AI & the Finance Function

I priced an AI replacement for a bookkeeper across nine entities. I walked away.

The pitch is everywhere: fire the bookkeeper, run the agent, keep the salary. I took it seriously and built the model. Here's the math the vendors don't show you — and the line item they leave off every slide.

The pitch is everywhere right now, and it's clean: fire the bookkeeper, run the agent, keep the salary. Autonomous general ledgers, AI close platforms, “your books, no humans.” I took it seriously. I'm a controller who already runs production AI workflows on six-figure-row GL extracts — forensic reconciliation, accrual audits, valuation models. I am not the guy who says AI can't do the work. So when the chance came to replace a bookkeeper with an agent stack, I didn't dismiss it. I priced it.

Then I walked away. Here's why — and why almost every CFO getting this pitch is being shown half the equation.

What I was actually replacing

The role was a lead bookkeeper costing $72.5K a year, fully loaded — salary plus benefits. Reliable, full-time, low-cost throughput across a nine-entity general ledger — intercompany activity, period-end accruals, deferred revenue, cash matching, the lot.

The proposed replacement was a stack: an autonomous GL agent to do the posting and an AI close layer on top to handle reconciliations and variance work. I talked to the companies. I pulled real pricing. I read the tiers. And I costed the one line item every vendor leaves off the slide.

The number the vendors price — and the one they don't

Vendors price the agent. They do not price the reviewer.

That omission is the entire ballgame, because the $72.5K you “save” does not vanish when the human leaves. It redistributes. Some of it becomes agent subscriptions and per-entity fees. The rest gets pushed up the cost curve — onto the controller — and that is the most expensive labor in the building.

Here's the mechanism the pitch skips. An autonomous bookkeeping agent does not fail by being slow. It fails by being confidently wrong. It posts a misrouted intercompany entry, a mis-mapped deferral, a broken accrual — and it does it at machine speed, at volume, without flinching. The only person qualified to catch a bad intercompany route or a mis-stated accrual before it poisons the financials is the controller or CFO. So the work doesn't disappear. It converts from cheap data entry into expensive forensic review.

That is the trade I was actually being offered:

• —Give up: $72.5K of reliable, full-time, low-cost throughput.
• —Take on: agent fees + implementation + a recurring slice of the highest-cost labor in the company, spent reviewing, approving, and unwinding machine errors.

Run that honestly and the “savings” evaporate. The agent stack plus implementation already eats a large share of the $72.5K before you cost a single hour of oversight. Add the review burden — and on a complex multi-entity ledger that burden is not trivial — and the project is underwater. I wasn't cutting a cost. I was swapping a known, contained cost for a variable one I'd absorb personally, at my own billing rate.

Why this only pencils in a world that doesn't exist

The replacement math works under exactly one condition: the agent's error rate is low enough that review stays near-zero. In a clean, simple, single-entity ledger with standardized transactions, maybe you get close.

That is not where most real businesses live. At the bottom of a multi-entity GL, the errors that actually hurt — misrouted intercompany transfers, incorrectly spread invoices, deferred revenue mapped to the wrong account, recurring templates pointed at the wrong target — are precisely the ones an unsupervised agent generates fastest and a non-expert reviewer is least likely to catch. I have spent real weeks unwinding exactly these errors when humans made them at human speed. An agent makes the same class of mistake faster and at higher volume. More throughput on a process that produces high-cost errors is not a saving. It's leverage in the wrong direction.

This is the real “AI cost crisis”

You've seen the viral version: companies blindsided by token bills, CTOs blowing the annual AI budget in four months, breathless takes about renting versus owning intelligence and which superpower wins. Most of it is theater bolted onto a mundane finding.

The actual cost crisis isn't hidden token bills. It's hidden labor transfer. The replacement pitch quietly moves work off a cheap, full-time line and onto your most expensive people — and calls the gap “savings.” Anyone reading invoices instead of building a model discovers this in April, after they've deployed. Anyone who builds the model first discovers it before they spend a dollar.

The diligence that catches it isn't exotic. It's two moves the pitch is designed to skip:

1. Talk to the vendor and pull the real, all-in price — seats, entities, implementation, the service tier you'll actually need, not the headline number.
2. Cost your own time as a line item. The reviewer is not free. If the design pushes hours onto a controller or CFO, those hours belong in the model at that person's rate.

Do both and the answer falls out on its own.

Where AI replacement does work — and where it doesn't

To be clear, this is not an anti-AI argument. I run AI through my own close every period. The distinction that matters is augmentation versus replacement, and the deciding question is always the same: who absorbs the cost of an error?

• —Augmentation wins where the human stays in the loop and the AI compresses their time — drafting flux explanations, surfacing anomalies, accelerating reconciliations a person still approves. The error cost stays contained because a competent reviewer was always going to look.
• —Replacement loses where you remove the human entirely from an error-sensitive, judgment-heavy process and quietly reassign the catch-and-fix work to someone who costs three times as much. The economics only invert when the agent is reliable enough that the review tends toward zero — and on a complex ledger, it isn't there yet.

The vendors selling “fire the bookkeeper” are selling replacement and pricing it like augmentation. That's the sleight of hand.

The takeaway

I'm not against the technology. I'm against the math being shown with one line missing. Before you replace a finance role with an agent, build the model the way you'd build any other: every cost in, including the hours that land back on your own desk. Count the reviewer. Count yourself.

When I did that for nine entities, the contrarian answer was the correct one — and I'd have never seen it from the invoice. Only from the model.

jiesen Li · Finance & accounting transformation · AI in the close · multi-entity controllership