Walk through any back office and count what is actually machine-readable. Supplier invoices arrive as scanned PDFs. Field engineers photograph meter readings and damage. Finance circulates charts as screenshots. Legacy systems can only be captured by taking a picture of the screen. For decades the answer was OCR software — good at turning pixels into characters, poor at understanding what the characters mean. Claude's vision capability changes the deal: you send an image in the same Messages API call as your instructions, and the model reads and interprets it in one step.
What vision handles well
Scanned and photographed documents. Invoices, receipts, delivery notes, signed forms — including layout-heavy ones where the meaning lives in tables, checkboxes, and stamps rather than running prose. Because the same model that reads the pixels also understands language, you can ask directly for what you want: "extract vendor, total, and due date," not "give me all the characters."
Screenshots. Error dialogs attached to support tickets, dashboards, configuration screens from systems with no export function. "What error is shown and what does the surrounding context suggest caused it?" is a legitimate prompt.
Charts and diagrams. Claude can describe trends in a chart, read values from labeled axes, and summarize a process diagram. It also handles handwriting with reasonable — not perfect — accuracy, which matters for annotated forms and margin notes.
Vision is an input capability: Claude reads images; it does not generate or edit them.
How it works in practice
An image is simply another content block in a message. Your code base64-encodes the file (or references it by URL where supported) and places it alongside text instructions in the same request; multiple images per request are fine, so a five-page scanned contract can go in as five page images with one instruction. Vision is a core capability available on all four platforms — Amazon Bedrock, Google Vertex AI, Microsoft Foundry, and Claude Platform on AWS — so nothing about your cloud choice blocks it.
One platform caveat does matter for pipeline design: the Files API, which lets you upload a file once and reference it across many requests, is not available on Bedrock or Vertex AI. On those platforms you include the image bytes inline in each request. That is a plumbing difference rather than a capability difference, but it affects how you architect high-volume flows, so check the feature matrix before you design one.
The pattern that makes vision useful: pair it with structure
The highest-value vision workloads are extraction jobs, and extraction wants machine-readable output. The standard pattern is a single call combining both capabilities: the image goes in, and a schema (via the tool-use mechanism) forces the answer out as typed JSON — vendor as a string, total as a number, due date in a fixed format. From there it is an ordinary data pipeline: validate the fields in code, auto-process what passes, and queue what fails for a person. This image-to-schema-to-validation flow is described in detail in the structured outputs article, and it is the backbone of most document-processing systems built on Claude.
Pitfalls and honest limits
Garbage pixels, garbage output. Blurry photos, skewed scans, and low-resolution images degrade accuracy quietly — the model still answers, just less reliably. Set minimum quality standards at the point of capture where you can.
Confident misreads. A smudged 7 can be read as a 1 without any signal of doubt. This is why numeric fields deserve cross-checks (do the line items sum to the total?) rather than trust.
Dense, tiny text. A full broadsheet page photographed at a distance pushes limits. Splitting documents into page-level images generally beats one giant composite image.
Privacy still applies. Images of documents contain the same personal data as the documents. Apply the same minimization, retention, and access rules you would to text, and confirm data-handling specifics with your cloud provider.
Where to go next
See document processing for a complete pipeline built on vision plus extraction, and structured outputs for the schema half of the pattern. The feature matrix lists exactly what each platform supports today.