The art of 'Auto data extraction from Images using ServiceNow AI'

[SujanDutta]

If you've ever submitted an expense report, you know the drill: upload the receipt, then manually type in the vendor name, the amount, the date, the transaction type — all information that's already sitting right there on the receipt. And then someone on the finance team has to eyeball both the receipt and the form to make sure everything matches. In 2026, that workflow is long overdue for an upgrade.
In this episode, Tushar Mishra — a ServiceNow MVP with over a decade of experience and deep expertise in AI integrations — walks through a solution he built that reduces the entire claim submission process to a single click.

How It Works
The user opens a ServiceNow catalog form (on mobile or desktop), attaches a photo of their receipt, and hits submit. That's it. Their job is done.
Behind the scenes, a business rule fires on record insertion. It grabs the attachment, converts the image file to base64 (since you can't send a file directly over a REST call), and sends it to OpenAI's multimodal vision model. The model reads the image — even if it's blurry, crumpled, or in a different language — and returns all the extracted text back to ServiceNow.
At that point, an analyst can click a "Populate Claims Data" button, which triggers a Now Assist analysis skill. The skill takes that raw extracted text and maps it into structured fields: transaction amount, date, type, vendor name, and location. The prompt is remarkably simple — just one line of instruction plus a format example — because Now Assist is fine-tuned on ServiceNow's own data and understands the platform's table structures natively.

Customisation Is Straightforward
Need to extract different fields? Just update the analysis skill prompt. Want to add a custom field like "cost center" or "project code"? No additional coding required — adjust the skill, and the structured output adapts. That's the advantage of building on Now Assist rather than rolling your own parsing logic.

Multi-Language Support Out of the Box
Tushar demoed a Spanish-language receipt during the session, and the system handled it without any configuration changes. OpenAI's multimodal model understands multiple languages natively, so receipts in Spanish, French, or other languages are extracted and parsed just the same. For teams that deal with vendors across geographies, this is a big deal.

A Quick Word on LLM Gateway
One of the practical challenges with AI integrations is managing different payload formats across providers. OpenAI expects one structure, Anthropic another, AWS Bedrock yet another. ServiceNow's LLM Gateway solves this by giving you a single standardized payload format. You connect your models to the gateway, and it handles the transformation. One integration to maintain instead of many — and it gives you a centralized point for security and governance over which models your instance talks to.

Why This Pattern Matters
This isn't just about expense receipts. The architecture — attachment intake, multimodal extraction, AI-powered structured parsing — is a pattern you can apply to invoice processing, onboarding document verification, insurance claims, or any workflow where humans are currently re-keying information that already exists on a document. The building blocks are all platform-native, and the development effort is surprisingly light.

Check out the full episode for Tushar's complete walkthrough, including the business rule code, the REST message configuration, and the Now Assist skill prompt.