Why On-Device AI: Remove the Typing Without Giving Up Privacy

The real cost of an expense tracker isn't the subscription. It's the typing — and every fix for the typing so far has meant handing your financial life to somebody's server.

The trade nobody should have to make

There are two standard ways to stop people typing. One is bank linking: an aggregator takes your banking login, pulls your whole transaction history and keeps a copy of it. The other is cloud AI: you paste a bank message or photograph a receipt, and that text goes to a model running in a data centre.

Both work. Both also move the most detailed record of your life — where you eat, what you treat yourself to, who you pay, when you stopped — off your phone and onto infrastructure you don't control, to save you about thirty seconds per expense. Pocket Clear exists because a lot of people aren't willing to make that trade, and we weren't willing to make it on their behalf.

The third option only became practical recently: run the model on the phone.

What runs where

On an iPhone 15 Pro or later with Apple Intelligence turned on, Capture and Ask use Apple's on-device Foundation Models. On an Android phone where Gemini Nano is available, they use that, through ML Kit. Where neither is available — an older phone, or a country or language the system models don't cover — a rules parser built into Pocket Clear reads bank messages and card alerts with no model at all.

That last case matters more than it sounds. The most common thing anyone captures is a bank SMS, and a bank SMS is structured text: an amount, a merchant, a date, in a shape that repeats. A parser handles it. The model earns its keep on the messier input — an order confirmation email, a photographed receipt, a sentence you typed yourself.

Whichever path a capture takes, it happens on the phone it was shared to. Nothing about the message, the receipt or the question is uploaded — not to us, not to an AI provider. The expense you confirm is saved like any other transaction, and it leaves the phone only if you have Cloud Sync switched on.

It is also the honest economic answer

Reading every capture through a cloud model means a bill that grows with how often the feature gets used — exactly the wrong incentive for something meant to be used several times a day. On-device inference has no meter running. That is a large part of why this can be free for everyone rather than metered, rationed or reserved for the top tier.

What it doesn't do

• No cloud model. There is no fallback that quietly ships your text to a server when the on-device path can't handle it. If no model is available, the rules parser tries, and if that can't read it either, you get the normal form.
• No bank linking. Pocket Clear never asks for banking credentials and never connects to an account. It reads what you hand it.
• No training on your data. Your transactions, your captures and your questions are not used to train or improve any model, ours or anyone else's. There is nothing collected to train on.
• No leaderboards or benchmarks. Nothing compares your spending to other people's. Building that would mean pooling everyone's data on a server, which is the thing this whole design is avoiding.
• No background reading. Pocket Clear does not scan your messages or your photo library on its own. It sees what you explicitly share, paste or photograph, and nothing else.

Opt-in, Beta, off by default

All of this lives behind one switch on the On-Device AI tab, and that switch starts off. It's marked Beta, and it means it: English only for now, results vary by phone, and you always get to confirm or undo what it read. It's a Pro feature, and it's free for everyone until 31 December 2026.

The full walkthrough is in the help centre: On-Device AI — Capture and Ask.