1.10Beginner7 min

AI Agents: FAQ for Decision-Makers

Blck Alpaca·8 June 2026

Definition

An AI Agent is a software-based system built on a (Large) Language Model that autonomously pursues a defined goal: it perceives its environment, plans across multiple steps, independently selects and uses external tools (tools, APIs, data sources), observes the results, and iteratively adjusts its plan until the goal is reached or aborted. The key point for decision-makers: an agent only pays off when the solution path cannot be planned in advance.

Key Takeaways

✓An AI Agent differs from chatbots, RPA, and workflow automation through four mandatory properties: LLM-based control, multi-step planning (Perceive→Reason→Act→Observe), dynamic tool use, and goal-oriented autonomy within guardrails.
✓Agents only pay off when the solution path cannot be planned in advance. For fixed paths, workflow automation or RPA are cheaper, more robust, and lower in compliance burden.
✓Market maturity is muted: according to McKinsey State of AI 2025, only 23 percent scale at least one agentic use case; Gartner (June 2025) expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027.
✓Compliance belongs in the picture from day 1: EU AI Act (Art. 50 transparency from 02 Aug 2026), GDPR DPIA (Art. 35), as well as the works council under BetrVG §87 (DE) or ArbVG §96 (AT). Note: informational, not legal advice.
✓Pragmatic start: 1 to 3 use cases with clear ROI, a low-risk pilot with human-in-the-loop, outcome KPIs instead of activity metrics, vendor-agnostic framework choice based on use case.
✓Costs are often underestimated: token and reasoning costs drive the effort; according to Bitkom 2026, 33 percent of companies report higher AI costs than expected. Routing and caching dampen this.

For decision-makers, the most important questions can be condensed into three core answers:

What is different? An agent plans dynamically and selects tools itself, instead of following a predefined path. This is exactly what distinguishes it from a chatbot, RPA, and workflow automation.
When does it pay off? Only when the solution path cannot be planned in advance. If the process is known, classic automation or RPA are cheaper and more stable.
What matters when introducing it? A clear use case with ROI, human-in-the-loop for irreversible actions, DPIA and works council from the start, thinking vendor-agnostically.

A concrete example

A typical L4 agent (autonomous agent) is a research assistant tasked with answering an open question. It breaks the goal down into sub-steps, independently invokes a web search, reads results, formulates follow-up questions, checks for contradictions, and finally summarizes an evidence-backed answer. No one prescribed the exact order of the search queries. It is precisely this dynamic control within the reasoning loop (Perceive→Reason→Act→Observe) that makes the difference from a hard-wired pipeline.

Where does the market currently stand?

Expectations are high, scaling is muted. According to McKinsey State of AI 2025, only 23 percent of companies scale at least one agentic use case, with a further 39 percent experimenting; in no function does the share of scaled use cases exceed 10 percent. Gartner (June 2025) expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027, often due to an unclear business case and underestimated effort. For decision-makers, this means: prioritize soberly instead of rolling out broadly.

Agent, chatbot, RPA, workflow, or assistant?

The following matrix classifies the terms that are frequently conflated in the market:

Criterion	AI Agent	Chatbot	RPA	Workflow automation	Assistant
Trigger	Goal event	Message	Time rule	Event	Prompt
Reasoning	LLM, multi-step	Intent matching	none (rule script)	conditional logic	LLM, single-step
Tool use	dynamic, many	usually none	UI bots, screen scraping	prebuilt connectors	limited
Memory	Short- and long-term	Session	none	Workflow state	Session
Autonomy	high	very low	medium, scripted	low	low to medium
Maintenance	high	low to medium	medium to high (UI breaks)	low	low

Key principle: workflow automation and RPA follow a predefined path, whereas an agent decides the path dynamically at runtime. This also entails the agent's higher maintenance and compliance effort.

How mature is "my" agent? The levels L1–L5

L1 Reflex: rule-based, no LLM (FAQ bot, thermostat).
L2 Augmented LLM: LLM plus one tool call, reactive (e.g. ChatGPT with web search).
L3 Workflow agent: LLM in a deterministic pipeline (prompt chaining, routing).
L4 Autonomous agent: LLM controls sequence and tool selection dynamically, full loop (e.g. Claude Code, Deep Research).
L5 Multi-agent system: several autonomous agents coordinate via A2A (orchestrator plus specialists).

This scale helps in vendor conversations: anyone who says "autonomous agent" but only delivers L2/L3 is engaging in "agent washing". According to Gartner, only about 130 of thousands of providers have genuine agent capabilities.

Standards: MCP and A2A

Two open standards structure the ecosystem. MCP (Model Context Protocol) standardizes the agent↔tool connection; the 2025-11-25 specification was handed over to the Linux Foundation in December 2025, and there are more than 10,000 MCP servers. A2A (Agent-to-Agent) standardizes agent↔agent communication and has been with the Linux Foundation since June 2025, with more than 150 participating organizations. For decision-makers, both are an argument for vendor-agnostic architectures that reduce lock-in.

