Invoice Coding Automation: A Controller's Real Story

A controller’s diary through the month she stopped fighting her invoices — and got her control back

A fictional narrative. Elena Marsh and Kestrel Property Group don’t exist. The close she describes happens every month, in some form, in every real estate finance department in the country.

---

Day 1 — The backlog

It’s the first business day after month-end, and there are four hundred and twelve invoices sitting in the queue. I know the number because I counted them twice, the way you re-count something you’re hoping will be smaller the second time.

I run accounting for Kestrel Property Group. Forty-one properties — office towers, three retail centers, a handful of light-industrial parks spread across four states. We run everything through Yardi Voyager, and on paper the process is clean: invoice comes in, gets coded, routes through Payscan for approval, gets paid. On paper.

The part nobody outside AP understands is the coding. Every one of those four hundred and twelve invoices has to be assigned a GL account, an expense account, and a property or cost center. A landscaping bill for a retail center in Ohio is not the same line item as a landscaping bill for an office park in Texas, even when the vendor is identical and the amount is within a dollar. Get it wrong and it doesn’t just sit wrong — it flows into the property P&L, into the owner reporting, into the reforecast the asset managers are already asking me about. A miscode isn’t a typo. It’s a small lie that travels.

So we code by hand. Marcus, who has been on my AP team for twelve years and knows our vendors better than our vendors know themselves. Priya, who started in March and is still learning that “utilities” means six different things depending on the property. And me, at the end, as the last set of eyes before anything posts.

Four hundred and twelve. It’s 8:40 in the morning. I already know I’m losing the weekend.

Day 2 — The one that got through

Here’s the thing about doing this job well: nobody notices. You catch the mispost, you fix it before it flows, and the reward is that nothing happens. The failures are the only visible part.

Last quarter one got through. A capital item at our Dallas office park got coded as a repair-and-maintenance expense instead of capitalized. On four hundred invoices a month, one wrong code. It surfaced three weeks later when an asset manager asked why the property’s operating expenses had spiked, and I spent a full day tracing it back, reversing it, re-explaining it to someone who now — reasonably — wondered what else we’d missed.

That’s the job I actually have. Not “code the invoices.” Guarantee the numbers. The coding is just the surface. Underneath it is a promise I make to the owners, to the auditors, to the CFO who signs off on statements I built: that what you see is real.

Which is exactly why, when people first told me an AI could code our invoices, my instinct was to close the tab. Not because the idea was stupid — because the people who usually make that kind of promise have never sat through a close, and I had no reason to believe this was any different.

Day 3 — Why I said no

I want to be precise about my resistance, because I don’t think it was irrational and I don’t think I’m the only one who has it.

Everything I read about it led with speed — fully coded invoices in seconds. And speed was the least interesting thing anyone could have offered me. I live in the world of the exceptions: the vendor who changed their billing name, the invoice that covers three properties on one PDF, the recurring charge that quietly doubled. Fast was never my problem. Being wrong quickly is worse than being slow — no number of saved seconds buys back a miscode that already flowed.

My second fear was the one I didn’t say out loud in meetings: if a machine codes four hundred invoices and it’s wrong on forty of them, in a way I can’t see, then I’ve traded a slow, visible process I control for a fast, invisible one I don’t. I’d rather have the backlog than the black box. At least the backlog is honest about what it is.

And my third fear was Marcus. Twelve years. If I bring in something that “codes the invoices,” what exactly am I telling the person whose job is coding the invoices?

So I said no. Then I said no again. Then the Dallas mistake happened, and the close after that ran until Sunday night, and by the second day Priya had stopped asking questions and started just nodding at them, and I stopped being able to afford my no.

Day 4 — The question that changed the shape of it

I finally took the call, and I came in loaded. I didn’t ask about speed. I asked the two questions that actually keep me up.

“When it’s wrong, how do I know?”

The answer reframed the entire thing for me. The system doesn’t replace my approval workflow — it feeds it. It sits in front of Yardi and Payscan, not on top of them. It reads the invoice, predicts the GL account, the expense account, the property, and it attaches a confidence level to each prediction. The high-confidence ones — the recurring, unambiguous, seen-it-a-thousand-times invoices — come through coded and ready. The uncertain ones get flagged for a human, surfaced, not buried. It’s not a black box that decides. It’s a system that does the obvious 80% and raises its hand on the 20% that needs judgment.

