Three of our products look at photos: a plant on a windowsill, a child’s dog, a supplier’s invoice. In all three the AI runs on the device that took the photo, and the photo is never uploaded. Here is how, and what it cost us.
PlantSync: naming the plant on the phone
PlantSync identifies a houseplant from a photo. The model runs on the phone with ONNX Runtime, so the identification works offline and the photo stays where it was taken. The store version carries no key for a cloud identification service at all, which makes the privacy promise easy to check.
Running a model inside an Android app has its own chores. Google Play now requires native libraries that support 16 KB memory pages, and the plugin we started with shipped older builds that did not. Play refused the upload until we swapped in current ONNX Runtime libraries.
Pletto: a photo becomes a colouring page
Pletto is a colouring app for children aged two to six. One tap turns a photo into a colouring page, and the subject of the photo into a sticker. A small neural network does the drawing on the phone’s graphics chip.
We tried a cloud image model first, and its drawings looked better. We chose the on-device route anyway. A children’s app that promises nothing leaves the device should not need a consent flow, a server holding an API key, or a longer data-safety form. Pletto does not even ask Android for internet access.
Inventyl: reading invoices in the browser
Inventyl restocks a shop from a photo of a supplier invoice and reads barcodes from the camera. Both run inside the browser, so the photo of the invoice, with its prices and supplier details, never leaves the device.
What on-device costs
- Size. Models ship with the app or load into the page, which makes downloads bigger.
- Quality. Small models that fit on a phone can be less accurate than the largest cloud models. Pletto is the clearest case.
- Speed varies with the device. An old phone takes longer than a new one, so the interface has to wait gracefully.
- Updates arrive with app updates, not instantly on a server.
What it buys
- Privacy you can show. No upload means no storage, no retention policy and no vendor reading the data.
- It works offline: on a shop floor, in a basement, on a plane.
- No cost per request, however much it is used.
- Simpler compliance. Less personal data processed means fewer obligations under privacy law and shorter answers on store data-safety forms and customer questionnaires.
When to choose it
On-device is the right call when the input is personal or confidential, when the task is narrow enough for a small model (recognising, classifying, extracting), and when the product has to work without a connection. The cloud is better when the task needs the largest models, changes daily, or depends on data that only exists on a server.
If you are weighing that choice for your own product, our AI & Intelligent Automation work covers both routes, and our AI Governance & Compliance service covers what each one means for privacy law and the EU AI Act.


