Snap a photo, skip the paperwork: AI that reads vaccine records and invoices

The newest decision models read photos for a fraction of a cent. A practical guide for kennels, groomers, and other local shops to checking vaccine records, supplier invoices, and intake photos, with a person still deciding.

Sections

It's the second week of December at a boarding kennel. Every run is booked from the 20th through New Year's. The front desk inbox has forty new messages, and a lot of them are the same thing: a photo of a vaccine certificate, taken on a kitchen counter, sometimes upside down, sometimes of the wrong dog.

Somebody has to open each one, find the pet's name, find the rabies expiration date, check it against the checkout date, then do it again for Bordetella. When a record is missing or expired, somebody has to text the owner, wait, and check the next photo they send. It's careful, boring work, and it piles up exactly when the phones are busiest.

This week two releases made that kind of work cheap to hand to software, at least the reading part. This post is about what changed, how I'd set it up for a kennel or a groomer, where else the same idea fits, and where a person still has to be in the loop.

What changed this week

Decision models are AI models that don't write paragraphs. You give them something to look at and a short question, and they give back a choice, a number, or a yes/no with a confidence score. I covered the developer side in Decision models, one month after Jev. The short version for a business owner: they're fast, they're cheap, and they're built for the small, repetitive judgments a front desk makes all day.

Two releases this week matter for paperwork:

Microsoft-Decision-1.1, November 17. Microsoft moved its decision model to general availability on Microsoft Foundry, OpenRouter, and Vercel. The new version accepts images, including photos and screenshots, and Microsoft says it runs more than twice as fast as the first version at about the same cost. Microsoft's chart puts that cost at roughly $140 per million decisions, which works out to about a penny and a half for every hundred. See Microsoft's announcement.

Liquid AI d1.2 and d1.2-ultra, November 16. Liquid added new kinds of questions its models can answer. The one that matters here is "read a value": point it at a document and ask for a name, number, or date, and it returns the value along with the characters it read it from. If the value isn't there, it answers "not found" instead of guessing. Liquid says d1.2-ultra scored 98% at reading values from receipts on its own test set, and it currently sits at the top of Artificial Analysis's decision model index. See Liquid's post.

That "not found" answer is the part I care about most. The biggest risk with AI reading documents isn't that it fails to read something. It's that it fills in a plausible date when the real one is smudged or cut off. A model that's built to say "I couldn't find it" is a much better fit for anything you're going to rely on.

One photo / Four small questions What the model actually reads off a certificate A made-up rabies certificate and illustrative answers. Each question has a short answer you can check against the photo.
  1. 1Pet name?BellaMatches the booking
  2. 2Which vaccine?Rabies, 3-yearFrom a fixed list
  3. 3Next due date?Dec 23, 2026Read, then compared in code below
  4. 4Tag number?not foundSmudged. It says so instead of guessing

The date math is plain code, not AI

To staffDue during the stay, and one field wasn't found. A person checks the photo and texts the owner for the updated record.

FIG. 01 An illustration, not real model output. The pattern is what matters: small questions, "not found" as a real answer, and dates compared by ordinary code.

A quick note on numbers: the accuracy figures above are the vendors' own. I haven't run either model against real kennel paperwork for this post. Your photos will be blurrier than any benchmark, so plan to test on your own.

What it costs to read a photo

Liquid prices images by size. A 1024 by 1024 pixel image counts as 1,536 tokens, and d1.2-ultra costs $0.042 per million input tokens, with no charge for the answer. That's well under a hundredth of a cent per photo. A thousand vaccine records would cost around six or seven cents to read.

Phone photos are bigger than 1024 pixels, so check how the API you choose handles large images, and shrink them first if it charges by size. Even with a few questions per photo, the model cost for a busy kennel's whole holiday season is less than a cup of coffee.

The real cost is setup: connecting the inbox or booking system, writing down the rules, and testing. That's a one-time job, and it's the same shape whether you have one location or five.

The kennel example, step by step

The mistake I'd avoid is asking the model one big question like "is this dog cleared to board?" That answer is impossible to check, and it quietly hands a health and liability decision to software. Instead, break the work into small questions, each with an answer you can verify.

The first job / A boarding reservation The model reads. Your code compares. A person decides. Each step is a small question with a short answer. Nothing in this path tells a customer their dog is cleared.
  1. 01What is this?Rabies certificate, Bordetella record, invoice, or something else. One choice from a fixed list.
  2. 02Read the fieldsPet name, vaccine, date given, date due. "Not found" is an allowed answer.
  3. 03Compare in codeDue date after checkout? Name matches the booking? Plain date math, not AI.
  4. 04A person confirmsAnything flagged or uncertain goes to staff with the photo beside the answer.

Send it straight to a person when

  • Any field is "not found" or low confidenceBlurry photo, handwriting, a cropped corner. Ask for a better copy rather than guess.
  • The name doesn't matchHouseholds with three dogs send the wrong certificate all the time.
  • Anything health-related is due or past dueStaff confirm with the owner or the clinic before the stay.
FIG. 02 The model never answers "is this dog allowed in?" It answers smaller questions you can check, and the decision stays with your staff.

