AI & privacy

No model ever calculates your money

A language model reads your message and proposes what it thinks happened: three wigs, fifty thousand each, part paid. Every figure after that is computed by ordinary, tested code, in whole kobo.

This matters because models are confident when they are wrong. Rekoda’s arithmetic is the kind you could check with a calculator, and the same code produces the number on your screen and the number on your invoice, so the two can never disagree.

Supported customer identifiers are replaced before any AI reasons about your business

Rekoda detects and tokenises supported personal identifiers before text reaches the AI that does your bookkeeping: recognised customer names, phone numbers, email addresses and supported financial identifiers are swapped for meaningless tokens. The model sees CUSTOMER_7, not Ada, and not her number. The real values stay in an encrypted store that the model layer cannot read. This is a precise promise about those supported identifiers, not a claim that every possible piece of personal information in free text is always detected: something unusual you type into a message, such as a full street address in an unexpected format, may reach the model as written.

One step works on raw material and cannot be tokenised first: turning a voice note or a photographed receipt into text. A recording is sound and a photograph is pixels, and whatever is spoken or printed in them reaches the transcription processor as it is. The two sections below name those processors and exactly what they are allowed to do; the text they return is tokenised like any typed message before the bookkeeping AI reads it.

The first time you mention a new customer by name, the model reads that one message to understand it. The name is then stored encrypted with its own token, nothing keeps a readable copy, and every later message that names them is protected before it leaves.

The substitution runs on the way out and is reversed on the way back, so your receipt says Ada while the provider only ever saw a token.

How a voice note becomes text

Your audio is sent to OpenAI, a transcription processor, for one purpose only: to come back as the sentence you spoke. Under the API terms we use, what we send is not used to train their models. The sentence is then treated exactly like a typed message: recognised customer names, phone numbers and other supported identifiers are replaced by tokens before any AI reasons about it, and we do not keep the audio.

An earlier version of this page said transcription ran only on infrastructure we run, and promised that if that changed, this page would change first and name the provider. This is that change, kept in those words: OpenAI is the transcription processor, it is the only one, and there is no automatic switching to anywhere else. If voice transcription is switched off for a deployment, voice notes are answered with an honest sentence and the audio is sent nowhere at all.

How a photographed receipt is read

The picture is sent to Anthropic with a single instruction: transcribe what the paper says, word for word, and nothing else. Under the API terms we use, what we send is not used to train their models, and we do not keep the picture. The AI that does your bookkeeping never sees it: it sees the transcribed text, with recognised customer names, phone numbers and other supported identifiers already replaced by tokens, exactly as a typed message is.

There is no quiet rerouting. If the reader is busy or cannot make out the page, we tell you and ask you to type the amount, rather than trying somewhere else you were never told about. Anthropic is the document reader, it is the only one, and which features are on is fixed when the system starts, never chosen per message.

Your business is not training data

Nothing you send is used to train a model, ours or anyone else’s. If we ever want to improve accuracy using real examples, we will ask for that specific permission, separately, and it will be genuinely optional.

Nothing is recorded until you say so

Rekoda shows you what it understood and waits. Where it is unsure it says so plainly rather than guessing quietly, and an unclear message becomes a question, not an entry in your books.