AI, agentic system, autonomous agent: three levels of automation for an owner

The brief arrives as “we need AI.” Under that sentence sit three different machines. Sometimes it is a chatbot that answers customers. Sometimes it is a system that assembles reports on a schedule. Sometimes it is a digital employee that is expected to decide without a person in the loop.

For an owner, the useful split is not brand names. It is level of automation: AI as a tool, an agentic system, an autonomous agent.

Level 1. AI as a tool

A program that answers, drafts, or analyses — and waits for a command.

Think of a very good calculator. The operator types the job, takes the result, and does the next step in the real world. There is no memory between sessions, and no action in the company’s systems.

Typical use: draft a letter in Claude or ChatGPT, then send it by hand. Generate an image, then place it. Drop a spreadsheet in and read the pattern, then decide.

The marker: the model works in isolation. It is not on the CRM, the inbox, or the warehouse. Every session starts blank.

The gain is speed on a manual step. The limit is that the person is still the bus between the model and reality. The invoice for this level is a subscription or token bill, not a project.

Level 2. An agentic system

A set of specialised agents, each doing one part of a process and handing the result to the next. A person designed the scenario. The fleet runs it.

Think of a factory line. Station one picks up the blank, station two drills, station three paints. Each station knows its operation. If the plan breaks, the line stops and calls the floor.

Typical use:

• Inbound. One agent reads mail and forms. The next classifies hot / warm / cold. The next writes the CRM and names an owner. A person reviews.
• Competitor watch. One agent walks competitor sites on a morning schedule. The next compares prices. The next files a report to Telegram or WhatsApp.
• Documents. One agent receives a supplier invoice. The next matches it to the order. The next prepares a payment for a person to confirm.

The marker: agents are wired into systems (mail, CRM, sites, messengers), they run on a schedule or a trigger, and they follow a written scenario. Off-script, they stop.

The gain is that a whole routine leaves the person’s hands, not just a paragraph. Add a competitor — the watcher can pick it up. The limit: an unfamiliar invoice format is a halt, not a creative improvisation. That halt is a feature.

Level 3. An autonomous agent

A program that is given a goal, chooses means, and acts without constant supervision. It may change approach when the current one fails.

The owner said: keep us competitive on price. The agent decides how often to check, which SKUs to watch, and when to raise a flag. A sharp competitor drop produces a counter-proposal, not a blank stare.

On paper this is a trading agent, an HR agent that searches and screens, a procurement agent that reorders. In practice a fully unsupervised loop is rare. What operators actually ship is semi-autonomous: the agent acts inside a fence, and a person confirms spend, delete, send, publish.

The marker: a goal and constraints, not only a script. Inside the fence it is free. Badly drawn fences produce expensive decisions. Early control is not optional.

One table

Which level to buy

Ask: what is eating the most hours?

People drafting similar text, scanning data, hunting facts — start with tools. That is a week, not a programme.

People copying between systems, sorting inbound, assembling reports, walking competitor prices — that is an agentic system. A process that can be written down can run on a trigger.

The wish that the system should set prices, run purchasing, or negotiate on its own — look at autonomy later. Autonomy is grown, not switched on.

Most firms get the useful work from level 2. It attacks a real cost (hours of routine) and stays inspectable. Autonomous agents are for operators who already automated the copying and are ready to delegate a class of decisions — with gates.

A catalog watch, as a fleet

A building-materials catalog: a little over a thousand SKUs, six active competitors. Purchasing used to walk sites by hand, type prices into a sheet, and decide on instinct. A full category took an hour or two. A thousand SKUs cannot be walked that way, so the team sampled, weekly.

What was built was not a chat window:

1. A crawler walks the operator’s own catalog and lifts every SKU.
2. One parser per competitor finds the analogue and the price. Two sites blocked naive bots, so those two run in a real browser.
3. An analyst agent fans out three views — market, risk, pricing — and a coordinator merges them.
4. A reporter writes an HTML brief and files history, so week-to-week movement is visible.

After the first full pass the picture was operational, not oracular: which share of the book was already cheapest, which SKUs were the expensive outliers, which competitor overlapped the catalog most heavily. The director still changes the price. The fleet does not. That is level 2: six live sites, a store of history, a Monday morning brief. Not an autonomous trader. Not a chatbot.

What to take away

• Tool — one request, one answer. The person is the integration.
• Agentic system — a scenario, wired to real systems, on a clock. A person still owns the exception.
• Autonomous agent — a goal, a fence, adaptation. Rare as a first project.

If the brief is fuzzy, start with an agentic system on one named workflow, with a stop switch. That is the middle that actually ships. Request a scope when the workflow and the failure you will not tolerate can be written down.