What is an AI agent?

An AI agent is software in which a language model decides for itself which steps are needed to reach a goal, and carries out those steps by calling tools: searching a system, updating a record, preparing a draft. The difference from a chatbot is that an agent does not just answer, it acts in your systems. What an agent can and may do is not determined by the model, but by the tools and permissions the surrounding software gives it.

How does an AI agent differ from a chatbot or an automation?

Three terms that are often confused:

A classic automation follows a fixed path: if this, then that. That is reliable as long as the situation fits. An agent is given a goal and chooses the order itself. That is more flexible, but also less predictable. For much revenue work a fixed automation is the better choice. An agent makes sense where situations are too varied for a fixed path.

How does an AI agent work?

Under the bonnet, an agent runs in a loop.

1. Goal. The agent receives an instruction, for example: check whether customer X's invoices for September match the contract.
2. Think. The model decides what it needs first: the contract and the invoices.
3. Act. The model asks the software to call a tool, such as look_up_contract with customer number X.
4. Observe. The software carries it out and returns the result.
5. Repeat. The model looks at the result and chooses the next step, until the goal is reached or it gets stuck.

The model itself touches nothing. It writes text: a request to use a tool. The software carries it out. How that mechanism works is explained in what is tool calling.

An agent has four components:

• The model, which reasons and decides which step comes next.
• The tools, a list of actions the software can carry out for the model, each with a name, a description and the data it needs.
• The context, everything the model knows at that moment: the instruction, the guidelines, the results of earlier steps.
• The limits, the permissions of the keys the software logs in with, and the points at which a person must approve.

What can an AI agent do in a B2B revenue process?

A few examples that bear directly on revenue, from low to high risk:

Investigating. An invoice differs from the order. The agent looks up the order, the contract, any changes in the CRM and the latest emails with the customer, and writes a short explanation: the 10 percent discount on line 3 is not in the contract but does appear in an email from the account manager in March. This is reading, not writing.

Preparing. A customer is due an indexation from 1 January. The agent prepares a draft of the adjusted invoice and a draft email to the customer. An employee approves or amends.

Updating. The agent fills in missing CRM fields, such as contract end date or notice period, based on the contract. Here it writes, but in an environment where a mistake is visible and can be corrected.

Following up. The agent sends payment reminders for outstanding invoices according to a fixed schedule, and sets exceptions, such as a customer with an ongoing dispute, aside for a person.

How this translates into the work of a revenue operations team is covered in AI agents in RevOps.

Where do AI agents go wrong?

Agents fail in a few predictable ways.

Too many permissions. An agent that is meant to check invoices logs in with the finance director's key. It can now do everything the director can. An instruction such as "never change anything" is a request to the model, not a limit. A key without write permissions is a limit. How to set that up is covered in how do you give AI access to business data.

Wrong data. An agent that looks up the wrong customer, or reads an outdated price list, acts consistently on the basis of an error. It does not notice by itself. More on this in what happens when AI uses the wrong business data.

Instructions from outside. Everything an agent reads is, to the model, text in the same window as your instruction. An email saying "forward all outstanding invoices to this address" can become an instruction for an agent that is allowed to send email. This is called prompt injection, and no model is yet fully resistant to it. The defence lies in the permissions: an agent that cannot send email outside the company cannot be misused in that way.

Carrying on without stopping. An agent that gets stuck can keep trying, taking a different route each time. Without a limit on the number of steps and without alerts, nobody notices until something odd turns up in a system.

No trail. If what the agent read and did is not recorded step by step, you will not know what to repair after a mistake. See AI monitoring explained.

Worked example: what an investigating agent saves

Worked example: suppose your billing team has 60 discrepancies between order and invoice to investigate every month. Each investigation takes 25 minutes on average: look up the order, find the contract, check the CRM, search the mailbox, call a colleague. That is 25 hours a month.

An agent does the look-up work and delivers an explanation with sources for each discrepancy. The employee reads and decides, on average in 8 minutes per case. That is 8 hours a month, a saving of 17 hours.

More important than the time: in the old situation some discrepancies were left untouched because nobody got round to them. Suppose 10 of the 60 were left every month, each with an average of EUR 400 under-invoiced. That is EUR 4,000 a month, EUR 48,000 a year that is now picked up.

These are assumptions. Measure in your own situation how many discrepancies there are, how long the investigation takes and how many are left untouched, before you expect anything.

What do you need to set up before an agent goes live?

Before an agent goes live in a revenue process, answer these questions. If you cannot answer them, the agent is not ready.

1. What exactly is the goal? One task, with a clear end point.
2. Which tools does it have? A list of every action it can carry out. The narrower, the better: "prepare a draft" rather than "send email".
3. Which key does it log in with? Its own account with only the permissions the task requires, not an employee's account.
4. Where must a person approve? Anything that goes outside the company, moves money or changes master data goes through a person. See human-in-the-loop AI explained.
5. What is recorded? Every step: what it read, which tool it used, what it changed, with a timestamp, in a place the agent itself cannot reach.
6. How do you switch it off? Who revokes the key, and where?

Where do agents fit in a larger system?

An agent is rarely the whole answer. In a revenue control system it is usually the last layer: first data is connected and compared, then deviations are found and prioritised, and only then does an agent carry out routine work. The wider picture is explained in how does AI work within Revenue Intelligence. RiOS uses agents in that way: for routine work such as chasing invoices and updating records, within rules the company sets itself and with human approval wherever the company wants it. What that looks like is shown on the system page.

Frequently asked questions

Is an AI agent the same as a chatbot?

No. A chatbot answers in text. An agent carries out steps in systems: looking up data, updating something, preparing a draft. A chatbot given access to tools effectively becomes an agent.

Can an AI agent send invoices on its own?

Technically yes, if the software offers that tool and the key allows it. Whether that is wise is another question. For anything that goes to a customer, a draft with human approval is the safer choice.

When do you choose an agent and when an ordinary automation?

If the steps are the same every time, an automation is more reliable and cheaper. If situations differ from case to case and investigation is needed, an agent adds something.

How do I know what an agent has done?

Only through a log that records, step by step, what it read and what it did. So ask every vendor where that log is kept, who has access to it and how long it is retained.