What the AI actually does with your photo
You photograph an invoice; seconds later it’s a filed record. Here’s the honest, step-by-step account of what happens in between — where it runs, what’s kept, and what your documents are never used for.
Checked on 28 Jul 2026
General information, not legal or tax advice.
At a glance
- Your photo is processed once, by Google Vertex AI’s EU endpoint, in the same EU region where your data is stored
- Under Google Cloud’s data-governance terms, your content is not used to train AI models — ours or Google’s
- Nothing is saved without you: extraction produces a review screen, and you confirm or correct before anything is filed
- Results are cached against the file itself, so identical uploads aren’t reprocessed
- Every extracted field links back to the exact photo it came from — anything can be checked against the source
“AI reads your paperwork” is either the most useful sentence in this product or the most alarming one, depending on how much you’re told about what actually happens. Here is the whole pipeline, honestly, step by step.
Step by step
1. You photograph a document — an invoice, a certificate, a rating plate — or forward an email. The file is uploaded over an encrypted connection and stored, encrypted, in the EU.
2. It’s read once, in the EU. The image goes to Google Vertex AI’s EU endpoint — in the same European region where everything else lives — with one job: pull out the structure. Make, model, serial number, dates, amounts, the supplier’s name, what kind of document it is.
3. It’s compared with what you already have. The extraction is checked against your existing record — is this the boiler you already track? The provider you already use? A duplicate of a document you already filed? This is what stops a second photo of the same invoice becoming a second boiler.
4. You decide. Nothing enters your record automatically. The results appear on a review screen — every field editable, every suggestion declinable — and only what you confirm is filed. The AI proposes; you dispose.
5. The result is cached. The extraction is stored against the file’s own fingerprint, so uploading the same document again doesn’t send it for reprocessing.
6. The provenance stays attached. Every field in your record links back to the photo it was read from. Years later you can tap any date or serial number and see the exact document that produced it — which also means any mistake can be caught and corrected against the source.
What your documents are never used for
Under Google Cloud’s data-governance terms, content processed through Vertex AI is not used to train models — not Google’s, and not ours. There is no secondary use: no advertising, no resale, no “anonymised insights” product built from your paperwork. The AI reads your document for you, once, and that is the entire transaction.
Why this design
Most of these choices cost us something — EU-pinned processing, human confirmation, provenance links, caching — and each exists for the same reason: this product asks you to trust it with the paperwork of your home. That trust has to be inspectable. The infrastructure details live on the security page; the ownership promise — export everything, any time, every plan — is written down here.
Accurate as of the checked date above; if the pipeline changes, this post changes.
Common questions
- Is my document used to train AI models?
- No. Content processed through Google Vertex AI is not used to train models under Google Cloud’s data-governance terms, and KeepUp doesn’t train models on your documents either. The document is read once to extract its details, and that’s the entire use.
- Where does the AI processing happen?
- On Google Vertex AI’s EU endpoint, in the same European region (Belgium, europe-west1) where KeepUp stores your data. Your documents don’t leave the EU for processing.
- Can the AI file something wrong without me noticing?
- Nothing is filed without your confirmation — extraction produces a review screen where every field can be edited or rejected. And every filed field keeps a link to the photo it came from, so anything can be checked against the source later.