The Week I Worked for an AI Boss

Last week I reviewed a client’s books alongside an AI. The client, running their numbers through an AI tool, came back and asked me to make an adjustment. I challenged it. I was fairly sure the AI had missed context, but I made the change anyway to keep things moving. In the end, I was right. We reversed the adjustments. Luke 1, AI 0.

But here’s the honest part: the AI did a pretty good job. It got most of the way there. Today it still gets confused, misses context, or doesn’t connect the dots the way a person does. How long until it catches every minor coding discrepancy cleanly? I don’t think it’s long at all.

The strange thing that actually happened

What stuck with me wasn’t the adjustment. It was the shape of the whole exchange.

For a week, I went back and forth with a client on an issue. The client was feeding my responses into an AI and sending the AI’s replies back to me. I found myself explaining, re-explaining, and defending correct work — not to the client, but effectively to a machine using the client as its mouthpiece.

I was taking orders from an AI. And it made me wonder how long before that’s just the arrangement. We talk about people using AI to get things done. But the flip is already visible: AI using select humans to execute what it wants — validating, confirming, doing the physical or relational parts it can’t. AI management, human input.

Where does the line sit?

I’m experiencing this in accounting, but it’s a general question. If I’m going in for surgery, do I start feeding every detail into an AI and telling the surgeon how to operate? Do lawyers begin defending their advice to an AI second opinion? Can architects design a building without software dictating what’s allowed?

Some of this is genuinely useful. Some of it is a person outsourcing their judgment and then asking a professional to argue with the output. The line matters, and I don’t think anyone has drawn it yet.

What AI still doesn’t do well

Here’s what my week as an AI’s employee taught me: the AI concluded. It didn’t explain. It handed over an answer without the reasoning that would let a professional check it, trust it, or overturn it.

That gap is exactly where a good accountant earns their keep — and it’s the gap we build around. Flying Ledger is designed to retain the work papers and the explanations behind every number, not just the number itself. When Ask Fly answers a financial question, it answers from the platform’s own live numbers, not a guess. When revenue recognition posts a deferred-revenue journal entry into QuickBooks, it’s broken down by category so you can see why. When intercompany reconciliation flags a mismatch, it’s telling you the two sides of the books don’t agree — and where.

That’s the difference between a conclusion and an audit trail.

Being correct isn’t the whole job

Even when AI gets the number right, businesses still need something to take all those numbers, confirm they’re correct, and turn them into operational drivers they can act on. That’s the work: a CFO Dashboard that shows cash, debt, revenue, net income and AP across every entity with drill-down into any figure. A financing snapshot a lender can actually read, with loan balances and DSCR. Cleaned-up vendor names so you know who you really owe and what’s coming due. Daily POS revenue reconciled against what the POS system itself reports.

None of that is a single conclusion. It’s a system of checks, each one explainable.

Where we’re headed

Flying Ledger uses AI where it makes sense to. We’re actively using the platform internally with real customer data today — it’s not open to external users yet — and that hands-on use is exactly why I keep coming back to explainability. Whether the person asking the question is a controller, a CFO, or an AI relaying instructions through a client, the answer needs to come with its reasoning attached.

AI will keep getting better. My guess is it gets very good, very fast. But the businesses using it will still need to know their numbers are right, understand why, and know what to do next. That’s the part we’re building for — and the part I don’t think AI has mastered yet.

Luke 1, AI 0. For now.