The difference between asking a chatbot and running agents is not the model. It is that the work is divided, that the steps run in an order that makes sense, and that each step knows the conclusions of the one before. The chatbot starts over every time. The chain does not.
You have just built a new strategy on the whiteboard. Everyone nods. Then comes the question that always comes: who makes sure it happens, and who watches whether it goes to plan?
The answer is usually that somebody hopefully remembers, that there is a weekly status meeting, or that a review happens once a quarter and arrives too late anyway.
The alternative is not one more manager or one more meeting. It is that the work is split into roles that each know their job, run in a sensible order, and do not need permission at every step.
The work divides into six roles
This is a way of organising the work, not a menu of tools. The value is in the division, whether the roles are carried out by people, by agents, or by both.
Gathering collects the context. What is happening in the market, what have we missed, what do the sources say? It produces a picture of the situation before anyone starts formulating a strategy.
Analysis interprets what was gathered. What does it mean for us, and where is the risk? It turns raw material into something management can respond to.
Development designs the options. Given what we now know, which three to five routes are worth considering?
Leadership prioritises. Which route do we take, and which risk can we live with? This is where the human belongs, and it is the one step of the six that cannot be delegated.
Execution plans. Who does what by when, with which resources, and what could go wrong?
Follow-up keeps track that it happens. A meeting was cancelled, a measurement was not taken, a deadline slipped. Somebody has to register that without it requiring a meeting.
The point is not that each role is perfect. It is that each knows its job, and that they do not work on top of each other.
The order is what makes the difference
On the surface it looks as though things simply get filled in. Underneath, something more precise is happening.
Each step depends on other steps. Prioritisation cannot happen before the analysis has said what the numbers mean. The execution plan cannot be made before somebody has chosen the direction. The order is not a preference, it is a logical necessity.
That principle is called topological sorting. In short: first what depends on nothing, then what can be built on it, then the layer after. Like building a house, where the foundation comes before the walls.
It also means the system knows what can run in parallel. The competitor analysis and the capability analysis do not depend on each other and can run at the same time. Prioritisation depends on both and waits.
That is the mechanical explanation for why one coherent story comes out instead of a stack of loose notes. The full chain from overview to execution is described in the strategy chain.
Set limits, or it gets expensive
This is where most people get nervous, and rightly. Agents running without limits can run up a cost.
So the limits are the first thing to settle. A ceiling per run, per week, or per role. The frame should be agreed before anything starts, not discovered on an invoice.
It is not about forbidding anything. It is the same principle as giving a marketing lead a budget: they run the detail, but the frame is fixed, and once it is spent, somebody has to take a position again.
If an analysis is too expensive, the answer is usually not a bigger frame. It is a more precise task. "Read the twenty most important sources" produces better results than "read everything you can find", at a fraction of the cost.
One process from question to plan
Say you have to decide whether to be present on a new channel.
Monday morning. The question is framed as a task: should we use this channel, and for what?
Gathering starts immediately. Who sells something like ours there? What do we know about the users? What have competitors tried, and what happened?
Analysis waits until gathering finishes, because it needs the result. Then it reads it and writes the conclusion: the users are markedly younger than our buyers, two competitors tried and stopped after a quarter, and the channel that actually produced enquiries for them was a different one.
Development formulates three routes: try it for three months on a small budget, buy the capability externally, or leave it and revisit in a year.
And here it stops. The choice requires somebody who knows the company from the inside, knows what else is competing for attention, and can carry the consequence. It is not an analytical question.
Execution starts once the choice is made: who does it, when, what we measure, and when we look at the numbers.
Follow-up registers that it is happening, and says so when it is not.
The whole thing takes less time than the meeting it replaces, and the hard part is still the hard part.
Why this beats asking a chatbot
When a manager today has a strategic question, they write a couple of sentences to a chatbot, get an answer, assess it, and usually forget it again.
Four things go wrong:
- It does not know what you found last month. Every conversation starts over.
- It does not learn your company. The context is the same every time.
- It has to be remembered. If nobody asks, nothing happens.
- It cannot prioritise. Every question weighs the same, because it does not know what matters most to you.
A structured chain solves the first three by storing the conclusions and passing them to the next step. How knowledge from one analysis becomes the basis for the next is covered in knowledge graph. The fourth you solve yourself, by deciding what matters most.
Why the context is built into the order
The weakest link in setups like this is that each step works in isolation and produces something that does not know what its neighbour 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. Progress is shown as it runs, and results collect in Insights with run history.
What the agents can and cannot do is covered in what AI agents can actually do, and how to decide how much they do alone is in autonomy levels.
None of this means you stop meeting. It means that when you meet, you have something to discuss instead of spending the first half hour rebuilding the basis from scratch.
The overview of when AI makes sense in strategy work at all is in AI for business strategy.