Agentic AI: definition, examples, and what we really hand to an agent
Agentic AI refers to AI agents that carry out a series of actions in your tools to reach a goal. A definition, two agents Idemys runs for itself, the grid to choose the first task to hand to an agent, and what we never hand to one.
The goal
A precise task
Observe
Read the file, the data
Decide
Following its written rules
Act
Through a declared tool, in draft
Check
The result, then back to “Observe”
The task is done
The agent stops and reports
A sensitive action
Sending, paying, signing, deleting: a person approves first
Agentic AI refers to artificial intelligence systems, called agents, that do more than answer: they take a goal, break the work into steps, use your tools to act and check the result. Where an assistant drafts a reply, an agent finds the file concerned, prepares the reply and submits it to a person for approval.
This article shows two agents Idemys runs for itself, the grid we use to choose the first task to hand to an agent, and what we never hand to one.
How does an AI agent work?
An agent brings together a language model, which understands the request and chooses the next step, tools, which it calls to read or to act, and a working memory. It works in loops: it observes, acts, looks at the result, then carries on, corrects itself or stops. Its tools and rights set what it can do: an action it has no tool for is out of its reach.
AI agents, generative AI and RPA: what is the difference?
What it does | Its limit | |
|---|---|---|
Generative AI | Produces content from an instruction: a text, a summary, some code | Waits for the next instruction |
RPA (robotic process automation) | Replays clicks and entries defined in advance | Stops as soon as a screen or a format changes |
AI agent | Pursues a goal: chooses its steps and acts through tools | Does only what its tools and rights allow |
The last two often work side by side: the agent handles what takes understanding a request or a document, RPA keeps the repetitive steps already automated on stable screens.
Two agents Idemys runs for itself
We automated our own processes first, then those of our clients. Here are two of our agents, described by the card we keep for each one, without the names of our tools:
The administration office | The agent that builds this site | |
|---|---|---|
The task | Running Idemys’s administration: activity reports, invoicing, supplier invoices, contracts and their signature | Writing and changing the site’s pages, articles and code, from written rules |
The method | Each process is a chain: read, check, draft, act. Always study in read-only mode first, then plan, then carry out | The site’s content rules and its architecture decisions, read again before each change |
Its rights | Read-only at first; then, writing drafts only | It works on a separate branch of the repository and puts nothing online itself |
What a person approves | Everything that goes outside and every payment: sending, approving, paying, signing, deleting is a person’s click, in the tool | Every change to the main branch goes through a pull request, reviewed on the local site and merged by a person, who also starts the release |
What is logged | Every run, in a journal that is only ever added to; every piece of feedback becomes a written rule the agent reads again | Every change and its review, in the repository’s history; automatic checks run before every proposal |
What we changed along the way. The office first asked for a confirmation at every step; it now goes to the end of the process in draft, with no question on the way, and reports once, with the list of what awaits a decision. Sending, approving and paying remain a person’s click. For the site’s agent, the written content rules became automatic checks: the linter refuses an em dash or a proper name outside its line, and a test refuses a title too long for search results.
Choosing the first task to hand to an agent: the grid
Before building anything, we run each candidate task through this grid, with the person who does it today:
Criterion | The question | Good sign | Warning sign |
|---|---|---|---|
Frequency | Does the task come back often? | Every week, or every day | A few times a year |
Clear rules | Can the rules the person applies be written down? | They fit on a page | “It depends”, with nobody able to say on what |
Available data | Is the information in tools the agent can read? | In your tools, up to date | In people’s heads or scattered files |
Reversibility | Can a mistake be fixed before it has an effect? | The result stays a draft | The effect is immediate and final |
Risk | What happens if the agent gets it wrong? | Some time lost, caught at review | Harm to a person, money gone |
Who approves | Is someone named to approve what goes out? | Yes, with time to review | Nobody, or everybody |
A good first task has a good sign on every line. A warning sign on reversibility or risk is enough to postpone it; on the other lines, it says what to prepare first.
What we do not hand to an agent
An agent can prepare almost anything. Here is what we do not let it decide alone, even when the technology would allow it:
- A decision about a person. Hiring, assessing, refusing a service: the agent gathers the facts, a person decides. The AI Act also lists employment among the high-risk uses.
- Money going out or a signature. The agent prepares the payment or the contract in draft; the click that commits stays a person’s.
- An action with no way back. Deleting, overwriting, sending outside: not without approval, and not without a backup.
- Checking its own work. An agent can check another agent, but the last check of what commits the company belongs to a person.
- A task whose rules nobody can write down. We first write the rules down with the person who does it; the agent comes afterwards.
Staying in control
Least-privilege rights, a person’s approval for sensitive actions, a log of every action and a stop button at any time: the full list of guardrails, with how to check that each one holds, is in our article on securing AI agents. To connect an agent to your tools, see also our article on MCP servers.
Where to start
With a task chosen with the grid, a first agent under control, measured on what it saves, then the next step if the results justify it. Our pages Automate software development and Automate your processes describe the two kinds of use, the offer AI agents under control the guardrails, and the page Agentic AI our work. To talk it through: the free scoping day.
Frequently asked questions
What is an AI agent?
A program that relies on an AI model to reach a goal in several steps: it chooses the next action, uses tools to act, checks the result and stops when the task is done or when it needs approval.
Agentic AI and generative AI: what is the difference?
Generative AI produces content in response to an instruction. Agentic AI uses that kind of model to act in your tools and chain several steps towards a goal.
Which first task should go to an agent?
A frequent task, with clear rules, whose data is in your tools, whose result stays a draft that a named person reviews: preparing a file, a reconciliation, a first reply to a request.
Can an agent act without human approval?
Technically yes, and that is exactly what needs a frame. Reading, searching and preparing can be automatic; sending, paying, deleting or releasing to production wait for a person’s approval.
Sources
Written with an AI assistant from the sources cited, facts checked and reviewed by Mehdi Chafik on 26 October 2026.