AI agents: where they take work off your team
AI agents are good at inbox triage, document intake and moving data between systems. They are bad at judgement. How to find the first one worth building.
Monday morning, half past eight. The shared inbox has 140 new messages since Friday. Someone on the team opens each one, decides whether it’s an order, a complaint, an invoice or spam, forwards it to the right person and copies a few fields into the ERP. By eleven the inbox is clean. By Tuesday it’s full again.
Nobody hired that person to sort mail. It just became part of the job, and in most companies with 20 to 300 employees there are several jobs like it. This is where AI agents earn their keep. It’s also where a lot of agent projects go wrong, because the word “agent” gets used for things it can’t do.
What an AI agent actually is
Strip away the marketing and an AI agent is a program that reads something, decides what it is, and takes an action in another system. It might read an email and create a ticket. It might read a PDF and fill in a form. It might look up a customer and write a draft reply.
The decision in the middle is what makes it different from a classic automation. A rule-based script breaks when the supplier changes their invoice layout. An agent reads the new layout and still finds the invoice number. That tolerance for messy input is the whole point.
Where agents take real work off a team
In my experience, four kinds of work are reliable candidates.
Inbox triage. Sorting incoming mail by type, urgency and recipient. The agent doesn’t need to answer anything. It reads, labels, routes, and pulls out the order number or customer ID so the person who opens it next doesn’t have to search.
Document intake. Invoices, delivery notes, contracts, applications, claim forms. Someone currently opens them, reads them and types the important parts into another system. An agent can read the document, extract the fields and hand over a structured record. A person checks the ones the agent is unsure about.
Data between systems. The CRM says one thing, the ERP says another, and every week someone exports a spreadsheet to reconcile them. Agents are good at this kind of glue work, especially when the data isn’t perfectly clean and a fixed mapping keeps failing.
First drafts of answers. Customer questions that come up again and again: delivery times, return conditions, how to reset an account. The agent writes a draft based on your actual rules, and a person reads it, fixes what’s off and sends it. The person stays in charge. They just stop typing the same paragraph twenty times a day.
What these have in common: high volume, a clear right answer most of the time, and a cost that someone can point to in hours per week.
Where agents don’t help
Judgement calls. Should we give this long-standing customer a discount even though the policy says no? Is this complaint a sign of a bigger quality problem? An agent can prepare the facts. The decision belongs to a person, and it should stay there.
Work without rules. If three people on your team would handle the same case three different ways, an agent will pick a fourth. Before you automate a process, someone has to be able to write down how it’s supposed to work. If nobody can, the problem isn’t missing AI. It’s a missing decision.
Rare, expensive exceptions. Something that happens twice a year and goes badly wrong when it’s mishandled is a poor first target. You won’t have enough cases to test the agent properly, and a mistake costs more than the time saved.
Things that are already fast. If a task takes one person ten minutes a week, an agent won’t change anything that matters.
Why the first agent often disappoints
A typical pattern: a company starts with the most ambitious idea on the list. “An agent that handles all customer service.” Six weeks in, it answers the easy questions well and the hard ones confidently and wrong. Trust drops, and the project quietly stops.
The agents that last start narrow. One inbox, one document type, one direction of data. They run next to the person who does the work today, so you can compare results on real cases before the agent touches anything on its own. Only when it’s reliably right do you let it act without a check, and even then you keep a log of what it did and why.
How to find the first agent worth building
You don’t need a strategy workshop for this. You need about an hour with the people who do the work.
- List the repetitive tasks. Ask the team what they do every day that feels like copying, sorting or retyping. Write down everything, without filtering.
- Put a number on each. Roughly how many times a week, how many minutes each time. This alone usually reorders the list.
- Check for rules. For each task, could someone explain in a few sentences how to decide what to do? If not, cross it off for now.
- Check the systems. Where does the input come from and where does the output go? Mail, a shared folder, an ERP with an API? The answer decides how hard the build is. I’ve written more about that in connecting AI agents to existing systems.
- Pick the boring one. The best first agent is usually the task with high volume, clear rules and an easy way to check the result. It’s rarely the most exciting idea on the list.
Then build that one, run it on real cases for a few weeks, and measure the hours it takes off the team. That number tells you whether the second agent is worth it.
If you’d rather have someone walk through the list with you, that’s what my AI automation work starts with: one process, measured before and after.