The problem
Operations teams know agents could take work off the floor. The first project usually goes to the most interesting workflow, not the most countable one. Six months later nobody can say what it saved, and the second project never gets funded.
What goes wrong
- The agent is built before the workflow is written down, so it automates the exceptions along with the rule and your team stops trusting it.
- Nobody owns the exception queue, so misses never become rules and the agent stays at week-two accuracy for a year.
- The metric is never baselined, so there is no before number, no after number, and no case for the next workflow.
The approach
- List every workflow that eats hours each week, usually eight to fifteen, and score each for impact, effort, dependencies, and how soon the number would move.
- Score the company you run today. If the system of record is three spreadsheets, plan the build around three spreadsheets.
- Pick the boring winner: countable volume, stable rules, one approver, and a metric your finance team already tracks.
- Write the approval line before you build: money, customer contact, and load changes always wait for a named person.
When it works
It works when the work arrives as an email, PDF, form, or row the agent can read, the rules fit on one page, and one named person owns the approvals and the exception queue.
Where agents pay off
Operations is a stack of small decisions that repeat. A freight bill comes in. Someone checks it against the quote, flags the overage, asks the manager, posts it, and tells the carrier. It takes a few minutes each time, it repeats all week, and the person doing it is the one who should be on the phone with the shipper.
An agent pays off where three things line up. The work arrives in a form the agent can read: an email, a PDF, a form, a row in a system. The rules are stable enough to write down. And somewhere in the chain a person has to say yes before money moves or a customer hears from you. That describes most of operations.
The places it does not pay off are the ones everybody picks first. The judgment call only the owner can make. The task two people do two different ways. The workflow nobody has written down, because there is no process to automate yet. Fix the process first, then put an agent on it.
How to pick the first workflow
Make the list
Write down every workflow that eats hours each week. Freight audit on carrier invoices. Pickup reschedules. Chasing proof of delivery. Claims intake. Onboarding a new carrier. Expect eight to fifteen. Do not pick yet.
Score each one
Four questions per workflow. How much time or money does it cost today? How hard is it to connect to the systems it touches? What does it depend on that is not ready yet? How soon would the number move? Score honestly against the company you actually run, not the one in the sales deck. If the system of record is three spreadsheets, the agent will read three spreadsheets, and the build is planned around that.
Take the boring winner
The first workflow is rarely the most interesting one. It is the one with a countable volume, a clear rule set, one approver, and a metric your finance team already tracks. Carrier invoice matching beats "fix dispatch" every time, because you can baseline it in a week and show the difference in a month.
Decide what a person must approve
Before anything is built, write the approval line. Anything that moves money, contacts a customer, or changes a load goes to a named person with the reason attached. Everything else the agent proposes until the shadow run shows it is right on that step. Then that step can run alone, and every run is logged. This is the whole design. Get it wrong and you build either a bot that asks permission for everything or one that nobody trusts.
What it looks like when it works
It runs on your real data, inside your security perimeter, on the channel your team already uses. Named people use it without being reminded. One person is accountable when it misbehaves, and a human approval sits wherever a mistake costs real money or trust. The test we use is simple: if you turn it off on Friday, someone's next week gets worse.
When it does not work
It does not work if nobody owns the exception queue: the misses never become rules, and the agent stays at week-two accuracy. And it does not work without a before number, because then nobody can say what the first workflow saved.
Where to start
If you already know the workflow, the Workflow Agent build runs in four steps: map, classify, confirmed writes, switch-over. If you have the list but not the pick, ask Roo, Webaroo's AI: Which of our operations workflows should an AI agent take first?
Book 15 minutes with Connor