Agentic AI: what it is, and what it changes for your business

What is agentic AI?

Agentic AI is AI that pursues a goal rather than answering a single request. It breaks the goal into steps, uses tools to act in other software, looks at the results and decides what to do next. When Gartner named it the first of its top strategic technology trends for 2025, in October 2024, it defined agentic AI systems as systems that “autonomously plan and take actions to meet user-defined goals”.

The word describes a property more than a product. A system can be a little agentic (the model only picks which of three routes a request takes) or very agentic (the model plans a multi-step job, chooses its tools and decides when it is finished). Anthropic, which makes the Claude models, groups both under the label “agentic systems” and separates them by who holds control. In workflows, models and tools follow “predefined code paths”. In agents, models “dynamically direct their own processes and tool usage”.

An AI agent is the unit of agentic AI: a concrete piece of software that works this way. Our guide What is an AI agent? covers the agent itself, its components and its types. This guide looks at the shift as a whole, and at what it means for a company deciding what to build. For concrete tasks by department and industry, see our guide to AI use cases.

Agentic AI vs generative AI

Generative AI produces content on request; agentic AI uses that ability to get work done. The models are often the same. What changes is what surrounds them: a goal instead of a prompt, tools instead of a text box, and a loop that runs until the goal is met.

The last two rows matter most for a buyer. A generative tool that makes a mistake produces a bad draft that someone will probably catch. An agentic system that makes a mistake can send the wrong email to a customer. That is why agentic projects need a design for permissions, approvals and logs, which a writing assistant never did.

From workflow to autonomous agent: the degrees of agency

Agency works like a dial with several settings. Anthropic’s engineering guide describes five workflow patterns, each giving the model a little more say, before reaching fully autonomous agents. Knowing where your use case sits on that dial is the most useful decision in an agentic project.

Anthropic’s advice is to climb the dial only when you must. It recommends “finding the simplest solution possible, and only increasing complexity when needed”, adding: “This might mean not building agentic systems at all.” Agents are for “open-ended problems where it’s difficult or impossible to predict the required number of steps”, and they come with “higher costs, and the potential for compounding errors”. Many business tasks are best served by the first three rows, with one agentic step where judgment is really needed.

Why agentic AI is taking off now

Three things changed at once: models can carry longer tasks, connecting them to business software has become standard, and the major vendors now ship the building blocks.

Longer tasks. In March 2025, the research group METR measured the length of software tasks, counted in the time they take a human professional, that frontier AI agents complete with 50% reliability. It found that this length had doubled about every 7 months over the previous six years. METR notes that the figure reflects its data at publication and depends on methodological choices, so treat it as a measured trend that may bend.

Shared connections. An agent is only as useful as the systems it can reach. The Model Context Protocol (MCP), open-sourced by Anthropic in November 2024, gives models a common way to connect to tools and data. When Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation in December 2025, it counted more than 10,000 active public MCP servers, and listed ChatGPT, Gemini, Microsoft Copilot and Visual Studio Code among the products that had adopted it.

Vendor building blocks. OpenAI, Google, Microsoft, Anthropic and Mistral all publish agent frameworks or agent APIs. Our guide on how to build an AI agent compares these routes with no-code tools and custom builds.

What the numbers say, hype included

Interest is real, and so is the failure rate. In a January 2025 Gartner poll of 3,412 webinar attendees, 19% said their organization had made significant investments in agentic AI and 42% conservative ones; 8% had made none, and 31% were waiting or unsure.

The same Gartner release, published in June 2025, predicts that over 40% of agentic AI projects will be canceled by the end of 2027, “due to escalating costs, unclear business value or inadequate risk controls”. It warns against “agent washing”, the relabeling of chatbots, assistants and robotic process automation as agents, and estimates that only about 130 of the thousands of agentic AI vendors are real.

Gartner still expects the shift to happen. It predicts that at least 15% of day-to-day work decisions will be made autonomously through agentic AI by 2028, up from 0% in 2024, and that 33% of enterprise software applications will include agentic AI by 2028, up from less than 1% in 2024. Both are forecasts.

What we take from these figures: projects rarely fail because the models are too weak. They fail when nobody priced the running cost, measured the value or designed for the risk, and each of those can be settled before the first line of code. That is where our method starts.

The risks that come with autonomy

The more a system can do, the more a mistake costs. Agentic AI adds two risks that plain generative tools do not carry in the same way.

The first is doing too much. The OWASP Top 10 for LLM applications (2025) calls it “Excessive Agency” and traces it to excessive functionality, excessive permissions and excessive autonomy. Its example is telling: an extension able to delete documents that “performs deletions without any confirmation from the user”. The remedies are simple to state: give the agent only the tools it needs, the narrowest permissions that work, and a human approval on high-impact actions.

The second is being steered. OWASP ranks prompt injection first in the same list: instructions hidden in the content a model reads, including external sources “such as websites or files”. It adds that, given how models work, “it is unclear if there are fool-proof methods of prevention”. For an agent that reads customer emails or supplier PDFs, this exposure is daily. The defense is architectural: treat everything the agent reads as untrusted, keep reading and acting separate, and make sensitive actions depend on checks the model cannot talk its way around.

OpenAI’s guide gives two triggers for handing control back to a person: when the agent exceeds a set number of failed attempts, and before actions that are “sensitive, irreversible, or have high stakes”, such as canceling orders, authorizing large refunds or making payments.

Agentic AI under European rules

The AI Act does not have a separate category for agents; it regulates AI systems by use. Its definition of an AI system, in Article 3, already speaks of a machine-based system “designed to operate with varying levels of autonomy”, so agentic systems fall squarely inside it.

Three points matter for most companies:

• Transparency. Article 50 requires that people be informed they are interacting with an AI system, unless it is obvious. The European Commission states that the transparency rules apply from August 2026. A customer-facing agent must say what it is.
• High-risk uses. An agent used in recruitment, credit scoring or access to essential services falls into the high-risk category, with obligations the Commission now dates from 2 December 2027, including “appropriate human oversight measures” and activity logging.
• Automated decisions. Under Article 22 of the GDPR, people have the right not to be subject to a decision based solely on automated processing when it has legal or similarly significant effects.

The rules are still being adjusted, so check the dates when you plan. Our guide to EU AI Act compliance tracks them. If your data or your model must stay in Europe, see our page on private LLMs hosted in Europe.

Where to start with agentic AI

Start from one task with a visible cost, and choose the lowest degree of agency that does the job. A good first project has a clear owner, a measurable result, and enough variation in its inputs that a script keeps failing.

In practice, that means three moves. Pick one process and measure how long it takes today. Decide which steps need judgment and which can stay fixed code. Give the agentic part narrow rights and a human approval on anything sensitive, then widen its scope as the error rate allows.

To see what a narrow, well-scoped system looks like, try our knowledge assistant demo: it answers from a fixed set of documents and cites its sources. When you are ready to build, see how we design AI agents for business and how we connect them to your existing tools through AI integration services. If your team wants to build its own agents, our AI training for employees is built around your tools and cases, and each team leaves with a first agent.

The quickest way to find your own first step is a conversation about one real process. Book your free 30-minute assessment and bring the task that keeps coming back: we place it on the dial above, tell you which steps an agent should own and which should stay fixed, and what a first version would take. We reply within one business day.