An AI Agent Just Found $260M in Overcharged Invoices

Freehand just raised $75 million on a single bet: that most businesses are being quietly overcharged, and nobody is checking.

The Invoice You Didn't Check

It's Friday afternoon outside Columbus, Ohio. Marcus runs a twelve-person commercial cleaning company. Forty vendors send him bills every month: supply distributors, uniform services, equipment rental, fuel cards.

He pays them the way most owners pay them. Open the PDF. Check the total looks about right. Approve. Move to the next one.

He is not lazy. He is one person doing the job of an accounts payable department, a dispatcher, and a sales team, before lunch.

So when a supplier quietly raises a per-unit price by four percent, or bills him twice for the same delivery because two warehouses both submitted the invoice, or rounds a fuel surcharge up instead of down every single month, nobody catches it. Not because the fraud is clever. Because nobody is looking closely enough to notice.

That gap, the space between what a vendor bills and what a business actually owes, is bigger than most owners think. And this week, an AI agent put a number on it.

An AI Agent Just Read 19 Million Invoices. It Found $260 Million.

On 29 July 2026, the startup Freehand posted that it had saved companies like Meta, Unilever, and Johnson & Johnson a combined $260 million in cash by reading more than 19 million invoices and finding the baseless charges buried inside them. It then raised $75 million to back a guarantee: audit a company's invoices and either find at least $500,000 in overpayments, or pay that company $10,000.

Corey Haines, a marketing operator who runs his own small agency, posted something smaller and more useful the same week. He built a Claude Skill called /company-cfo that pulls his bank feed, Stripe, payroll, and card statements, reconciles them against each other, and catches the traps: double-counted transactions, internal transfers mistaken for income, a missing partner draw. It computes his end-of-month cash to the penny and writes the snapshot, on its own, every month.

Same idea. Different size. Freehand is doing it for enterprises with nineteen million invoices. Corey Haines is doing it for one agency with a laptop and a Claude subscription. Marcus, and every business owner like him, sits in between: too small for enterprise spend-management software, too busy to check every bill by hand, and finally, small enough that an AI agent can do the job for free.

How Do I Get an AI Agent to Check My Vendor Invoices for Overcharges?

You give it three things: the bill, the record of what you ordered, and the record of what you were charged last time. Then you ask it to compare all three before you pay.

In practice, that looks like this. You forward or upload a vendor's PDF invoice. You point the agent at your purchase orders, delivery notes, or simply last month's bill from the same vendor. You ask: does this match what we ordered, does the unit price match what we paid before, and has this invoice number already been billed?

A general-purpose chatbot can do a version of this once, in one conversation, and then forget everything the moment you close the tab. That was true of Claude eighteen months ago, and it is still true of a plain ChatGPT window today. What changed is the second half of the system: a place for the agent to keep score.

Jason Staats, a CPA who used to run a forty-person accounting firm, demonstrated this on his channel, Jason On Firms, by pairing Claude Cowork with Beancount, a free, open-source, double-entry ledger that stores every transaction as plain text. He fed it twelve months of bank statements, as scanned images, for a file with 6,700 transactions. Claude built about fifty pattern rules on the fly, the kind of "every time you see this vendor, categorize it this way" logic a bookkeeper writes by hand, and those rules alone handled 96 percent of the transactions correctly. The whole file took seven minutes.

The Beancount file is the part that matters for Marcus. It is not a party trick that evaporates when the chat ends. It is a running ledger the agent can check every future invoice against, and a paper trail a human can open and verify in plain text, line by line.

Why This Works Now, and Didn't Two Years Ago

Three things changed at once. Claude and its peers can now read a scanned PDF or a photographed receipt directly, no separate OCR tool required. They can hold a working file, a Claude Skill, or a text ledger like Beancount as long-term memory instead of starting from zero in every chat. And they can show their working, the steps between the invoice and the flag, in plain language.

That last point is the one owners underrate. In the same video, a feature from the expense platform Ramp called Stack showed an agent preparing month-end journal entries: a prepaid-expense accrual, a payroll entry built from a Google Drive file a teammate supplied. Every entry came with a workpaper. You could open the agent's session and read exactly why it booked the number it booked, before you approved it.

An agent that hides its reasoning is a liability. An agent that shows it, in a format you can check in under a minute, is a second set of eyes that never gets tired on invoice number 340 of the month.

Build Your Own Overcharge-Catcher This Weekend

1. Pick one vendor category to start. Not all forty vendors. The one that bills you most often, or the one you already suspect creeps up in price.
2. Give the agent your baseline. Upload the last three to six months of that vendor's invoices, plus your purchase orders or delivery notes if you keep them.
3. Ask for a plain-text ledger, not a one-off answer. Tell Claude (or your assistant of choice) to record each invoice, the line items, and any flags in a simple text file or spreadsheet it updates every time. This is your Beancount-style memory.
4. Write the three checks into a reusable prompt or Skill. Does the unit price match the last three invoices from this vendor? Does the invoice number already appear in the ledger? Do the line items match what was actually delivered?
5. Review only the flags, not every invoice. Once the ledger exists, a new invoice takes the agent seconds to check. You only need to look at the ones it flags, with its reasoning attached.
6. Expand vendor by vendor. Once the first category is running clean for a month, add the next one. Corey Haines' whole company-cfo system started as one reconciliation habit before it became a monthly close.

What You Have After This

A week in, you have one vendor category with a running ledger and a habit of forwarding invoices instead of just approving them.

A month in, you have a paper trail across your busiest vendors, and probably one overcharge, one duplicate, or one price creep already caught and refunded.

Six months in, you have what Corey Haines has: a monthly close that writes itself, catches its own traps, and hands you a snapshot instead of a Sunday spent in spreadsheets.

Same invoices. Same vendors. A completely different level of scrutiny, for the cost of a Claude subscription instead of a spend-management contract Marcus was never going to sign.