Compliance: what DACH decision-makers need to clarify early

Note: informational, not legal advice; legal deadlines partly provisional.

EU AI Act: Transparency obligations under Art. 50 apply from 02 Aug 2026. The AI literacy obligation (Art. 4) has already applied since 02 Feb 2025. High-risk classifications under Annex III may, according to the Digital Omnibus (07 May 2026), be postponed to 02 Dec 2027; until formal adoption, 02 Aug 2026 remains legally decisive.
GDPR: For agentic systems, a Data Protection Impact Assessment (Art. 35) is usually required; Art. 22 (automated decisions) and Art. 28 (processing on behalf) are also relevant.
Co-determination: Before introduction, the works council must be involved — in Germany under BetrVG §87, in Austria under ArbVG §96.

Common mistakes when introducing agents

The most expensive pitfalls are rarely technical. Frequently the framework is chosen before the use case, data preparation is underestimated, the DPIA is skipped, or human-in-the-loop is omitted for irreversible actions. Equally risky: a token cost explosion without routing and caching, treating agents as deterministic, and a lack of observability. A pragmatic start addresses exactly these points: 1 to 3 use cases with measurable ROI, a low-risk pilot with human approval, outcome KPIs instead of activity metrics, as well as DPIA and works council from day 1.

FAQ

What is the difference between an AI Agent and a chatbot?

A chatbot responds to individual messages via intent matching and usually works without tools and without long-term memory. An AI Agent pursues a goal: it plans across multiple steps, independently selects tools, and iteratively adjusts its plan. The chatbot replies, the agent acts in a goal-directed way.

What is the difference between Agentic AI and an AI Agent?

Agentic AI denotes the overarching paradigm of autonomous, goal-directed AI action. An AI Agent is the concrete component or the individual system that implements this paradigm. In short: Agentic AI is the concept, the AI Agent is the manifestation.

Do we need our own LLM to deploy agents?

No. Agents can be run via API LLMs (for example from Anthropic or OpenAI); a self-trained model is not required. What matters instead is data connectivity, tool integration, guardrails, and data protection, not model ownership.

When does an agent pay off compared to workflow automation or RPA?

Workflow automation and RPA follow a predefined path and are cheaper and more robust for known, stable processes. An agent only pays off when the solution path cannot be planned in advance and must be decided dynamically, i.e. for open, variant-rich tasks.

How do we choose the right agent framework?

The selection follows the use case, not the other way around. Common options are LangGraph, CrewAI, OpenAI Agents SDK, Anthropic Claude Agent SDK, Microsoft Agent Framework, n8n, and Pydantic AI. The decisive factors are the required level of autonomy, tool integration, observability, and a vendor-agnostic architecture.

What are MCP and A2A and why are they relevant for us?

MCP (Model Context Protocol) is the standard for the agent↔tool connection, with more than 10,000 servers and with the Linux Foundation since December 2025. A2A (Agent-to-Agent) standardizes communication between agents and has been with the Linux Foundation since June 2025. Both reduce lock-in and promote interoperable architectures.

Which legal obligations apply in the DACH region? (informational, not legal advice)

Relevant are the EU AI Act (transparency under Art. 50 from 02 Aug 2026, AI literacy under Art. 4 since 02 Feb 2025), the GDPR (DPIA under Art. 35 usually required, Art. 22 and 28), as well as co-determination via BetrVG §87 in Germany or ArbVG §96 in Austria. Deadlines for high-risk are partly provisional.

Is a Data Protection Impact Assessment (DPIA) required for agentic systems?

In most cases, yes. Agents frequently process personal data autonomously and across multiple tools, which triggers a DPIA under GDPR Art. 35. It should be carried out before the pilot phase, not afterwards. (informational, not legal advice)

Does the works council need to be involved?

As a rule, yes. If systems are deployed that can capture the behavior or performance of employees, co-determination applies — in Germany under BetrVG §87, in Austria under ArbVG §96. Involvement should take place from day 1. (informational, not legal advice)

Why are agents often more expensive than expected and how do you dampen the costs?

Multi-step reasoning and many tool calls generate high token consumption. According to Bitkom 2026, 33 percent of companies report higher AI costs than expected. Routing (smaller models for simple steps) and caching reduce the costs noticeably, as do loop limits and clear abort criteria.

Why do so many agent projects fail?

According to Gartner (June 2025), more than 40 percent of agentic AI projects are likely to be cancelled by the end of 2027. The main reasons are an unclear business case, choosing the framework before the use case, underestimated data preparation, a lack of observability, and a missing human-in-the-loop. Success comes from focused use cases with measurable ROI.

How do we measure the success of an agent?

Through outcome KPIs instead of activity metrics: which business goal was achieved, at what cost, with what error and escalation rate. Activity figures such as the number of tool calls say little about value. A low-risk pilot with clear target values is the pragmatic entry point.

Want to go deeper?

Get new analyses straight to your inbox – or see how we put this knowledge to work for companies.

Subscribe to newsletter →Our services

Previous← The History of AI Agents