That was the word that unlocked it: judgment. It wasn’t taking the judgment away. It was taking the typing away and handing me back the judgment, concentrated on the invoices that actually deserve it.

“Where does it get the idea that Dallas landscaping is different from Ohio landscaping?”

From us. From our own history. It learns each client’s coding from that client’s own past invoices and vendors — not some generic rulebook, ours. The Dallas-versus-Ohio distinction that lives in Marcus’s head is exactly the pattern it’s built to learn. And it’s real-estate specific — it understands property and cost-center coding, which is precisely the layer that generic OCR and horizontal AP tools always left for us to do by hand. OCR reads the invoice. This codes it. That’s a different job.

I asked about the auditors. SOC 2 Type II. I asked whether it would fight Yardi. Certified interface partner. I asked what happens in year one versus year three — the honest answer was 70 to 80% automation from day one, climbing toward 99% accuracy as it learns more of our history. Not a magic 100 on Monday. A curve I could actually believe, because I’ve never trusted anyone who promised me a hundred.

I didn’t feel sold. I felt something more useful. I felt like I’d been handed a framework I could stay in control of. Freedom within a framework. I keep the approval. I keep the exceptions. I keep the promise to the owners. I just stop typing four hundred property codes by hand.

Day 5 — Telling Marcus

This was the conversation I dreaded most, and it turned out to be the one that mattered most.

I didn’t dress it up. I told him we were bringing in a system that would code the high-confidence invoices automatically. I watched him do the math I knew he’d do — twelve years, and here’s the machine.

So I told him the part I’d figured out at 2 a.m. His value was never the typing. His value is that he knows — which vendor is really three vendors, which charge is about to be disputed, which property manager always miscategorizes their utilities. For twelve years we buried that knowledge under data entry, spending his judgment on keystrokes. The system does the keystrokes. What it can’t do is be Marcus on the exceptions. It flags the hard ones; he’s the one who resolves them.

His job wasn’t disappearing. It was getting promoted out of the parts that were beneath it.

He didn’t fully believe me that day. He believed me a month later, when he spent the close reviewing eighty flagged exceptions instead of hand-coding four hundred invoices, and left the office on Friday.

---

The close after

I want to tell you it was magic. It wasn’t magic. It was the first close in four years that ended on a Friday.

The invoices came in coded. Not all of them — the uncertain ones surfaced, flagged, waiting for a human, which is exactly what I’d asked for. Marcus worked the exceptions. Priya stopped drowning; she spent her week learning the properties instead of guessing at them. The arrival-to-approval cycle that used to run eleven days ran closer to three. And the number I actually cared about — the one that isn’t on any slide — the number of miscodes that flowed into owner reporting: zero. Not because a machine is perfect, but because the machine did the volume and the humans did the judgment, and that’s the combination that was always supposed to be the point.

We didn’t hire. We’d been about to. The team absorbed a growing portfolio without adding a head, because the same people could now handle two to five times the invoice volume when the coding stopped being manual. On paper that’s a capacity story for the CFO. On my floor it was Priya getting through her first clean close, Marcus no longer resenting the machine or me, and me not losing a Sunday to a mistake I could have caught if I hadn’t been buried.

Here’s what I’d tell another controller — the version of me who said no three times:

The pitch will be about speed, and speed is not your problem. Don’t evaluate it on speed. Evaluate it on control. Ask where the errors surface. Ask where the knowledge comes from. Ask what happens to your approval workflow, your audit trail, your people. If the honest answers are the errors surface to a human, the knowledge comes from your own history, your controls stay yours, and your people move up the stack — then it isn’t taking your job away.

It’s giving you back the job you were actually hired to do.

I was hired to guarantee the numbers. For four years I spent that job typing property codes. I don’t anymore. And the promise I make to the owners — that what you see is real — I can finally keep it without losing my weekends to it.

The invoices still come in four hundred at a time. I just don’t count them twice anymore.

---

Kestrel Property Group is fictional. The month it describes is not. If your close still runs to Sunday night, the question worth asking isn’t how to code faster — it’s how to get your control back. That’s the question PredictAP was built to answer: AI-powered invoice coding built for real estate AP, that feeds your Yardi and Nexus workflows instead of replacing them, learns your coding from your own history, and surfaces the exceptions to the people whose judgment they deserve.