An assistant answers questions. An automation runs fixed steps. An agent takes several steps with judgement in between, using your systems. They cost more to build and run in that order, and in our experience most problems brought to us as agent problems are automation problems with one messy input.
Five questions
- Is the output an answer or an action? If a person reads the output and then does something, you need an assistant. If the system itself changes a record, you need automation or an agent.
- Can you draw the steps on a whiteboard, the same every time? If yes, it is automation. Rules are cheap, fast and predictable.
- Where is the mess? If the only hard part is reading the input, an email, a scanned invoice, a photo, add a model at that one step and keep the rest as rules.
- Does a competent person have to decide something between the steps, differently each time? That is the one case for an agent, and it needs a harness.
- What happens if it is wrong? If the answer is a refund or an apology, build in a human approval before you build anything else.
Three examples
Admissions enquiries on WhatsApp: an assistant. It answers from approved documents and hands over when it does not know.
Supplier invoices into the accounting system: automation, with document intelligence at the reading step and a queue for the ones that do not match a purchase order.
Customer quotation requests that need stock checks, a pricing decision and a drafted reply: an agent, with a person approving every quotation until the harness shows it can be trusted on the routine ones.
The honest default
Start with the cheapest system that solves the problem. We say this knowing that agents are what people ask for. An automation that runs a thousand times a day without a mistake is worth more than an agent that is impressive on Tuesday.