The AI never knows whose books these are.

The rule the product is built on

Code does every number. A model only reads, maps or drafts, on masked data. A person in your firm confirms. Totals, ties, casts and journals are calculated in code, so they can be checked and they come out the same every time. The model is used where judgement on wording helps: matching a ledger account to your chart, explaining why an entry looks unusual, drafting a query to the client.

Masked before any AI call

Before anything leaves PaperHQ for a model, identifiers are found in code and replaced with stable tokens: names of people and companies, bank account numbers and sort codes, UTRs, VAT and company numbers, postcodes and email addresses. Fenwick Timber Ltd becomes ORG_017 on the way out, and your reviewer reads Fenwick Timber Ltd again on the way back.

• Ledger codes and balances are not sent when a model suggests a mapping; only the masked account name and your firm's chart.
• Amounts are scaled by a private factor when a model reviews entries, so real values are not sent by default.
• The client's name and year end are not part of any request.

Every payload re-checked

A second check scans each request just before it is sent. If anything still looks like an identifier, the item is dropped or the request is not sent at all. A failed or blocked call never stops the work: rules, name matching and your team carry on without it.

Encrypted, and purged at sign-off

The pairs of token and real name are kept in an encrypted vault for each engagement. The vault is deleted when the partner signs off, when the engagement is deleted, or when someone in the firm clicks Purge. After that, the tokens in any log cannot be turned back into names.

Every call logged

Model calls are recorded in the engagement: which step made them, when, and what kind of data they carried, so a reviewer can see what an AI was shown for a file.

Works without AI at all

With no model configured, nothing is sent anywhere and PaperHQ still prepares the pack: rules, prior-year mappings and name matching do the work, and more is left for a person to confirm.

Regulation

Matching invoices, mapping codes and drafting notes are minimal-risk uses under the EU AI Act. UK GDPR and professional confidentiality do apply; masking turns them from a promise in a contract into a property of the system. Pilot firms get the full technical assessment, including the known limits of the name recogniser. This is not legal advice.

Security questions, or a request for the technical assessment: founders@paperhq.ai.