Owners keep asking me whether they should let an AI agent into the books. My answer is usually yes, with rules. The tool can be useful. Handing it the run of your accounts without deciding what it can see and what it can do is how useful turns into expensive.
Elena owns a medical billing company with about eighteen people. Last spring she set up an AI agent to pull bank activity, match deposits to claims, draft replies to payer emails and flag invoices that looked off. She connected it to her bank login, her email, her practice management system and the shared drive where contracts live. Within a week the bank locked her online profile. The agent had been logging in as her, at odd hours, faster than any person works. Her agreement said not to share credentials. She'd never thought of the agent as "sharing."
She was frustrated, and I get it. The agent was doing real work. The bank was protecting the account the way it protects every account, and she was the one who'd given the tool her keys.
That lockout was a gift. It forced a conversation about what the agent was allowed to do before something worse happened. It also made her look at what the agent could already see. Patient names, payer contracts, her personal email threads. Once data leaves your systems, getting it back is hard.
Here's how I'd set the rules if it were my company.
Decide what it can see, and what it can do. Reading transactions is one kind of access. Starting a payment, changing a vendor, or sending money is another. Most of the help Elena wanted was on the reading side. Keep the agent there unless you have a clear reason not to.
Never give it the owner's password. Ask your bank for a connection it supports, ideally read-only, or set up a separate user with limited rights. If the only way the tool gets in is by logging in as you, that is a red flag, not a feature.
Ask the vendor plain questions before you connect anything. Where is our data stored? Who at your company can see it? What happens to it if we cancel? Can we turn logging on and get a record of every action the agent takes? If they can't answer simply, that is an answer.
Keep a person on payments and approvals. Let the agent draft, match and flag all it wants. A person still approves every payment, and someone other than the person who set the agent up looks at the bank activity on a schedule. Spot-check its work the first month the way you'd spot-check a new hire.
Write down what to do if it goes wrong. Who can shut it off? Who calls the bank? Who reviews what it touched in the last forty-eight hours? You do not want to invent that plan at 9 p.m. on a Friday.
Elena kept the agent. She cut its access to read-only bank data through a connection her bank approved, pulled it out of her email and the contract drive, turned logging on, and she still approves every outbound payment herself. The work got faster. The keys stayed hers.
An AI agent is a new hire that never sleeps and never gets offended when you limit its permissions. Treat it that way. Useful is fine. Unwatched is not.