1. What is this document? Give the model a fixed list: rabies certificate, Bordetella record, combined vet record, something else. A choice from a short list is exactly what decision models are good at, and "something else" catches the photo of the dog instead of the paperwork.

2. Read the fields you need. Pet name, vaccine name, date given, and expiration or due date. Ask for each one separately, and accept "not found" as a real answer.

3. Compare in plain code. Is the expiration date after the checkout date? Does the pet's name match the reservation? This is date math and text matching, and ordinary software does it perfectly. There's no reason to ask an AI model to compare two dates.

4. Sort into two piles. Records where every field was read with high confidence and everything checks out go into a "looks current" pile. Everything else, including any "not found," any low confidence score, any name mismatch, and anything expired or due during the stay, goes to a person, with the photo shown right beside what the model read.

5. A person confirms. Staff look at the flagged pile first, because that's where the follow-up calls come from. And at check-in, a person still glances at the record. The software did the reading and the sorting. It didn't make the call.

That last point is the one caution I'd underline. Vaccine records are health records, and in Florida rabies vaccination isn't optional. State law requires it for dogs, cats, and ferrets four months and older, with a booster 12 months after the first shot and then on the schedule the vaccine maker sets (Florida Statutes 828.30). That's why you'll see both one-year and three-year certificates. Bordetella is usually your own facility's rule rather than state law, but you set it for a reason. Any record the AI flags gets confirmed by a person before you tell an owner anything about it, and nothing the AI says counts as clearing a pet.

Where it goes wrong

A few things I'd test for specifically, because they're how real paperwork comes in:

Also treat these photos as customer data. A vaccine certificate usually has the owner's name, address, and phone number on it. Read the provider's data terms before you send anything, and don't keep copies longer than you need.

The chasing is where the time goes

Reading the records is only half the job. The other half is following up: the owner whose rabies shot expires on the 23rd of a stay that runs to the 27th, or the one who sent nothing at all.

Once the reading is automated, the follow-up can be too, with a person approving the message. The system drafts a short text: "Hi Dana, we have Bella's rabies record on file but it shows due December 23, which falls during her stay. Could you send the updated one?" Staff skim the drafts in the morning and send them in one batch. The owner replies with a new photo, and it goes back through the same check.

This pairs naturally with an AI receptionist that already answers texts. If you're thinking about that side, AI voice agents for the calls you're missing covers what to hand off and what to keep.

The same idea for other paperwork

Vaccine records are the clearest example for a kennel, but the pattern is the same anywhere someone reads a photo or PDF to pull out a few fields.

Supplier invoices. A pet boutique gets invoices from distributors as PDFs, photos, and emails. Read the invoice number, total, and due date, then let code match them against the purchase order and flag anything that doesn't line up. The person approves payment. Liquid's receipt numbers suggest this is the strongest use today, because invoices are printed, structured, and mostly the same layout from a given supplier.

Intake photos. Groomers often photograph a dog at drop-off. A decision model can sort those photos with simple yes/no questions, such as "is the coat heavily matted?", so the groomer sees the tough appointments first and can mention the possible extra charge before starting. That's a sorting job, not a diagnosis, and anything that looks like a skin problem or injury goes to a person to raise with the owner.

Trades and service shops. A plumber's tech photographs a water heater's label, and the model reads the model and serial numbers into the work order. An HVAC company reads the nameplate on a condenser. A salon or auto shop sorts customer-uploaded photos so the right person sees the right request. Same steps: a fixed question, a short answer, code that checks it, a person on anything uncertain.

Receipts for bookkeeping. Staff snap receipts for supplies, and the model reads vendor, date, and total so your bookkeeper starts from a filled-in row rather than a pile of photos.

What all of these have in common: the document already exists, a person currently reads it by eye, and the answer you need is a few fields, not a summary.

How I'd start

  1. Pick one document. For a kennel, rabies certificates for boarding. Don't start with everything.
  2. Write down the rules your staff already use. What counts as current, what name has to match, what happens when something's missing. This list matters more than which model you pick.
  3. Gather last month's records. Run them through the setup and compare the results with what your staff actually decided. Count the misses, especially any expired record that landed in the "looks current" pile. That number has to be zero before you go further.
  4. Run it next to your staff for two weeks. The software sorts, people still check everything, and you see how often they disagree.
  5. Then let it cut the work. Staff confirm the flagged pile and spot-check the rest, and the drafted follow-up texts start going out with a person's approval.

You don't need a new booking system for this. Most of the work is connecting the inbox or upload form you already have to a small service that asks the questions and puts the answers somewhere your staff already look.

Worth trying before the holidays

If your front desk spends the busy season opening photos and squinting at dates, this is one of the cheapest, most contained places to start with AI. The model costs pennies, the questions are small and checkable, and a person stays responsible for every decision that matters.

Setting this up is part of the workflow automation and integration work I do, whether that's connecting it to the booking software you already use or building a small intake form for it. I'm in Southwest Florida and work remotely across the US. If your inbox looks like that kennel's in December, tell me what paperwork you're checking by hand.

More notes

All notes · Contact