AI Transparency Statement

Elizabeth Bookkeeping uses artificial intelligence. You should not have to reverse-engineer where, or wonder what it is allowed to decide. This statement sets both out plainly.

Where AI is used

• Reading documents. AI extracts details — dates, amounts, counterparties, line items — from receipts, invoices, bank statements, and similar evidence you provide through upload, connected email, our browser extension, or messaging.
• Drafting ledger entries. From that evidence, AI proposes double-entry transactions for your books. Every draft carries a confidence score: the model's own estimate of how sure it is.
• Ordering review work. Confidence scores are used to sort the human review queue so the most doubtful drafts get attention first, and to help flag items — such as possible duplicates identified by deterministic checks — for a person to examine.

That is the extent of it. AI in our platform does not move money, does not submit filings, and does not communicate with HMRC on its own initiative.

The oversight around it

Our operating rule is one sentence long: AI drafts. AgentLedger validates. People approve.

Every AI draft must first pass AgentLedger, our ledger kernel, which mechanically checks hard invariants — balanced postings, accounts that exist, sane dates — and rejects failures rather than repairing them. Every validated draft then meets a human decision: an accountant approves it, corrects it, or queries it under the review policy agreed with you.

The narrow exception is explicit automation: where you have agreed a per-account policy, routine entries below a materiality limit and above a confidence threshold may post without queueing. Those policies are recorded and versioned, such entries remain visible and reversible, and the default for every account is manual until you decide otherwise.

Drafted, not final

The principle underneath all of this: AI output in our platform is a proposal, never a decision. No AI output stands in your books on its own authority — it stands because it passed deterministic validation and a person (or a policy a person explicitly set) accepted it. Suspected duplicates are likewise held for a human decision, never auto-posted and never auto-discarded.

We pair that with honesty about limits. AI extraction and categorisation are good but not infallible; models can be confidently wrong. The controls above exist because of that fact, not despite it, and we do not claim error-free bookkeeping — we claim errors are made rare, visible, attributable, and correctable.

How to query a decision

Every entry in your ledger carries its full history: the source document, the AI's draft as originally proposed, its confidence score, any edits, any queries and answers, and the identity of the approver — or the exact policy that permitted an automatic posting. If you want to understand or challenge any entry:

1. Ask about the entry through the portal or your accountant. A person, not a model, will answer.
2. You will receive the trail — what the AI proposed, what changed, who decided, and why.
3. If it is wrong, it is corrected, and the correction joins the same permanent record.

The ICO publishes guidance on AI and data protection, linked in this page's sources, which informs how we approach transparency and accountability. Questions this statement does not answer are welcome at [email protected].