Agentic AI is an AI paradigm in which software systems – typically based on foundation models – pursue goals autonomously or semi-autonomously: they perceive, plan, use external tools and data, take actions and learn from the outcome, with minimal ongoing human intervention. In contrast to generative AI, which reactively produces content, agentic AI acts proactively and in a goal-oriented manner. An AI agent is the concrete technical artefact; agentic AI is the overarching class.

This distinction is not an academic detail but the decisive filter against what market observers call "agent-washing" – the relabelling of existing chatbots or RPA tools as "agentic". This article accompanies the pillar "Agentic AI vs. classic AI" and clarifies the concept, characteristics and market development.

The three most important points up front:

• Paradigm vs. artefact: Agentic AI is the class of systems, an AI agent is the concrete implementation within it. Vendor marketing often uses both terms synonymously – which is incorrect.
• Five core characteristics: Autonomy, goal orientation, tool use, reasoning/planning and memory separate agentic AI from reactive GenAI and classic machine learning.
• A market between hype and reality: McKinsey reports 62 percent experimenting, but only 23 percent scaling; Gartner forecasts more than 40 percent project cancellations by the end of 2027 (explicitly a forecast).

What does "agentic AI" mean? The concept in detail

The core of the term lies in the word "agency" – the ability to act independently and purposefully. IBM formulates the canonical definition as follows: unlike traditional AI models that operate within predefined boundaries and require human intervention, agentic AI exhibits autonomy, goal-driven behaviour and adaptability. The term "agentic" refers to the capacity of these models to act independently and with intent.

Important for editorial accuracy: even in mid-2026, "agentic AI" is still a contested term that has not been finally consolidated. There is no ISO/IEC standardisation in place (as of 2026). Providers vary in the details – particularly on the question of whether deterministic workflows already count as "agentic". Anthropic groups both fully autonomous systems and predefined automated workflows under "agentic systems" and, within this, distinguishes workflows (orchestrated) from agents (self-directing). Databricks, by contrast, strictly separates generative from agentic AI. This definitional fuzziness is not a shortcoming of the concept but a market risk that decision-makers need to be aware of.

The five core characteristics of agentic AI

Across the various vendor definitions, five recurring characteristics can be isolated. Together they constitute what makes a system "agentic" – individual features alone are not sufficient.

• Autonomy: The system acts independently towards a goal without every step having to be instructed. Gartner explicitly includes semi-autonomous systems as well – relevant for regulated DACH industries in which human-in-the-loop remains mandatory.
• Goal orientation: Rather than answering a single prompt, the agent pursues an overarching goal, derives sub-goals from it and adapts its approach.
• Tool use: The agent calls external tools – APIs, databases, search systems, entire software interfaces (computer-use). Only this enables it to act on the real world rather than merely producing text.
• Reasoning and planning: Multi-step inference, the decomposition of complex tasks and self-correction. NVIDIA describes this as a loop of Perceive → Reason → Act → Learn.
• Memory: A persistent state across multiple steps and sessions – beyond the transient context window. Databricks explicitly names the retention of memory across steps as a defining characteristic.

Agentic AI vs. AI agent: keeping paradigm and artefact cleanly separate

The most common terminological confusion concerns the relationship between agentic AI and AI agent. IBM puts it succinctly: agentic AI is the broader concept of solving a problem with limited oversight, whereas an AI agent is a specific component within that system which takes on tasks with a certain degree of autonomy.

A rule of thumb for practice: agentic AI is the paradigm, the AI agent is the artefact. Anyone who conflates these levels in tenders, pitches or strategy papers invites inflated expectations.

Distinction from classic AI: the comparison table

The fastest way to understand agentic AI is through the contrast with classic AI – that is, with generative AI and classic machine learning, which operate within predefined boundaries.

The decisive leap lies in the "External action" column: as soon as a system not only produces content but acts on live systems, the risk shifts from the informational (a wrong answer) to the operational (an incorrectly triggered transaction). This is the real reason why agentic systems need different governance from a chatbot.

Market development and forecasts – cleanly contextualised

When it comes to market figures, discipline in labelling sources is essential: survey data (measured) and forecasts (projected) must not be mixed.

Measured adoption (survey): In November 2025, McKinsey reports, based on a survey of 1,993 organisations in 105 countries: 88 percent use AI in some form. With regard to agents, 62 percent are experimenting, yet only 23 percent are scaling an agentic system in at least one business function – and in no single function is the scaling rate above 10 percent. For the DACH region, Bitkom measures AI usage broadly at 36 percent (German companies with 20 or more employees, 2025, a doubling compared with 20 percent in the previous year). There is no separate agentic AI metric for DACH; the global McKinsey figures serve only as a proxy.

Forecasts (explicitly a forecast, not fact): The most-cited figure comes from Gartner (press release of 25 June 2025): more than 40 percent of agentic AI projects will be cancelled by the end of 2027 – owing to escalating costs, unclear value contributions and inadequate risk controls. This is an analyst forecast that Gartner can revise. In the same breath, Gartner estimates that of thousands of vendors only around 130 possess genuine agentic capabilities – a non-reproducible analyst assessment, not a measured value. Gartner further forecasts that by 2028 around 15 percent of day-to-day work decisions will be made autonomously by agentic AI (up from 0 percent in 2024) and that 33 percent of enterprise software applications will contain agentic functions (up from under 1 percent in 2024).

The message for decision-makers: the market sits between broad experimentation and narrow productive scaling. Anyone who cites forecast figures as established facts undermines their own credibility.

Practical example: how to recognise genuine agentic AI

A concrete identifying feature beyond the marketing is the execution profile. Real agentic sessions – for example in software engineering with coding agents such as Claude Code – reach, according to NVIDIA telemetry, context windows of 10,000 to over 150,000 tokens per session. This is the measurable signature of multi-step, tool-using goal pursuit, whereas a pure question-and-answer service makes do with a fraction of that.

In pseudocode, the logic differs fundamentally:

```

Classic GenAI (reactive, single step)

answer = model.generate(prompt)

Agentic AI (goal-driven loop)

goal = "Repair reporting pipeline"
state = memory.load()
while goal not reached:
plan = model.plan(goal, state)
result = tool.execute(plan.next_step) # Tool use
state = model.reflect(result, state) # Reasoning
if plan.is_irreversible:
wait_for_approval() # Human-in-the-loop
```

It is precisely this loop of planning, acting, reflecting and replanning – combined with persistent state and tool calls – that makes a system agentic. Where it is absent, what you have is classic GenAI or automation, despite the "agent" label.

For agencies and B2B decision-makers

For marketing agencies and DACH B2B decision-makers, clean terminology is a concrete competitive advantage: anyone who can explain to clients why not every "agent" in a vendor pitch is genuinely agentic, and which five characteristics matter, positions themselves as a reliable advisor rather than a hype amplifier. Always treat forecast figures as forecasts, label version and market states with "as of 2026", and keep paradigm and artefact separate – this protects your credibility and that of your clients.

Want to receive terminological clarity and well-founded context on agentic AI on a regular basis, before the next buzzword does the rounds? Subscribe to our newsletter – concise, source-based and free of marketing waffle.