Last updated: March 2026 · 5 documents tested, testing ongoing
Why general-purpose AI fails on veterinary records
We tested 5 real veterinary documents across Pawssier and three leading AI tools.
Here is what we found.
0 of 3
General AI tools that detected multi-record documents
When a PDF contained 5 separate clinic visits, all three tools returned a single record with no indication data was merged.
100%
Rate at which Gemini flagged unambiguous dates as ambiguous across all 5 documents
Every document tested triggered at least one false positive ambiguity flag from Gemini, including dates with explicit year values.
0 of 3
General AI tools that resolved relative dates like 'In 10 days' to a scheduled date
All three tools returned the raw phrase or null. Pawssier calculated the absolute date using the visit date as an anchor.
0 of 3
General AI tools that extracted microchip numbers from handwritten records
Microchip number (the primary persistent identifier linking a pet across providers) was missed by all three tools.
Documented failure patterns in general-purpose AI
Multi-record documents collapsed into one
Veterinary history PDFs often contain multiple visits in a single file. When we submitted a 5-page document with 5 separate invoices, all three tools returned a single record with no indication that pages had been merged. Five visit dates, five invoice numbers, and weight data tracked across months were all lost. The tool had no way to tell a platform that anything was missing.
Field-level decisions made with no explanation
When a document contains conflicting or ambiguous values -- multiple weights across visits, two phone numbers, an expired reminder alongside active ones -- general-purpose AI picks one and moves on. There is no confidence score, no flag, and no indication that a choice was made. Pawssier returns a confidence score and a reason code for every field so platforms know which values to trust and which to review.
Output format varies across documents
In testing across 5 documents, Gemini flagged unambiguous dates as requiring review in every single case. Claude returned output wrapped in markdown formatting in 3 of 5 documents, which breaks any parser expecting clean JSON. Both behaviors were consistent across the test set, not limited to specific documents.
Side-by-side extraction comparison
Document type: A 5-page PDF containing 5 separate clinic invoices from the same practice for the same patient across multiple visits from December 2024 to February 2025. Each page is a distinct record with its own invoice number, visit date, and service items.
Summary: All three general AI tools collapsed 5 separate visits into a single record with no indication that data was merged. The weight discrepancy is the clearest example: the puppy grew from 9.1 lb to 24 lb across visits, but all three tools returned only the most recent weight as if it were the definitive value.
| Field | Pawssier | ChatGPT | Gemini | Claude |
|---|---|---|---|---|
| Records detected | 5 (segmented by invoice number + date change, 0.98 confidence) | 1 | 1 | 1 |
| Multi-record awareness | Yes — segmented by invoice number + date change | No | No | No |
| Pet name | Pippin | PIPPIN | PIPPIN | PIPPIN |
| Species | Canine | Canine | Canine | Canine |
| Breed | Mix, Poodle | Mix, Poodle | MIX, POODLE | Mix, Poodle |
| Weight | 9.1 lb (Dec 2024 visit) → 24.0 lb (Feb 2025 visit) | 24.0 lbs (last page only) | 24.0 pounds (last page only) | 24.0 pounds (last page only) |
| Visit dates | 5 dates: 2024-12-16, 2025-01-06, 2025-01-26, 2025-02-09, 2025-02-13 | 2025-02-13 only | 2025-02-13 only (flagged ambiguous ✗) | 2025-02-13 only |
| Invoice numbers | 346366, 347467, 348599, 349353, 349570 | Not extracted | Not extracted | Not extracted |
| Reminders | 4 structured (with categories + confidence); visit instructions excluded | 6 mixed (includes visit instructions as reminders) | 8 raw text strings | 6 mixed (includes visit instructions as reminders) |
| Clinic name normalization | Sunset Paw Animal Clinic | Sunset Paw Animal Clinic | SUNSET PAW ANIMAL CLINIC (raw OCR) | Sunset Paw Animal Clinic |
Methodology
- 5 documents tested
- Document types: multi-visit combined invoice (5 pages), discharge summary (3 pages), multilingual vaccine booklet (image/JPG), payment history invoice with expired reminders (2 pages), handwritten vaccination and health record
- Tools tested against: ChatGPT (GPT-4o), Claude (claude-sonnet), Gemini (Gemini 1.5 Pro) — using each tool's standard chat interface with a consistent prompt
- All documents were real veterinary records, anonymized before publication
- Confidence explanations and reason codes referenced on this page are part of the PSRF v1.5 specification (preview — schema and reason codes are documented, but the v1.5 parsing logic is not yet wired into the production extraction pipeline, which still returns
api_version: "1.4.0")