How to capture US business receipts with AI and stop typing

TLDR: Dext is the specialist when receipt volume or line-item extraction matters. FreshBooks is the simple choice for a service freelancer, with OCR from its $43 Plus plan. QuickBooks Online makes sense when the business already keeps its US books there. The bottleneck is not reading the document, it is getting the document in.

Thirty-four lines that say “need documentation”

Your accountant sends the list in February. Thirty-four transactions from last year with no receipt attached, most of them between forty and ninety dollars, spread across eleven months. You recognize maybe a third. For the rest you are looking at a card charge from a Tuesday in June and trying to remember which client you were with.

Every one of those lines is money you already spent. Some of it will end up unsupported, and all of it costs you an afternoon and costs your accountant billable time to chase. Nothing about that afternoon is an accounting problem. It is a capture problem that happened eight months earlier and stayed invisible until somebody finally counted.

Which is why the question when picking a tool is not how many fields it extracts. It is how much friction there is between paying for something and the document being inside the system, while you are standing at a counter with a phone in one hand.

The route, and who handles each leg

Notice the drawing opens in lavender rather than lime. That is backwards from almost every other process in this category, where the machine starts and a person signs off. Here a person starts, which is why the best product on the market performs exactly as badly as the worst one if the habit is not in place.

Build two inlets and no more

One for paper, which is the phone camera, and one for digital, which is a single email address anything in PDF gets forwarded to. Two inlets get remembered. Five do not, and five eventually becomes zero.

The forwarding address has an advantage the camera does not: you can hand it to somebody without explaining anything. Give it to whoever receives vendor bills and the circuit works without that person ever logging into the accounting product or needing a seat.

Specialist or built in

A built-in tool removes steps because the document is born inside the ledger. FreshBooks starts at $43 and includes receipt scanning from Plus, deliberately narrower than a full accounting copilot but enough for a service business. QuickBooks Online starts at $75 and keeps capture next to the rest of the US workflow, so there is no handoff if the books already live there.

Dext starts at $25 and specializes in pulling vendor, date, tax, totals and line-item detail. It feeds several accounting systems and is the strongest fit for a firm or a business with real volume. Its limit is plain: it is not a ledger, so the integration is mandatory rather than optional.

The practical rule is short. If volume is the problem, go specialist. If the habit is the problem, go built in, because owning one fewer tool matters more than extracting two more fields.

The four exceptions that always jam

The faded thermal receipt. Six weeks in a wallet and the ink is gone. There is no technical fix, because there is no text left to read. The process fix is capturing at payment. If a recurring vendor always fades, ask them for an emailed invoice and stop dealing with the paper.

The bill that only exists in a vendor portal. Utilities, telecoms and some platforms never email anything. They post a PDF in your account area and send a notice that reads like marketing, so nobody opens it and nobody forwards it. These have to be pulled by hand once a month, and it pays to keep the written list of which vendors they are.

The restaurant receipt with a handwritten tip. The printed subtotal and tax extract fine. The tip you wrote in pen, and the new total you wrote under it, are the two numbers OCR gets wrong most often, and they are the ones that decide the amount. Route these to review even when the system reports high confidence.

The expense on a personal card. Somebody on the team pays for something out of pocket. There is no company card charge behind it, so reconciliation will never chase it and the document floats. It needs its own reimbursement path with the receipt attached, or it disappears.

What to count before deciding

With that batch in front of you, count three things. How many documents came out right without you touching a field. How many minutes you spent on the ones that did not. And how many reached the ledger with the image attached rather than the data alone, because an extracted figure with no image behind it is an entry you cannot defend.

The following month, count a fourth: how many expenses closed with no supporting document at all. That number tells you whether you fixed the problem or moved it somewhere else.

Do not use the accuracy figure on the vendor’s own page. It was measured on their documents. Your real mix is the only benchmark that decides anything.

Set the approval boundary

Extraction proposes and a person confirms, every time, in three specific cases: when the system flags low confidence, when tax is broken out, and when the amount is large. Everything else should clear on rules, and has to clear on rules, or review becomes the typing you were trying to remove.

A cleanly extracted field does not turn a blurred photo into evidence either. Keep the readable image next to the entry, with the date and a record of who approved what. Retention requirements depend on the document type and your situation, so confirm the rule with your accountant rather than inferring it from a product page.

This flow feeds bank reconciliation and the month-end close directly. If your bookkeeper is the one reviewing documents, agree the split first in working with an accountant. And keep it separate from the invoices you send out, which run the other way around the circuit and get chased in collections.

Frequently asked questions

Should I keep the paper original?

That depends on the document type and the retention rules that apply to you, and your accountant confirms it. What you should demand from the system either way is a readable image with a date, an audit trail and an export, because without those it does not matter what you decide about the paper.

Dext or a built-in feature?

Dext when the bottleneck is document volume or when several people capture. A built-in feature when you want fewer tools and volume is ordinary. If you are unsure, start built in: moving from a built-in feature to a specialist later is easy, and the reverse is not.

When is it not worth it?

Under about fifteen documents a month, most of them arriving as emailed PDFs, automated capture saves very little. The unsupported-expense argument still holds, but at that volume it is an argument about discipline rather than about software.

Can I just use a general chatbot to read the receipts?

You can, and it will read them reasonably well. What you lose is everything around it: permissions, an audit trail, the connection to the ledger and the image stored beside the entry. You are uploading financial documents somewhere that has none of those four things in order to avoid a subscription.

How long until it pays for itself?

Compare capture and review minutes across one full month, before and after, and include the corrections your accountant makes, because you pay for those too. That calculation decides it, not the sales demo.