I tracked a month of spending using only receipt photos: the radical experiment

By Daniil Kozin · ex-banker · Florianópolis

For 30 days I logged every expense with a photo of the receipt and nothing else: no typed chat messages, no manual entry, no bank statement upload. Capi's vision model read 42 receipts across two currencies. It got 33 right on the first try, hallucinated a year twice, and misread a currency three times. Four days it could not cover at all.

I run Capi, so this was partly a stress test on my own product and partly a real question I had never answered for myself: if a photo is the easiest way to log a purchase, why do I still type half my transactions into the chat? The rule was simple and I broke it exactly four times, for reasons that turned out to be the most useful part of the month.

What was the one-photo-a-day rule?

The rule was one photo per purchase, sent to Capi's Telegram chat, with no typed amount and no correction unless the app flagged the entry as pending. A day with three purchases meant three photos. Over 30 days I sent 42 photos: groceries, coffee, a haircut, a hardware store run, two dinners out, a co-working day pass, and a handful of Uber receipts screenshotted straight from the app.

I tracked two currencies during the month, Brazilian reais for daily life in Florianópolis and US dollars for a work trip. Total spend across both came to the equivalent of $2,140. Each photo took about 15 seconds to take and send, against roughly 10 seconds for a typed chat message with amount and category. The overhead was real but small, and it disappeared entirely on days with more than one purchase, since the photo replaced typing out a merchant name I would have half-remembered anyway.

How accurate was the vision model at reading receipts?

Capi's vision model read merchant, amount and currency correctly on the first try for 33 of the 42 receipts, about 79 percent. The other nine needed a manual correction, and every one of the nine came from the same two causes: faded thermal paper past its first week, or a receipt photographed at a sharp angle in bad light.

That number lines up with what independent OCR benchmarking has found through 2026: current vision-language models cut character error rates roughly three to four times lower than older OCR engines on messy scans and receipts specifically, with clean printed totals near one to two percent error and much worse rates on faded or crumpled paper. My own 21 percent first-pass failure rate sits inside that range once you weight for how much of my sample was thermal paper carried in a pocket for a few days before I got around to photographing it. The fix in every case was the same: Capi flagged the entry as pending rather than committing a guess, and confirming it took one tap.

Which categories were hardest for the vision model to read?

Groceries and restaurants were the easiest categories, misread zero times out of 19 receipts combined, because supermarket and restaurant printers use fresh thermal paper with a clear total line. Coffee shops and hardware stores caused the most trouble, six of the nine failures between them, because small receipts get folded into a pocket and photographed later, once the print has already started to fade.

The breakdown by category: groceries 11 receipts, zero misreads. Restaurants 8 receipts, zero misreads. Coffee 9 receipts, four misreads, all faded paper. Hardware store 3 receipts, two misreads, one angle and one fade. Transport, mostly Uber screenshots rather than paper, 6 receipts, one misread, a screenshot cropped too tight to catch the currency symbol. Co-working and services, 5 receipts, two misreads. The pattern held all month: paper age and photo angle explained every failure, not the merchant type or the amount involved. A $4 coffee and a $340 co-working invoice failed for the identical reason, a shot taken too fast in bad light.

What went wrong: the year hallucination and the currency mixups?

Two receipts got their printed year misread, once as 2025 and once as 2027, and three photos produced the wrong currency, always on a blurry or low-light shot where a peso sign and a dollar sign look close enough for a model to guess wrong. Neither mistake was silent. Both showed up as a flagged entry waiting for confirmation.

The year hallucination sounds worse than it is, because the transaction date in Capi comes from the message timestamp, not from parsing the printed date on the receipt. So a receipt with a hallucinated year still landed on the correct day in my ledger; the OCR error only showed up in a secondary text field I never actually use to sort or filter anything. The currency misreads were the sharper problem, because a wrong currency changes the number, not just a label. All three were caught before they ever touched my monthly total, which is the entire point of a pending-by-default design: a model that is right 79 percent of the time is only safe if the other 21 percent asks before it writes.

