Assistant
Answers questions and drafts texts. Every answer needs a new request; nothing is acted on.
Insights · AI
Agentic AI: when AI does not just answer, but acts
AI agents plan steps, use tools and complete tasks on their own. That promises a lot and often fails at the same points: unclear value, missing interfaces, no rules for approvals.
By Michele Mangieri · Updated
In short
Agentic AI refers to AI agents that do not just answer but act: they receive a goal instead of a question, plan the steps, call systems and tools and check interim results themselves. That sets them apart from assistants, which answer questions. Projects often fail at the same points: unclear value, missing interfaces and no rules for approvals.
From chatbot to agent
Four levels that are often confused.
Embedded assistant
Works directly inside an application, such as Microsoft 365 Copilot or SAP Joule, and makes suggestions. Whether a suggestion is applied is up to a person.
Agent
Receives a goal instead of a question. Plans the steps, calls systems and tools and checks interim results itself.
Multi-agent system
Several specialised agents share one task. A control unit distributes the work and brings the results together.
What Gartner expects
High potential, high cancellation rate.
33%of enterprise software will include agentic AI by 2028 (2024: less than 1%)
15%of day-to-day work decisions will be made autonomously by AI agents by 2028 (2024: 0%)
over 40%of agentic AI projects will be cancelled by the end of 2027
Gartner names rising costs, unclear business value and inadequate risk controls as reasons for the cancellations. On top of that comes “agent washing”: vendors rebrand chatbots and conventional automation as agents. Of thousands of vendors, Gartner says only around 130 offer genuine agentic features.
Source: Gartner, predictions on agentic AI, most recently in the press release of 25 June 2025. Gartner is a registered trademark of Gartner, Inc. and/or its affiliates.
What matters
An agent is only as good as the landscape it works in.
- A measurable case first: one process with a clear goal, known effort and a metric for success.
- Interfaces instead of screen clicks: agents need stable APIs and clean master data. Standards such as the Model Context Protocol (MCP) and Agent2Agent (A2A) define how agents address tools and other agents.
- Own identity, narrow rights: each agent only gets the permissions its task requires, and every action is logged.
- Approvals by risk: anything touching money, contracts or customer data goes past a person before it is executed.
- In the SAP world, SAP brings its own agents into the processes with Joule Agents. How well they work depends on a clean core and documented interfaces.
From our work
We work AI first and use various AI agents ourselves, for example with Claude Code. Our rule: the agent assists, the responsibility stays with people. The fundamentals teams need for this are covered in AI literacy. How agents fit into an overall architecture is covered in AI reference architecture.
Wondering where AI agents could add value for you? AI Decision Readiness →
Frequently asked questions
Questions about Agentic AI
What is the difference between an AI assistant and an AI agent?
An assistant answers questions and drafts texts, but nothing is acted on. An agent receives a goal, plans the steps, calls systems and tools and checks interim results itself. In between sits the embedded assistant, which makes suggestions inside an application.
Why are so many agentic AI projects cancelled?
Gartner names rising costs, unclear business value and inadequate risk controls. It expects over 40% of agentic AI projects to be cancelled by the end of 2027. On top of that comes agent washing: vendors rebranding chatbots and conventional automation as agents.
What do AI agents need to work in a company?
A landscape they can work in: a measurable case, stable APIs and clean master data, their own identity with narrow rights and logged actions, and approvals by risk for anything touching money, contracts or customer data.
What are MCP and A2A?
Standards that define how agents address tools and other agents. MCP stands for Model Context Protocol, A2A for Agent2Agent. For this, agents need interfaces instead of screen clicks, meaning stable APIs, along with clean master data.
30 minutes, and you will know whether we can help.
You describe the decision ahead. We tell you which package fits, or that none does.