Ramp demos on‑device receipt scanning with Apple Intelligence

Ramp showed a live workflow that pulls receipts from photos already stored on an iPhone and matches them to card charges inside Ramp, without sending a user’s full photo library to the cloud. (x.com) The setup leans on Apple Intelligence, Apple’s system for running some artificial intelligence tasks directly on a device and sending harder requests to Apple’s Private Cloud Compute only when needed. Apple says Apple Intelligence first decides whether a request can stay on device, and sends only the data relevant to the request when cloud processing is required. (apple.com) For finance teams, the point is mundane and specific: receipts are often missing, delayed, or buried in camera rolls. Ramp already lets companies collect receipts by text, email, browser capture, and manual upload, and its help center says the product can automatically verify whether a receipt matches a transaction. (ramp.com) (support.ramp.com) That makes the demo less about a new consumer assistant and more about a narrow office task with a clear privacy constraint. Apple has spent the past two years arguing that on-device processing should be the default for personal data, while software vendors have been looking for features that fit that model. (apple.com 1) (apple.com 2) Apple widened that opening at Worldwide Developers Conference on June 9, 2025, when it said developers would be able to access the Apple Intelligence on-device foundation model inside their apps. Apple framed that release around “private, intelligent experiences,” not general-purpose chatbots. (apple.com) The separate debate around Apple’s machine learning stack is more technical but hits the same question: can developers ship these features reliably on Apple hardware. In a widely shared thread, developers pointed to model conversion into Core Machine Learning, Apple’s format for running models on its chips, as a recurring failure point between a working prototype and a production app. (x.com) Apple’s own documentation describes Core Machine Learning and Private Cloud Compute as pieces of a layered system: run what you can locally, and fall back to Apple silicon servers for larger jobs. That architecture gives companies like Ramp a privacy story, but it also means deployment depends on model packaging, compatibility, and device support lining up cleanly. (apple.com 1) (apple.com 2) Ramp is not pitching receipt capture as a side feature. Its main site says more than 50,000 finance teams use the platform, and its expense product markets “expenses that submit themselves,” which is exactly the kind of repetitive workflow this demo targets. (ramp.com 1) (ramp.com 2) The demo lands at a moment when Apple is still trying to prove that “personal intelligence” can produce everyday software features, not just polished demos. Ramp’s pitch was simple enough to test in public: find the receipt, match the expense, and keep the private photos private. (x.com) (apple.com)