Did skipping manual chat logging actually work?

Mostly. Card and formal retail purchases photographed well because they print a receipt: groceries, restaurants, the hardware store, the co-working pass. Four days broke the rule entirely, because the purchase had no receipt to photograph: a tip in cash, a street food cart, and two small favors paid back in cash to a friend. Nothing to point a camera at meant nothing got logged unless I did something else.

What I did on those four days was send a ten-second voice note instead, the same input Capi uses for hands-free logging covered in the voice-note tracking test. That is the honest finding: photo-only tracking is not a full replacement for chat or voice, it is the best option specifically for spend that generates a paper trail. My actual daily habit, which I wrote up separately in what I track every day, already blends photo, voice and the occasional typed line for exactly this reason. This month just proved why the blend exists instead of assuming it.

How does photo-only tracking compare to Copilot Money and Monarch?

Copilot Money has no receipt-photo capture at all in 2026. It categorizes bank transactions well but there is nothing to point a camera at inside the app. Monarch added native receipt scanning that reads merchant, amount and line items, but it sits behind the Plus tier at roughly $199 a year, above the base $99.99 plan. Capi ships photo capture on every tier, including free.

If your bank already syncs cleanly and you never touch cash, Copilot's categorization is fast and you may never miss photo capture. If you want receipts to build a real paper trail behind a card statement, Monarch's version is polished but paywalled. The gap this test exposed is not about which vision model is smarter, it is about which product lets you start photographing receipts from day one without upgrading first. The full field is ranked in the 2026 money tracker guide, and the multi-currency side of my own setup is covered in Capi vs Monarch.

One caveat before the verdict: this is a sample of one person, one month, 42 receipts, run on the product I built, so treat the 79 percent figure as a single data point rather than an industry benchmark. What I can vouch for is the failure pattern, because it matches the wider 2026 OCR research on messy receipts and it matches complaints I hear from Capi users every week: fresh paper reads clean, old folded paper does not, and a model's confidence in a wrong currency guess looks identical to its confidence in a right one until a human checks.

Is photo-only expense tracking worth it in 2026?

Yes, for the roughly three-quarters of purchases that print a receipt, and no, not as your only method. The realistic outcome of this month was not "photos replaced everything," it was "photos replaced most of my typing, and voice quietly covered the rest." That combination cost me less attention than either method alone.

The honest number to remember is 79 percent first-pass accuracy and zero silent errors, because every miss surfaced as something to confirm rather than a number that quietly went in wrong. A vision model does not need to be perfect to be useful. It needs to know when to ask, and for a month that mattered more than raw accuracy did.

Frequently asked questions about photo-based expense tracking

What is photo-only expense tracking?

Photo-only expense tracking means logging every purchase by sending a photo of the paper or digital receipt to a finance app, instead of typing the amount or connecting a bank feed. The app's vision model reads the merchant, amount and currency off the image and creates the transaction automatically.

How accurate is AI at reading receipts from photos?

In this 30-day test, Capi's vision model read merchant, amount and currency correctly on the first try for 33 of 42 receipts, about 79 percent. The rest needed a manual correction, most often on faded thermal paper or a receipt photographed at an angle.

Can you track all your expenses using only photos, with no manual entry?

Not entirely. Card and formal retail purchases photograph well because they print a receipt. Cash purchases with no receipt, a street vendor or a tip, cannot be photographed, so four of the 30 days needed a quick voice note instead to stay logged at all.

What are the biggest risks of photo-only expense tracking?

The two real risks are a misread currency on a blurry photo and a misread year on faded print, sometimes called a year hallucination. Both showed up in this test. Neither corrupted the ledger, because misreads land as a pending entry to confirm and the transaction date comes from the message timestamp, not the printed text.

How much does Capi cost for photo-based tracking?

Photo logging ships on every Capi tier, including the free plan up to 30 transactions a month in one currency. Capi Core costs $9.90 a month or $69.90 a year and removes that limit. Capi Together, for two people, costs $99 a year for the household.