AI and documents
AI, OCR and your documents
When you send Bully a bill, something has to read it. Here's exactly what that is, what leaves us, and what we do when it gets something wrong.
What reads your documents
One service: Google Gemini, reached through the Lovable AI gateway. It is used for reading bills and statements, reading a keypad or meter photo, classifying a document that arrives by email, answering your questions in chat, and understanding a travel request.
There is no separate OCR company, no human transcription service and no other AI vendor. If that ever changes, this page changes with it.
Where the processing happens
Your document is sent from our servers to the gateway over an encrypted connection and processed on Google's infrastructure. Processing may take place outside the UK and EEA under the transfer safeguards in the providers' own terms. Everything we store stays in our own UK/EU-region database and storage.
Is your data used to train models?
No. We use these services on paid API terms, which exclude your content from model training. We do not licence, sell or share your documents with anyone for training, research or advertising.
What we keep, and for how long
- Documents you upload or email in: kept in your filing cabinet while your account is open, because the point of a filing cabinet is that things stay in it. Delete any one of them yourself, at any time.
- Keypad and meter photos: never stored at all. We keep the number that was read, not the picture.
- Emails to your filing address: the message record is kept for 12 months, then removed automatically.
- Chat messages: kept so Bully remembers what you told it, and deleted with your account.
- Page-visit measurement: kept 90 days, then removed automatically, and only collected if you allowed it.
Documents are not deleted automatically while your account is open. Close your account, or ask us to delete your data, and they go — files and all.
When the reading is wrong
It happens. A crumpled bill, a glare on a keypad, a supplier who changed their layout. So nothing extracted by AI is treated as fact:
- Every extracted field arrives marked as unconfirmed and carries a confidence score.
- Anything we're not sure of is labelled Not certain on screen, with the value beside it so you can see what we read.
- You can correct any field in one tap. Your correction wins from then on and overrides future readings.
- Where a new document disagrees with something you already confirmed, we raise it with you rather than quietly overwriting it.
- A reading too poor to trust is thrown away and we ask you instead of guessing.
- An incorrect figure that has already been used in a comparison is a correction request — tell us and we'll fix the record and re-run it. You can also request a correction formally.
Information, not advice
Bully tells you what a document says and what the numbers work out to. That is information, and we label it as such. It is not financial, insurance, legal or tax advice, and it is not a personal recommendation.
You'll see the difference in the wording. "Your bill says your renewal is 14 March" is something read off a document. "This tariff would cost you an estimated £120 less over a year, based on the usage you gave us" is a calculation with its assumptions attached. Bully will never tell you a product is best for you, guaranteed, or the right choice — because that would be advice, and giving it isn't what we do.
What goes to anyone else
Nothing extracted from your documents is passed to a supplier or provider unless you approve that specific application. The full list of who receives what is on the data sharing page.