What makes an AI system agentic?

"Agentic" has become one of those words that gets attached to everything. A chatbot on a website is called an agent. A workflow that sends an email when a form is filled in is called an agent. That makes it hard to tell what is actually being offered, and what it can do for a business.

Here is the plain version, and the one used on every project here.

Three kinds of AI system

1. A chatbot answers

A chatbot takes a question and gives an answer. It can be very good at it. But it waits for a person to ask, and it usually stops at the answer. It does not go and update the CRM, check the order, or send the follow-up.

2. An automation follows a script

A classic automation is a fixed sequence: when this happens, do that. It is fast, cheap and predictable, which is exactly what you want for rule-based steps. It breaks when the input is messy, when something unexpected happens, or when a step needs judgment.

3. An agentic system pursues a goal

An agentic system is given a goal rather than a script. It plans the steps, uses tools and data across your systems, checks the results and adapts when something unexpected happens. It does this within the permissions and rules you set, and it brings in a person when a decision needs one.

The loop that makes it work

What separates a system that can be trusted with real work from a demo is the loop around the model:

1. Plan. Turn the goal into steps. For an incoming request, that might be: read it, find the customer, check the policy, prepare the update.
2. Execute. Do the steps in the real systems: email, CRM, documents, databases, helpdesk.
3. Verify. Check the result against your rules before anything is final. Did the record update? Does the answer match the policy?
4. Escalate. When something falls outside the rules, hand it to a person with the full context, instead of guessing.

Without the verify and escalate steps, you have a clever assistant that can make expensive mistakes quickly. With them, you have something you can put into production.

Multi-agent systems

Larger processes are often split across several specialist agents, each with a clear role and only the access that role needs. A coordinator agent manages them: it plans the work, assigns each step and collects the results. One agent reads and sorts incoming requests, another checks them against your rules, a third updates your systems.

Splitting the work this way makes each agent easier to test, easier to limit and easier to improve, the same way a team works better when roles are clear.

Agentic does not mean "AI everywhere"

One of the most useful guides on the subject, Anthropic's Building effective agents, makes a point worth repeating: start with the simplest solution that works, and add complexity only when it clearly improves the result. Plenty of steps in a business are better served by a plain automation. Agents belong where the work needs judgment and the input is messy.

That is why every build here starts by mapping how the business runs, end to end, and deciding step by step where an agent belongs and where simpler automation is enough.

A quick test

When someone offers you an "AI agent", ask four questions:

• What goal is it given, and who sets it?
• Which of my systems can it act in, and with what permissions?
• How does it check its own work before anything is final?
• When does it stop and bring in a person?

If the answers are clear, you are looking at an agentic system. If they are vague, you are probably looking at a chatbot with a new name.