Three questions decide whether an AI agent can handle a job
When someone on your team says an AI agent could handle a job, three questions decide whether it is worth trying: can the steps be written down, does the information live in your systems, and is a mistake cheap to catch and fix?
In plain terms, an agent is software that carries out a multi-step job across your tools. Chasing unpaid invoices is a good example. Find the invoices past 30 days, look up the contact, draft a reminder in your tone, send it, log it, and flag the ones that need a phone call.
Each question rules out a lot of ideas.
If the steps live in your office manager's head, the agent has to guess, and guesses on your customers are expensive. Writing the steps down is the first project, and it often shows where the real problem is.
If the customer details sit in a spreadsheet one person updates on Fridays, the agent will act confidently on old information.
If a mistake means a wrong quote to your biggest customer, someone has to approve every action, and the time saved shrinks fast.
Then measure before anyone sends a proposal. Have whoever does the job log the time and the misses for a week. That week tells you what the agent would have to beat.