure is 22 %. Our sister project tracks these and other figures every quarter in the Agentic Organization Index (in German).
What Makes AI "Agentic"?
- Tool calling: the model calls APIs, databases and other systems, today often via the open MCP protocol
- Decision logic: clear rules for when the agent acts and when it hands over to a human
- State: context is kept across several steps and sessions
- Error handling: failures are caught without derailing the process
The Tool Landscape in Autumn 2026
For code-based projects, providers today mostly work with LangGraph, the Microsoft Agent Framework (successor to AutoGen and Semantic Kernel), the OpenAI Agents SDK, Google's ADK, the Claude Agent SDK or CrewAI. For process automation without much code, platforms such as n8n or Microsoft Copilot Studio are common. Which one fits depends on your existing stack, not on trends.
The Hard Part: Control Without Losing Autonomy
- Too much autonomy: the agent makes expensive mistakes
- Too much control: the agent is just a chatbot with extra steps
Good providers define permissions per tool, set limits on cost and number of steps, log every action and build in approvals for critical steps.
Chatbot Development vs. Agent Development
| Chatbot development | Agent development |
|---|---|
| Prompt engineering | Software architecture |
| Conversation design | API integration and permissions |
| Answer quality | State, orchestration and recovery |
| Manual testing | Automated evaluation and tracing |
Typical Use Cases
- Finance: invoice processing, pre-checking approvals, reconciliation
- Customer service: resolving tickets end to end, not just answering
- HR: onboarding workflows, document generation
- Operations: supply chain monitoring, quality checks
Compliance
Besides the GDPR, the EU AI Act applies. Which obligations affect you depends on the risk class of the use case, and the timeline is being phased in. The Digital Omnibus on AI (Regulation (EU) 2026/1744, in force since 27 July 2026) postponed the obligations for high-risk systems: to 2 December 2027 for stand-alone systems under Annex III and to 2 August 2028 for systems in products under Annex I. Ask providers how they document data flows, decisions and human oversight.
Finding Providers
In our directory you can see for each entry whether and when we checked existence, location, website and agent offering. We do not rate quality. Use the checklist to check technical depth yourself.