What is agentic AI?
An agentic AI system receives a goal, decides on the steps, calls tools to carry them out and checks the result. It reads the state of a record before acting, and it writes to the software afterwards.
The difference with classic generative AI comes down to one question: what comes out of the system? An assistant produces text, which someone reads and applies. An agent produces an effect: an entry created, a field filled, a message sent. An assistant’s mistake costs a proofread. An agent’s mistake costs a correction in the database.
That is why agentic AI is judged on different criteria. You no longer measure the quality of an answer, you measure whether an action was right, and what a wrong action costs. The shift is covered in detail on the AI agent page.
What changes compared with a chatbot or RAG?
A chatbot answers from what it knows. RAG adds a search through your documents: it answers from your sources, and can cite them. Both remain systems that talk.
An agent reuses those building blocks and adds three more:
- tools, the functions of your software it is allowed to call, with parameters validated before execution;
- state, what it knows about the current record, before and after it acts;
- a path, the steps it follows, with mandatory checks and the points where a human takes over.
RAG does not disappear. At Archipelia, the agent first searches the ERP documentation to answer, then opens a ticket when it does not have the information. Document search is one step on the path, not the whole system. How RAG works is explained here.
What are the autonomy levels of an agent?
Autonomy is earned, not granted. Four modes cover most of what we put into production:
| Mode | What the agent does |
|---|---|
| On demand | The user asks, the agent executes and shows what it did |
| Triggered | On an event or a schedule: a document arrives, the agent processes it |
| Deterministic | A fixed path, no interpretation, for anything compliance-related |
| Autonomous | The agent reads the state of the business and acts, under supervision |
The same agent can combine several modes depending on the step. What does not change: every move towards more autonomy is decided on measurements, not on a successful demo.
Why does an agent need guardrails?
Because a free model finds a correct path on common cases and invents one on rare cases. And rare cases are exactly where a business has rules.
The guardrails that matter are structural, not instructions added to a prompt:
- a workflow graph that states which steps are permitted, in which order, with which checks;
- the user’s rights: the agent acts under the rights of the person using it, never beyond;
- a confidence threshold below which the agent hands over to a human, with the context already gathered;
- a log of every action: original question, tools called, result.
Our graph approach comes from our research, published at ICMLC 2026. The page on securing an AI agent goes through the protections one by one.
What does agentic AI change for a software company?
Your software records your users’ work. With agents, it can do part of it. That is a feature your customers see, use every day, and that you can bill for.
A public example: Lemon Learning helps companies adopt their software with in-app guides. Its agent navigates enterprise software on its own, understands the interface and creates complete step-by-step guides. Lemon Learning generated its first AI revenue three months after the first conversation. The details are in the case study.
Three conditions come back on every project we deliver:
- the agent works in your screen, under your brand, where your users already are;
- it goes through your APIs, under your roles and rights, without rewriting your product;
- every successful action is counted, which lets you bill it by usage. The credit model explains how.
Where to start?
With an activity, not a model. Pick a task your users perform often in the product, whose result can be checked, and whose mistakes can be corrected. Build the evaluation set before anything else: it is what will tell you whether the agent is ready.
On a known activity with API access, allow six to eight weeks from workshop to production. The steps are laid out in integrating an AI agent into a SaaS product, and the way we run them with you in how we work.
