What is a good first agent workflow for a business?

The first one is chosen for what it teaches rather than for what it saves, and the two are not in conflict if the choice is right. Frequency matters because a daily task gives you thirty observations in a month and a monthly one gives you one. Read-only matters because it removes the permission question entirely, so the project is not blocked on a decision nobody wants to make. And producing something wanted matters because a workflow whose output nobody reads is not being observed, which defeats the purpose of the exercise.

The shape that satisfies all three in most businesses is a daily summary of what arrived. What came in, from where, in what category, what is outstanding, what looks unusual. It reads records the business already has, writes nothing, produces something a person actually opens, and its errors are visible immediately to whoever knows the business. It also produces the enquiry data that most small businesses lack, which is the input to every subsequent decision about what to automate.

The second candidate is preparation. Assembling what someone needs before they do something: the customer history before a call, the outstanding items before a review, the relevant material before a meeting. This has the same properties and a stronger immediate saving, because it removes a search that currently happens under time pressure.

The third is a checking workflow, which is under-used. Comparing two sources that should agree, finding records missing a required field, identifying items that have been outstanding too long. These produce a short list of things to look at, they are checkable because each item can be confirmed, and the value is realised the first time the list contains something nobody had noticed.

What to avoid first is anything customer-facing, anything that writes, and anything infrequent. The first two raise the permission and disclosure questions before you know how the system behaves, and the third gives you too few observations to know anything. This ordering is not caution for its own sake — it is that the information you need to make the next decision comes from watching, and watching requires volume and low stakes.

The transition to acting should follow from what the observation showed rather than from a schedule. When you can predict what the report will say before reading it, the model of its behaviour is accurate enough to extend, and extending should be narrow: one action, in one place, reversible. That sequence takes weeks rather than months and it produces a system somebody actually understands.