“AI agent” currently covers everything from a chatbot with a calendar plugin to a system that runs a purchase order end to end with nobody watching. Somewhere under the word there is a real and useful pattern: software that reads, decides and acts across several steps, with a model making the judgment calls a regular expression can’t.
We put these into ordinary companies rather than labs, which mostly means we see them again a year later, either still running or quietly switched off.
The ones that survive
The agents still in production after a year look alike. The task repeats. The inputs are messy. The list of things the agent is allowed to do is short. And a person signs off on anything that leaves the building.
Inbox triage is the clearest case. Mail arrives, gets classified, gets enriched with whatever the last several tickets from that customer said, and lands in front of a human as a draft. The human still decides. The typing is gone.
Document intake is the same shape. Invoices and delivery notes read, fields extracted, pushed into the ERP, and anything the model is unsure about routed to a person instead of guessed at. That last clause is the entire trick. An extraction agent without a confidence threshold isn’t an agent, it’s a random number generator with good manners.
Data hygiene is the least exciting item on the list and probably the best return on it. Duplicate records merged, addresses normalized, missing fields flagged for someone to chase. It’s the job nobody defends in a meeting and everybody quietly needs done.
The ones that fail, and how
The failures repeat just as reliably.
Give an agent broad tool access and a vague goal and you get a spectacular demo followed by production behaviour nobody can audit. When it does something strange in week six, there’s no log that explains why, because the reasoning happened inside a model and the model doesn’t keep one.
Let it take irreversible actions on its own, sending, paying, deleting, and it will eventually take a wrong one. A single wrong send to a customer list costs more than a year of the typing you saved.
Then there’s the quiet failure, which is economic rather than technical. Somebody bolts a model onto a task a plain script already handled. The inputs are structured, the rules haven’t changed since 2018, and now there’s an API bill and a nondeterministic step in the middle of a process that used to be boring and correct. A model earns its keep where judgment is genuinely required, and nowhere else.
Start with the queue, not the model
Every agent project of ours that worked started the same way. Find the queue a person empties every day. Sit with them while they empty it. Automate the reading and the drafting, leave the deciding alone.
Pick the model last. It’s the least important decision in the stack, it will be obsolete long before the project is, and swapping it is cheap as long as you put the fence in code rather than in the prompt.
Here’s the version that costs us work. Most companies who ask us for an agent want a form, a rule and a report, and would be better served by plain automation with no model in it at all. We say so, on the call, before anyone has signed anything. Some of those conversations end there. We’d rather lose that project than maintain a language model doing arithmetic.
If you do have a queue like the ones above, finding out what an agent would save is usually a small piece of work. Describe it to us, rough is fine, or start with a single consulting session and we’ll give you an honest read on whether it’s agent-shaped at all.