What an AI agent can and cannot do

An AI agent is a program with a task, a short list of actions it is allowed to perform, and a model that picks which action is next at each step. It reads something, picks a step, carries it out, looks at the result and continues until it is finished or stuck.

That is less magical than the word suggests, and less unpredictable too. What an agent can do is exactly what is on that list. If sending is not on it, it sends nothing, however convinced it is that it should. That list is written down before anything runs.

Volume is the clearest example. Five hundred emails, two hundred PDFs, a folder of work orders going back six months: an agent reads the last one with the same attention as the first. People do not, and that is not a criticism. It is why things slip through at four in the afternoon that would have been caught at nine in the morning.

It is also useful for a first version: an answer to a recurring question, the outline of a quote, the summary of a call. Not the finished piece, but the version somebody goes over. In written work most of the time goes into getting past the empty screen, and here that screen already has something on it.

And it turns mess into structure. Out of a PDF that every supplier lays out differently it pulls the fields you need: number, date, amount, job. That includes the repeated decision whose rule can be written down, such as: this invoice belongs to that job, because the number matches and the amount falls inside the margin.

Judgment. Not the hard cases, but the cases whose rule is written down nowhere. Whether you cut this client some slack this time depends on what happened last year, on what the account manager knows and on how next quarter looks. A model that sees only the document sees none of that.

Thin context is the quiet version of the same problem. The agent reads the email, not Tuesday's phone call. It reads the system, not the arrangement made over coffee. As long as half the truth lives outside your systems, an answer that sounds complete is still based on half of it.

And then the nastiest one: a model that does not know sounds exactly like a model that does. Confidence in the tone is not a signal. So we build the doubt in on purpose. If something deviates from what the system expects, it puts the case forward with the reason attached, rather than offering a guess as an answer.

Responsibility cannot be handed to software. If a wrong invoice goes out, there is a person who explains that to the client and who issues the credit note.

Anything that goes out in your name or moves money waits for a person to approve it. On most routes that means reading through a list of thirty lines once and clicking thirty times, rather than redoing the same work thirty times.

That gate does something else as well: it makes mistakes cheap. Clicking away a wrong draft costs seconds, a wrong email at a client costs a conversation and sometimes the client.

If something breaks, the system falls back on the manual route and the owner of that route gets a message, so the work carries on by hand that day.

Work that runs differently every time is the first category. If you cannot write the steps down without adding that it depends at every turn, there is no route to automate yet. Sometimes that is a reason to tidy the process up first, and that is a perfectly good outcome of a conversation.

The second is work where the contact itself is the value. A difficult phone call with a client who has just found a mistake of yours is not expensive because it takes time.

And then there is work that comes back once a quarter. Saving ten minutes four times a year does not weigh up against a system you have to keep maintaining while the process around it changes.

An agent that is set up well gives hours back. Not a department fewer, but the pile that used to be there every morning and is not there now. The systems we publish carry that figure.

And it can be walked back. Every run leaves a trail: the message that came in, the steps the system took, the answer that went out, with the date and time attached.