AI Checks on Every Invoice and Claim Just Got Cheap (2026)

Short answer: many businesses check a sample of their invoices, claims or support tickets because checking all of them takes too many people. On 11 November 2026 Google released an AI service built only for closed questions (yes/no, pick-one and graded checks), priced low enough that running a dozen checks on every single document costs very little. The work that remains is deciding which checks matter and what happens to the documents that fail.

What Google released

Google's Gemini Decisions API runs a new model called Gemini Decision 1. It doesn't write answers. You hand it one input, which can be text, images, audio, video or PDFs, along with up to 1,000 separate questions, and it scores each one.

The questions come in three kinds:

• Yes/no: "Is the GSTIN on this invoice filled in?"
• Pick one: "Which team should handle this ticket: billing, delivery or product?"
• Grade on a scale: "How bad is the damage in this photo, from none to severe?"

Every answer carries a confidence score between 0 and 1. Google puts the typical end-to-end time at 150 to 250 milliseconds for inputs up to about 16,000 tokens, and the price is $0.10 per million input tokens, with zero output tokens billed. The input is billed once per request no matter how many questions you ask.

Why the price changes the job

Sampling exists because checking is expensive. When every invoice needs a person to read it, a busy accounts team checks a slice and hopes the rest are fine.

Here's the rough arithmetic on the new rate. Google's Gemini documentation counts each PDF page as 258 tokens. Ten thousand one-page invoices come to about 2.6 million tokens, roughly 26 US cents at the paid rate, plus the tokens for your questions. Even if your real number ends up several times that, it's small next to an hour of anyone's time.

That makes a different setup possible: the machine reads every document, and people read only the ones it flags.

Where it fits in a business

This works for any checking job you could hand to a careful new joiner with a written checklist:

• Accounts payable: does the invoice have a PO number, does the vendor name match the PO, is the GST break-up shown, is it signed?
• Claims and returns: does the photo show damage, does the damage match the complaint, is the item the one that was sold?
• Support tickets and emails: which queue, how urgent, is the customer threatening to leave?
• Sales calls: did the rep mention the payment terms, did the customer raise a competitor? Audio goes in directly.
• Site and store photos: is the shelf stocked, is the safety signage up, is the work finished?

Google's launch post gives pairing a scanned PDF claim with damage photos in one request as an example. That's close to what an insurer's back office, a logistics firm or an electronics retailer handles every day.

What it won't do

Google's Decisions API guide is direct about the limits:

• It doesn't extract or write. If you need the invoice amount pulled out into your ledger format, or a reply drafted to the customer, the guide points you to Google's Interactions API with structured outputs. Many workflows will use both: one to read the fields, this one to run the checks.
• Pick-one answers always pick something. The guide notes that choice questions spread probability across your options "even when none fits". Add an "other" option or a separate yes/no check for whether the document is in scope.
• Leave sums and dates to your own code. The guide says to compare numbers, dates and durations in code and pass in the result. Totals, due dates and duplicate checks belong in your accounts software, not in a question.
• It's on the Gemini Developer API only for now. Support for Gemini in Google Cloud is "coming soon", per the launch post. Batch processing and caching aren't supported, so repeated input counts in full each time.

There's one more thing for Indian owners. Google's pricing page says data on the no-cost tier is used to improve its products, and data on the paid tier isn't. Prototype with dummy or masked documents. Real invoices, customer photos and call recordings belong on a paid key.

How to set it up so it holds

1. Write the checklist first. Ask the person who checks today what they look for. Turn each item into one narrow question. The guide recommends keeping each question to a single decision and combining answers in your own code.

2. Label a few hundred real documents. Have your checker mark pass or fail on each item for a batch of past documents. Google's guide says to set confidence thresholds "using labeled examples from your own traffic". This batch is how you find out whether the checks match what your team would decide.

3. Decide what a wrong answer costs. Passing an invoice that should have failed can cost money. A wrongly flagged clean one costs a few minutes. Set the bar for each check to match, and send everything below it to a person.

4. Keep a person on the flagged pile and a small random sample of passes. The sample tells you if something has drifted, such as a vendor changing its invoice layout. Google's guide also says to treat input as untrusted and keep a human or a second check on high-stakes actions such as payouts.

5. Track one number. Errors caught per month before and after, or hours spent checking. If it doesn't move in a month, fix the questions before adding more.

We wrote earlier about where AI goes wrong in operations, and most of it applies here. A cheap check that nobody reviews can let a run of bad invoices through faster than any sampling routine.

The gap

Much of the talk about AI at work is about writing: emails, reports, posts. A big share of back-office work is sorting and checking instead, done by people who could be doing something better. That part just got cheap enough to run on everything, and it's easy to miss.