Where should my company start with AI agents?
The short answer
Start with one task, not a company-wide program. Pick work that repeats often, whose information is already in order, whose result a person can judge in a minute, and whose mistakes are cheap to correct. Let the agent prepare and a person approve. Widen its work only after it has earned your trust.
One task, not a program
A company-wide AI program sounds bold. In practice it spreads attention across many tasks at once, and when results disappoint, nobody can tell which part failed. One task, done well, teaches you more than ten started together.
Four tests for the first task
- It repeats often. A task done every day pays back the effort quickly. A task done twice a year does not.
- Its information is already in order. An agent is only as good as what it knows about your business. If the documents it needs are scattered or out of date, fix that first.
- A person can judge the result in a minute. If checking the agent’s work takes as long as doing it, you have saved nothing.
- Its mistakes are cheap to correct. A wrong draft costs a minute. A wrong payment, a wrong price sent to a client or a deleted record costs much more.
A good first task, for example: preparing a first draft of replies to routine client questions, from documents that already exist, for a person to review and send. A poor first task: anything that pays, changes prices or speaks to clients with no person in between.
Keep it as simple as the task allows
Often the first step is not an agent at all. It may be a report, or a single well-written instruction to a model that answers questions. Add more only when the simple version is not enough.
My own system follows the same rule. The agents read and prepare. They never send, never delete and never change a record. I act.
The question to ask your team
“Which task would we be glad to hand over, and could a person check the result in a minute?” Start there.