"AI agent" has become one of the most frequently used phrases in the enterprise software world. Yet the term is applied to very different products: a simple chat assistant and a piece of software that connects to systems and completes work on its own can both be marketed as an "agent." This article explains what an AI agent is, how it differs from an assistant and from classic automation, and the steps a company can follow when setting up its first agent.
What is an AI agent?
Gartner defines agents by setting them apart from AI assistants. According to Gartner, assistants make the user's work easier but depend on human input and do not work independently. When an assistant is given specialist capabilities for a particular task, an agent emerges: an agent can carry out complex work end to end.
Gartner's example is a cybersecurity threat-response agent: it scans network traffic, system logs and user behavior in real time, assesses the situation and, if necessary, initiates a response. In other words, an agent does not merely produce answers; it reads data in pursuit of a goal, makes a decision and initiates an action.
The definition of an AI system in the EU AI Act (Regulation (EU) 2024/1689) also helps in understanding this distinction. The Act defines an AI system as a machine-based system that is designed to operate with varying levels of autonomy and that infers, from the input it receives, how to generate outputs. Agents are systems that sit toward the upper end of this spectrum of autonomy.
Assistant, automation, agent: which one, and when?
The recommendation in Gartner's 2025 announcement is simple: agents when decisions are needed, automation for routine workflows, assistants for simple information retrieval.
| Tool | What it does | Suited to |
|---|---|---|
| Assistant | Answers the user's questions and generates text; a human directs it | Finding information, summarizing, drafting |
| Automation (e.g., RPA) | Applies predefined rules the same way every time | Repetitive tasks with clear rules |
| Agent | Plans the steps itself toward a goal, uses tools and makes decisions | Multi-step tasks that call for case-by-case judgment |
In the same announcement, Gartner also stresses that many use cases presented as "agents" today do not actually require an agent. The first question should therefore be not "which agent should we build?" but "does this work really require a decision?" The selection framework is covered in detail in the article RPA or AI agent?.
The building blocks of an agent
Regardless of product or platform, the elements that make up an agent are, in broad terms, the following:
- Goal and instructions: What the agent should achieve and what it must never do.
- Model: The language model that interprets input and decides on the next step.
- Tools: The connections that let the agent take action, such as reading a record in a system, filling in a form or creating an email draft.
- Context and data: The documents, records and previous steps the agent consults when making decisions.
- Permissions and approval points: Which systems the agent accesses and under which identity, and at which step it waits for human approval.
- Logging and monitoring: The ability to review afterward what the agent did.
What separates an agent from classic automation is the combination of the second and third items: the model itself decides which tool to use and when.
How do you set up an AI agent? Six steps
The sequence below is not an installation guide for a particular product but a general framework for planning a first agent inside a company.
1. Choose the use case. At this early stage, Gartner recommends pursuing the agent approach only where it delivers clear value or return on investment. A good first candidate is a high-volume process whose decisions rest on defined criteria and whose errors are recoverable.
2. Document the process and the success metric. Before the agent goes live, how the work is done today, how long it takes on average and which errors occur should be recorded. Otherwise, improvement cannot be measured at the end of the pilot.
3. Keep tools and permissions to the minimum. OWASP traces the root of the "excessive agency" risk, which leads to uncontrolled agent actions, to three causes: excessive functionality, excessive permissions and excessive autonomy. Its recommendation is to give the agent only the tools its job requires and the lowest level of access. This topic is covered separately in the article on the permissions granted to AI agents.
4. Define the human approval points. OWASP recommends that high-impact actions be approved by a human before they are carried out. Steps such as initiating a payment, deleting a record or sending a formal reply to a customer can be set up, in the initial phase, so that the agent prepares a draft and a person approves it.
5. Build risk management in from the start. The AI Risk Management Framework (AI RMF), published on January 26, 2023 by the U.S. National Institute of Standards and Technology (NIST), is a reference resource designed for voluntary use. The core of the framework consists of four functions: Govern, Map, Measure and Manage. NIST notes that Govern is a cross-cutting function that runs through the other three, and that risk management should continue throughout the life cycle of the AI system. For an agent, this means assigning an owner, listing the risks, measuring performance and defining in advance how to respond when problems arise.
6. Start small, measure, then expand. Gartner notes that integrating agents into legacy systems can be technically complex and that, in most cases, redesigning the workflow from scratch around the agent is the sounder path. A pilot that starts with a limited group of users and a narrow scope makes both these integration problems and the real value visible early.
Why start cautiously?
Expectations are high. Gartner forecasts that by the end of 2026, 40% of enterprise applications will be integrated with task-specific AI agents; the share in 2025 was below 5%. According to Gartner, by 2028 at least 15% of day-to-day business decisions will be made autonomously through agents; in 2024 this share was 0%.
On the other hand, the same firm forecasts that more than 40% of agent projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. Gartner also draws attention to assistants, RPA and chatbots being rebranded as "agents" without genuine agentic capability, and estimates that only about 130 of the thousands of vendors are real. The reasons behind the cancellation risk are covered in detail in the article cancellation risk in AI agents.
Read together, these two pictures send a clear message: agents are becoming an ordinary part of enterprise software, but whether they create value depends on the right use case and the right controls.
A brief note on regulation
The EU AI Act defines a "deployer" as a legal entity that uses an AI system under its own authority. In other words, a company that buys a ready-made agent product and runs it in its own processes also falls into one of the roles the Act defines. Companies that do business in the EU market should assess with their own legal counsel which use case the agent they use operates in and what obligations follow from it.
Conclusion
Unlike an assistant, an AI agent is software that plans toward a goal, uses tools and makes decisions. This capability creates value on the right task; on the wrong one, it creates needless cost and risk. For a company thinking about setting up its first agent, the soundest starting point is to choose a single process that requires decisions, keep the agent's authority narrow, retain human approval at critical steps and measure the result by comparing it with today's process.
Why logistics experience matters here. The six steps above will look familiar to anyone who has run a warehouse or a transport network. Exception handling, such as deciding what to do when an inbound delivery runs late, is a natural first use case (step 1): it is high-volume, governed by defined criteria such as service level, cost and cut-off times, and usually recoverable, because a decision can still be reversed before the truck leaves. Logistics teams also tend to track cycle time and error rate already, so the baseline that step 2 asks for often exists. Role-based access in warehouse systems, supervisor sign-off on write-offs and rolling changes out one site at a time are everyday versions of steps 3, 4 and 6. One of our consultants brings more than 30 years in operations, logistics and supply chain: at a technology distribution company of roughly one billion dollars in scale, he ran logistics, warehousing, technical service and administrative operations end to end, building and leading teams of 100+ people and delivering lasting gains in operational performance through KPI-driven programs. He has also managed nationwide warehouse and transport operations, large-scale distribution networks, and network and capacity planning at leading retail, logistics and distribution organizations. An industrial engineer, he now brings that supply-chain discipline to SMEs through AI and automation. Meet our consultants.
#interimmanagement #forwarddeployedengineering