Every software company in the country relabelled something an agent this year. Most of what got relabelled is a chatbot with a longer memory, and the gap between that and the thing worth paying for is the whole subject of this post.
The distinction is not intelligence. It is permission and reach
A chatbot answers when spoken to. It has no standing instructions, no access to anything except what you paste into it, and no ability to change the state of your business. Ask it to chase your overdue invoices and the honest answer is that it cannot: it does not know who owes you money and could not send an email if it did.
An agent, used properly, has three things a chatbot does not. It has a job it does without being asked. It has read access to the records that job depends on. And it has a written boundary around what it may do unattended versus what it must put in front of a person.
If it cannot act, it is not an agent. If it can act without limits, it is a liability.
Why reach is the part everybody underestimates
The most common disappointment we see is an agent bolted onto one tool. It can read the inbox but not the job system, or the accounts package but not the diary. Ask it anything that crosses those boundaries and it either guesses or apologises.
That matters because almost every question worth asking a business crosses a boundary. "Which work are we winning, and is it work worth having?" needs sales, delivery and finance in the same sentence. An agent that owns a third of the picture gives you a third of an answer, confidently, which is worse than none.
The boundary is a business decision, not a technical one
This is the part suppliers rush and owners should not. For every job an agent does, somebody has to decide whether it acts alone or asks first. Acknowledging an enquiry: act alone, because being slow costs you the job and being wrong costs nothing. Sending a price with your name on it: ask, every time.
Get that line wrong in one direction and you have built an expensive notification system nobody reads. Get it wrong in the other and something goes out to a customer that should not have. The line should be written down in plain words, visible on the screen, and moveable by you rather than by whoever built it.
Three questions worth asking a supplier
- What can this do without a person, and where is that written down where my team can read it?
- Which of my systems can it actually see, and what happens when I ask it something that spans two of them?
- How will I know if it silently stops working? An automation that fails quietly looks exactly like one that had nothing to do.
The third one catches more suppliers out than the other two combined, and it is the one that costs you most when nobody asked it.