Booking a journal entry with a robot

Ask an accountant to let AI post journal entries on its own, and you’ll probably get a strong negative reaction. They are picturing a scenario where a bot runs rogue and books hundreds of incorrect entries, costing them time and getting them in trouble with auditors.

I had a similar fear. In accounting, we answer to auditors and regulators. Incorrect numbers aren’t just a “hallucination,” they will affect how the business operates and fail an audit. When accountants fear AI, it’s not resistance to change, it’s pragmatic.

Yet today, my team at an insurance company uses agentic operations to post to NetSuite every single day.

My first attempt at an automated entry involved an AI driving my Chrome browser, clicking through the NetSuite website. It took ten minutes to do a one-minute job. Worse, it had no inspectable steps, no approval gates, and no limits on what it could do. It was an impressive demo, but it wasn’t enough to use.

What actually earned our trust was building strict, mandatory checkpoints into the process. Today, the operations we’ve built prepare our journal entries autonomously in the background. They then pause, send us a secure form to review, and a human accountant approves it. Only then is the system able to post it to the ledger.

These checkpoints don’t just protect the ledger, but solve one of the biggest hidden time-sinks in accounting: chasing people down. Month-end close usually involves hunting for approvals, waiting on partners, and dealing with bottlenecks when someone is out of the office. Now, if an operation needs an approval on an overdue bill, the agent flags the right people in Slack and waits. It doesn't forget, and it will escalate to someone else if they are out of office. Once the right person responds, the agent immediately finishes the job. It completely removes the burden of manual coordination.

We also rely on heavily scoped access. Our connection to Ramp can’t charge our credit cards or send payments, but it can flag overdue bills. Just because one workflow step can check for overdue bills, doesn’t mean it can check for reimbursements.

We’ve built this trust incrementally, over a few months. Our first autonomously-posted journal entry was two lines long. It didn’t save us time, but it proved a hypothesis that we could do this in a way that we trust. By August, we were confidently posting 500-line entries with the system.

The biggest takeaway for me was realizing that a human reviewer isn’t just at the end, but throughout, similarly to when I bring someone new onto my team. Because I can shape the checkpoints for each operation, and I’ve done this work for years, I can have AI show its work in a way that lets me spot mistakes instantly. Some AI systems show me code and ask if they can run it. But I don’t know code. I know our accounting books and what different parts of the spreadsheet should look like.

When you bring AI into a regulated space, the goal isn’t to take your domain experts out of the loop. It is to move them from doing the rote work to actually guiding the operation.