Think in actions, not pages

Most business software is organised in pages. There is one for customers, one for orders, one for invoices, and each page shows what a table holds. Getting something done means moving across them: open the customers, find the orders, filter what is due, create the invoices, send them one by one. The software holds the data. The path between the pages lives in someone’s head.

That path is where the time goes. A task that takes a few seconds to describe takes minutes of clicking – and only for the person who knows the way. When such a system is replaced, the new one is often built the same way, the old tables as nicer pages, and the path survives the migration. AI makes that kind of copying cheaper than ever, which makes it the least useful thing to do with it.

A different starting point is a single question: when someone opens the application, what is it they want to do? Not which data they want to see, but what they want to have done. Point A is where they are. Point B is the action completed. The interface gets them from one to the other.

Take an order management system at the end of a month. Someone wants to invoice the open orders. So they type exactly that: invoice open orders. A page opens with the invoices for the open orders of the current month. They switch to the previous month. They click invoice, and the invoices go out to the customers. Three or four words, a look, one adjustment, one click. Finding the action from those words needs no AI either: a fixed vocabulary and a few rules turn them into the action and its settings, and show how they were read.

That is automation – not the kind that runs out of sight. It is not a workflow nobody can control, and not human in the loop either, where a person approves what a machine has decided. It is human-driven action: the person decides, the system does the work.

None of it needs AI. Behind the action sits ordinary, deterministic logic – find the open orders, create the invoices, send them – chained together to complete one objective. What it takes is thinking about the system as a set of actions rather than a collection of pages, each built from steps that already exist. The same input gives the same result, every time, in seconds rather than minutes. The earlier note Owned, not rented described deterministic software as the bones of a business. This is what those bones look like to the person using them.

This is also why AI feels so powerful. Someone describes in a short sentence what they want to have achieved, and the model takes on the task and chains together the steps they would otherwise do by hand. It is the same idea, reached from the other side. But a lot can go wrong – a model mixing two steps that should never be mixed, for example, or doing the same task differently the next time. For the work that runs a business, a chain that is fixed and reliable is the better foundation, and the AI can simply execute it.

That is where the two meet. The chain of steps behind the action can be offered as a tool an agent can call, through an MCP server. The agent then has exactly the same controls as the person: the same action, the same settings, the same checks. Neither can do something the other cannot, and the software and the AI do not drift apart. And if an agent can execute the action, there is no reason a person should not do the same with three or four words – and keep more control over the result.

The pages stay. Every record can still be opened, every field checked. An order that needs correcting by hand is where they come in. The edge case gets the pages. The everyday case gets the action.

Designing this way moves most of the effort to the beginning: understanding what people actually do, which steps an outcome really needs, and which checks have to run every time. Anything that needs no screen does not need to become its own feature. What remains is a small number of actions that cover most of the work, and the pages for everything else.

Business software should start there. Not with more pages, not with more AI for its own sake – with a short sentence, a clear result and the same controls for everyone who uses it.