Why can't more data fix the veterinary AI data problem?

Pet care platforms deal with health records constantly. A boarding facility needs proof of vaccines before a dog can check in, a rescue gets a folder of paperwork with a surrendered pet, a vet sees a new patient and needs to know what another clinic did six months ago... the list goes on. The records involved (vaccination certificates, vet invoices, intake forms, health certificates for travel) come faxed, photographed, scanned, exported from a dozen different systems in a dozen different ways.

Platforms using Pawssier don't spend much time thinking about any of this. They send any document and get back structured data, without needing to worry about everything in between. The problem and solution seem simple enough, right? Not quite. The data problem in pet care has a lot more to it than it first appears.

How most platforms are solving it today

Manual entry. Front desk staff, rescue volunteers, account managers, and oftentimes the pet owner themselves, copy records by hand into whatever system their platform uses. Pawssier co-founder Nick volunteered at a humane society where surrendered pets would come in with records that nobody had time to go through until the moment of adoption.

As a result, critical health history gets missed, or has to be pieced together quickly right when someone needs it. Manual entry is slow, it doesn't scale, and it introduces errors at every step.

Pointing a general AI model at the problem. Some platforms route their documents through ChatGPT, Claude, Gemini, or a similar general AI model with a prompt designed to extract the fields they need. On the surface, this looks like it solves the problem. But if you've ever told ChatGPT not to use em-dashes and watched it use them anyway, you already know the core problem: general AI does what it wants with your instructions, and there's no guarantee it'll follow them consistently. It works for email-writing but is a real problem when the output is feeding a travel approval or an insurance claim.

We've documented this in a previous post and benchmark against it regularly. General AI models pointed directly at veterinary records without a purpose-built validation layer fail in specific, predictable ways. To name a few: they guess units instead of flagging ambiguity, grab the wrong date off a certificate, fill gaps with plausible-sounding values rather than admitting uncertainty, and give you no way to know which parts of the output you can actually trust. When those failures compound across thousands of records, a compliant pet could look overdue (resulting in frustrated pet owners and wasted time), a missing lot number passes QA unchallenged, or travel approvals get issued on unverified data.

The instinct at this point is usually to add more data: train the model on more records, cover more cases, get better results. But that just makes the underlying problem harder to fix.

Why more data doesn't fix it

The scarcity in veterinary AI is about variety, not volume. There are a lot of records out there, but the distribution is heavily skewed. Most of them cluster around the same few practice management systems, the same document layouts, and the same regions. Even if your solution works on 80% of the most common records, that's 2 out of every 10 records that breaks the system and requires manual intervention.

Early on at Pawssier we went looking for records to train on. We asked friends and family, gathered what we could from our own networks, and built a small corpus. What we got was a pretty accurate reflection of who we know: similar regions, more dogs than cats, records almost entirely in English, and documents from a handful of the same practice management systems.

When you train on a corpus like that, the model learns those layouts well. But point it at a record from a different system, a different country, a handwritten note from a rural clinic, a health certificate in a format it's never seen, and performance drops because the model learned a layout, not the domain.

The next instinct is rules: encode everything you know outside the model and catch what the model misses. The problem with rules is that they only handle what you've already seen and miss everything you haven't. Every new document type, every new language, every new clinic's formatting quirks requires a new rule or a patch. Over time you end up with a patchwork of rules that handles the cases you've already seen, while the model underneath hasn't actually improved.

What actually works

The data scarcity problem in veterinary AI is structural. It doesn't go away with more compute, a better base model, or a cleverer prompt. What solves it is genuine variety built up over time from real-world documents across the full range of what pet care actually looks like.

At Pawssier, we don't train on customer data. What we've built instead is a pipeline that gets smarter the more we run it, without that dependency. A foreign date format that caused failures when we first encountered it is now handled correctly in every record we process, because it's built into what the model knows, not bolted on as a rule afterward. A new entrant hitting that same format tomorrow has to discover it, build against it, and test it before they can move past it. And the longer we run, the harder that gap gets to close from a standing start.

Platforms using Pawssier today don't have to worry about any of this. They send a document, they get back structured data with confidence scores and field-level provenance, and they build on top of it. The variety problem, the edge cases they haven't seen yet, the records from systems they've never heard of, all of that is our problem, not theirs. The platforms still on approach one or two are solving it themselves, one rule at a time, one format at a time, every time a new edge case shows up that their model hasn't seen before.

Ask the Pawssier bot →