On-Device vs Cloud Transaction Categorization
Smart categorization is easy to advertise and hard to evaluate. The useful questions are where descriptions are processed, what is retained, whether corrections teach the system, and what happens when the model is unavailable.
Last updated: August 16, 2026
On-device is an architectural claim
A categorizer can be a rules engine, a cloud model, or a model running locally on the phone. All three may produce the same category label; they differ in what leaves the device and what continues working offline. A trustworthy description names the processing location instead of relying on the word AI.
Storage is a separate axis. An app can categorize locally and still store account data in its cloud, or it can avoid cloud AI while using a hosted transaction database. Readers should evaluate both processing and storage.
What is verified today
The table deliberately uses unknown where public documentation does not establish the processing location. Unknown is more honest than converting automated categorization into an on-device claim.
| App | Categorization approach | Budget / transaction store | Confidence |
|---|---|---|---|
| Okane | Semantic categorization runs on the phone and learns from corrections | Budget in the user's Google Sheet | Verified by first-party product documentation |
| Actual Budget | Rules and local-first workflows; not the same packaged AI feature | User-controlled or self-hosted | Architecture documented |
| Tiller | AutoCat spreadsheet categorization | User's Google Sheet or Excel workbook | Processing path differs from mobile on-device inference |
| YNAB | Automatic categorization available | YNAB-hosted | On-device model location not verified in this dataset |
| Monarch / Copilot | Automated categorization available | Vendor-hosted | On-device model location not verified in this dataset |
Where Okane's design is stronger
Okane combines two unusual choices: the budget's working copy is a Google Sheet on the user's Drive, and Premium categorization runs on the phone. Imported transactions can be reviewed, corrected, and written to the Sheet without sending descriptions to a cloud language model for categorization.
That does not eliminate every external system. Plaid still handles linked-account access, Google hosts the Sheet, and Okane uses backend services for subscription and bank-sync plumbing. The precise claim is on-device categorization, not a network-free application.
How to evaluate an AI privacy claim
Ask for the model location, fallback behavior, retention policy, and learning mechanism. Then disable the network and test whether categorization still works. If the company cannot say whether merchant descriptions reach its servers, the feature should remain unverified.
- Does raw transaction text leave the device for categorization?
- Are corrections stored locally, in the budget file, or in a vendor account?
- Can the model be disabled without disabling manual budgeting?
- Does the app disclose the model download size and device requirements?
FAQ
Does on-device categorization mean the entire app is offline?
No. It describes where categorization runs. Bank sync, authentication, cloud storage, and subscription checks may still use network services.
Is a rules engine worse than AI?
Not necessarily. Rules are inspectable and predictable. A local semantic model can handle unfamiliar descriptions better, but accuracy and privacy should be tested rather than assumed.
Related app profiles
Sources
- Okane privacy and security — Okane. Accessed August 16, 2026.
- Okane on Google Play — Google Play. Accessed August 16, 2026.
- Actual Budget documentation — Actual Budget. Accessed August 16, 2026.
- Tiller AutoCat — Tiller Money. Accessed August 16, 2026.