What is agentic AI? A practical guide for business leaders.

A working definition of agentic AI

Agentic AI describes software systems that are given a goal, a set of tools and a mandate, and that work toward the goal over several steps. Instead of producing one answer to one prompt, an agent plans what to do, calls tools such as a search index, a database, an email account or your ERP, looks at what came back, and decides on the next step. It keeps going until the task is done, blocked, or handed to a person.

The reasoning core is usually a large language model. What turns a model into an agent is everything around it: the loop that lets it act repeatedly, the tools it may use, the memory it keeps and the limits you set. Autonomy is therefore a design decision, not a property of the model. The same model can power an assistant that only drafts and an agent that closes tickets on its own.

A useful test: if the system can complete a task you would otherwise assign to a colleague, including the tedious middle part of looking things up in three systems, it is an agent. If it only answers, it is a chatbot. If it helps you while you do the work yourself, it is a copilot.

Agents, chatbots and copilots compared

Vendors use the three terms almost interchangeably, which is why so many companies are surprised when their chatbot turns out to be unable to do anything. The difference lies in who drives the work and what happens between the question and the result.

Chatbots and copilots are interfaces to a model. Agents are workers with a mandate. Both have their place, and in practice a chat window is often the front door to an agent with real tools behind it. What matters for your risk and your business case is not the label but the last row of the table: who acts, and who checks.

The six building blocks of an AI agent

Every production agent, whatever the framework or platform, consists of the same six parts. Knowing them helps you ask vendors and your own team the right questions.

• Model. The language model that reads, reasons and writes. It is the most visible and, increasingly, the least differentiating part. Model choice matters for cost, response time, language quality in German and where your data is processed.
• Tools and APIs. What the agent may do: search a knowledge base, read a ticket, query the ERP, send an email, create a record. Tools are where the agent meets your systems; standards such as the Model Context Protocol make this connection reusable.
• Memory. Short-term memory is the context of the current task. Long-term memory stores what the agent should remember across runs, such as customer preferences or earlier decisions. Both need rules about what may be stored and for how long.
• Orchestration. The logic that runs the loop: plan, act, observe, repeat. It handles retries, splits large tasks into smaller ones, hands work to specialised sub-agents and decides when to stop or escalate.
• Guardrails. Permissions, spending limits, approval steps, forbidden actions, checks on inputs and outputs. Guardrails turn a capable model into a system you can trust with real accounts. We wrote a separate guide on guardrails for AI agents.
• Evaluation. Test cases, quality metrics, logs and monitoring. Without evaluation you cannot tell whether a prompt change made the agent better or worse, and you cannot show an auditor what the agent did and why.

When a demo looks impressive but the project stalls, it is almost always because the last three parts were missing. Our AI agent development work starts with orchestration, guardrails and evaluation for exactly this reason.

Four levels of autonomy

The most consequential decision about an agent is not which model to use but how much it may do without asking. We describe this on four levels, and we set the level per action rather than per agent: the same agent may draft freely, propose with reasoning, and require approval before it sends anything to a customer.

Most agents in production sit at level 2 or 3, and that is not a compromise but the design. Level 4 is earned action by action, when evaluation data shows that the agent's error rate on that step is lower than the cost of a review. For some uses the EU AI Act requires meaningful human oversight regardless; our plain-language guide to the EU AI Act explains which.

What agents do in practice, department by department

The pattern repeats across functions: messy input, rules that need some judgement, several systems, and a person who currently does the glue work. The following examples are typical scenarios, not a catalogue.

• Marketing. An agent turns a product brief into channel-specific drafts, checks them against brand and legal guidelines and prepares the campaign in your tools for a marketer to approve. See AI agents in marketing.
• Sales. Inbound leads are researched, qualified against your criteria, enriched in the CRM and routed to the right person with a suggested first reply. See AI agents in sales.
• Customer service. Tickets are classified, the relevant order and contract data is pulled together, a reply is drafted, and simple cases such as address changes are executed after approval. See customer service.
• Operations. Orders that arrive as PDFs or emails are read, validated against price lists and stock and entered into the ERP; discrepancies go to a person. See operations.
• Finance. Incoming invoices are matched to purchase orders and goods receipts, exceptions are explained, and postings are proposed for the accountant to release. See finance.
• HR. An assistant answers policy questions from the handbook, prepares onboarding checklists and drafts documents, while every decision about people stays with people. See HR.
• Legal and compliance. Contracts are reviewed against your playbook, deviations are flagged with the clause and the reason, and standard agreements are pre-filled. See legal and compliance.
• IT and engineering. Incidents are enriched with logs and history, runbook steps are executed after approval, and internal requests are resolved without a ticket queue. See IT and engineering.

When not to use an agent

Agents are expensive when misapplied. In four situations we advise against them, at least for now.

• The process is fully deterministic and stable. If the input is structured and the rules never need judgement, a script, an integration or RPA is cheaper and more reliable. Our comparison of AI agents, RPA and chatbots goes into detail.
• A wrong action is irreversible and there is no review step. Payments, terminations, public statements. Either add a level-3 approval or keep the agent at level 2.
• The decision is legally sensitive. Hiring, credit, access to essential services and similar uses are treated as high-risk under the EU AI Act. Agents may assist there, but design, documentation and oversight follow different rules.
• There is no baseline and no volume. If nobody can say how long the process takes today or how often it runs, you cannot show that the agent helped, and a process that runs twice a month rarely justifies the effort.

How to start

Start with a diagnosis, not a demo. An AI readiness assessment tells you in two to three weeks where your data, systems, skills and governance stand and which two or three use cases are realistic. Then pick one deliberately: our guide on choosing your first AI agent use case contains the scoring framework we use ourselves.

From there, a pilot of typically six to eight weeks should end with a production-grade agent for one workflow, a measured result against the baseline and the governance you need to run it. Not a slide deck about what an agent could do.