Most conversations about AI in business start with a chatbot. The bigger opportunity in finance and operations is quieter: agents that do the checking, matching and reporting a person does by hand every morning, and hand anything that needs judgment back to that person.

The hard part is not the model. It is deciding what an agent is allowed to do. These are the principles we use, and the order we recommend starting in.

Start where mistakes are cheap

The best first agents only read and report. They cannot break anything, and they prove their value within a week.

  1. A morning brief. Cash collected yesterday, what failed overnight, and what is waiting for a decision. Nothing else.
  2. A health monitor. Every couple of hours, check the integration error log, payments stuck in processing, and automations that stopped running.
  3. Plain-English questions. "Which customers over 60 days have not been called this month?" answered with a list and a spreadsheet.

Then let agents fix the safe cases

Once the monitor has run for a few weeks, patterns appear. Some errors have one obvious, reversible fix every time. Those are the candidates for an agent to correct on its own, with a log entry for each change. Everything else stays a ticket for a person.

Four guardrails that are not optional

  • Run as a real user. The agent signs in as a named Salesforce user with its own permissions. If that user cannot see a record, neither can the agent.
  • Read-only by default. Write access is granted per task and per object, deliberately, never as a blanket setting.
  • People approve money. Anything that creates, changes or sends a financial record waits for a named person's approval.
  • Log everything. What the agent read, what it changed and who approved it, stored where an auditor can find it.

These rules belong in code, not in the instructions you give the model. A prompt can be ignored; a permission cannot.

What good looks like after three months

Problems surface the same day instead of at month end. The team starts the morning with a short list instead of five reports. And the people who used to do the checking spend that time on the exceptions that actually need them.