What are AI agents? And what do they do at the management table?
An AI agent is software that independently watches one question: it reads your sources, compares what is happening with what should be happening, and reports what deviates, with the evidence attached. A chatbot waits until you ask something, a dashboard shows a number, an agent comes to you on its own with a signal and a prepared decision. The deciding is yours.
What exactly is an AI agent?
An AI agent is a piece of software with an assignment, access to sources and a rhythm. The assignment is narrow: watch the margin on projects, watch the payment terms, watch capacity utilization. The sources are your own systems: ERP, CRM, time tracking, accounting, contracts and folders of documents. You set the rhythm: every night, every Monday morning, or just before the leadership team meeting.
Within those boundaries it works independently and does not ask for instructions every time. That is the difference with virtually all the AI you have seen so far.
What is the difference with a chatbot or a dashboard?
Three tools, three roles. They do not rule each other out, but they solve different problems.
| Chatbot | Dashboard | AI agent | |
|---|---|---|---|
| Starts with | Your question | Your click | Its own assignment |
| Looks at | What you supply | Preselected numbers | All connected sources |
| Delivers | An answer | A status | A signal with evidence |
| Notices a deviation | No | Only when you look | Even when nobody looks |
| Prepares a decision | No | No | Yes, with scenarios |
A dashboard tells you margin is 16,2%. An agent tells you margin is 16,2%, that this is 2,4 points below your budget, that 71% of that gap sits in two customer groups, and that you have three levers to pull. The first is information. The second is a prepared decision.
What does an agent read, and what does it do with that?
Most agents do three things, in this order.
- Read. Orders, invoices, timesheets, quotes, contracts, stock levels, payment behavior, service agreements. Also what is not in tables: an extra work order in a PDF, a clause in the fine print of a framework contract.
- Flag. The agent compares reality with the standard you have recorded: the estimate, the budget, the contract, last quarter. Anything outside the range becomes an alert.
- Prepare. Every alert states where the number comes from, how large the effect is, who owns it and which options you have. Including the option to do nothing.
So what an agent delivers is not another screen. It is a shorter road from signal to decision.
Example: the margin agent that found € 54.000 in extra work
Take a technical services firm with 140 employees and about 60 active projects. Margin was under pressure, but nobody could point to where. The monthly report arrived on working day nine and showed the total, not the cause.
The margin agent got one assignment: for each project, put the logged hours and purchased materials next to the estimate, and next to the invoices sent. It read the time tracking, the purchase invoices, the project administration and the extra work orders in the project folders.
Within two weeks there was an outcome. On three projects, 62% of the recorded extra work had never been invoiced. That came to € 54.000. No fraud, no ill will: the extra work was agreed on site, written up on an order form and then never passed on to the project administration. Two project managers did pass it on, three did not.
The problem was never that the information was missing. The problem was that nobody put the three sources side by side at the same moment.
The executive team decided two things. The outstanding extra work was invoiced after all, in consultation with the customers. And the agent kept running: every Monday a list of extra work older than fourteen days that is not yet on an invoice. Across the whole portfolio the leadership team estimated the annual value at € 180.000. That amount is an expectation, not a promise. Only what is actually invoiced counts in the evaluation.
What does an AI agent not do?
This is the most important part, and sales pitches usually skip it.
- It does not decide. The agent has no mandate, no budget and no vote. It does the groundwork for a decision that a person takes and signs.
- It does not know your context by itself. That one customer is strategic, or that a project manager just came back from sick leave, is something it knows only if you record it.
- It does not repair your data. If hours get logged three weeks late, the agent sees it and reports it. Fixing it remains a management task.
- It does not replace your craft. Interpreting a signal takes experience. The agent makes sure you spend that experience on the conversation, not on the digging.
Hoe je die grens hard maakt, lees je in AI at the table, people at the helm.
How do you start with one agent?
Not with a platform. With a question that is costing you money or time right now.
- Pick one question whose answer currently arrives too late. For example: which projects are running over the estimate?
- Write down what the standard is. Without a standard there is no deviation, and so no signal.
- Name the sources and the owner of each source. Usually there are two to four.
- Record the baseline measurement: how long it takes today before you know this, and what that costs.
- Laat de agent zes weken meedraaien en beslis daarna op feiten. Die aanpak staat in the 45-day pilot.
Eén agent is een hulpmiddel. Vanaf een stuk of vijf ontstaat iets anders: een managementlaag over je hele bedrijf, beschreven in Digital Business Twin. Waarom dit iets anders is dan de rapportage die je al hebt, staat in From Business Intelligence to Management Intelligence. En wat je ermee wilt beantwoorden, vind je in Six questions every leadership team should be able to answer.
Frequently asked questions
- What is the difference between an AI agent and ChatGPT?
- ChatGPT is a chatbot: it waits for your question and works with what you give it. An AI agent has a fixed assignment, fixed access to your business sources and a rhythm of its own. It comes to you with a signal on its own, even if nobody asks a question that day.
- Do my systems have to be in order first?
- No. An agent reads your sources as they are today, without you changing your ERP or CRM. Poor data does lead to weaker signals, so in the first weeks you measure how complete and current your sources are. Often that is a valuable outcome in its own right.
- How much time does an AI agent cost my team?
- During setup, count on two to three hours a week from one closely involved employee, plus two one-hour sessions with the leadership team. After that, watching the signals takes less time than the digging it replaces.
- Can an AI agent send an invoice or place an order by itself?
- Technically it can, but we do not do that without explicit approval. Flagging, advice and decision stay separate. The agent prepares the invoice with source and evidence, a person reviews it and sends it.