An AI agent is an analytical colleague that does not sleep: it gathers data, finds patterns and proposes. It cannot decide what you should want, handle a situation resembling nothing it has seen, or have the difficult conversation. The value lies in choosing which tasks belong where.
You hear it everywhere. AI agents will automate everything. Competitors are using them. Your staff are asking when you will.
The problem is that the word agent has stretched until it means almost nothing. To some it is a chatbot that got better. To others it is software acting on its own with no supervision.
Neither is right.
An AI agent is closest to an analytical colleague that works around the clock, does exactly what it was instructed, and never gets bored of a tedious task. It also has limits, and they are worth knowing before you build anything on top of it.
What it actually does
An agent can gather data from several sources, process it systematically against rules you defined, summarise what it means, and make proposals.
It cannot make the final decision. That is not purely a limitation of the technology. It is a choice you have to make deliberately, and it is the subject of this whole article.
The work falls into six roles
It helps to think of the work as roles rather than as one tool, because it makes clear where the human belongs.
Gathering. What is happening in the market, and what have we missed?
Analysis. What does it mean for us, and where is the risk?
Development. Which three to five options are worth considering?
Leadership. Which do we choose, and which risk do we accept?
Execution. Who does what by when, and what could go wrong?
Follow-up. Is it happening, and what do we do when something slips?
An agent can contribute to five of the six. The fourth it cannot, and that is deliberate. How the six work together in practice is covered in from chatbot to strategic department.
How much does it do on its own?
This is the decision that determines whether it succeeds, and it often never gets made explicitly.
Insight only. The agent analyses and presents. You read and decide. The model is a statistician, not an adviser.
Proposal with approval. The agent proposes a concrete action, and a human says yes or no before anything happens. In practice the approval takes thirty seconds.
Full automation. The agent acts within rules you set, and reports afterwards.
Most start in the first and move to the second once they have seen enough to trust it. The third belongs to routine work where the rules will still make sense in two years. How to choose the level task by task is in autonomy levels.
The four things it cannot do
It cannot decide what you should want. An agent can say that if you do A, B probably follows. It cannot say you ought to do A. That question is about values, risk appetite and what you want to become, and it is not an analytical problem.
It cannot handle what it has never seen. If regulation shifts, a crisis lands, or a competitor does something unexpected, the agent still follows the rules it was given. Those rules were written for the world that was. Those are exactly the moments a human has to think, and also where most people discover they had trusted the system too far.
It cannot carry the relationship. An agent can prepare you thoroughly for a difficult conversation with an important customer. It cannot have the conversation, and attempting to let it is obvious to the other party.
It cannot judge whether something should be done. It can find the route that earns most. Whether that route is acceptable is not a question it was asked, and it does not get asked by itself.
An example
A company had over two hundred articles on its site and did not know which ones produced sales. Doing it by hand would take about a working week, which is why it never happened.
They ran it as an analysis at the first level: what gets written, what draws visitors, and what leads to an enquiry. The result was a report with a dozen concrete findings.
The marketing lead read it and made the call: write more of what works, drop one topic entirely. That is the part a human has to do, because it is about what the company wants to be known for and not only about what gets clicked.
The agent then proposed a clean-up of the internal link structure, and got a yes with one exception the lead could see was wrong.
The division of labour is the whole point. The agent did the thorough work nobody else was getting to. The human made the two choices that had consequences.
How to start
- Begin by reporting, not by acting. Choose something important but routine: what customer feedback says, how contribution margin has moved.
- Read the reports for a month. Do they change any decision? If not, it is the wrong task, not the wrong tool.
- Move to proposal with approval once you can predict what the report will say before opening it.
- Use full automation sparingly. Only where the rules are stable and a mistake is easy to spot and undo.
Good tasks: data-heavy passes, recurring analyses, gathering market information, following up on something falling between two people.
Bad tasks: what should we bet on, how do we handle this relationship, and is this acceptable to do.
Why the order is built in on our side
An agent working in isolation produces an analysis that does not know what the others found.
In 360° Sprint the connections between nodes on the board determine the order the analysis runs in, and each node receives its context from the ones upstream through the knowledge graph. The capability analysis therefore knows the market analysis's conclusion, because it came earlier in the chain.
Fields classified RESTRICTED are never sent to any provider, regardless of who is working in the system. What that means in practice is in the four classification levels.
An agent is like a strong analyst who never sleeps. A strong analyst does not run the company, and you should not ask it to.
The whole picture, from what suits it to a week you can run yourself, is in AI for business strategy.