How It Works: From Evidence to Approved Books

Our whole platform is one sentence long: AI drafts. AgentLedger validates. People approve. This page walks that sentence end to end, then shows it handling a real, slightly awkward receipt.

Step 1 — Evidence arrives

Your paperwork reaches us through whichever channel suits the moment: direct upload, a read-only connection to your email (official Gmail or Microsoft APIs, disconnectable in one click), a browser extension that collects invoices from vendor portals, or a photo sent over WhatsApp or Telegram. Bank statements come in for reconciliation alongside.

Every original is retained and every incoming file is checked for duplicates — the same invoice arriving by email and by photo is common, and suspected duplicates are held for a person to decide, never silently discarded or double-posted.

Step 2 — AI drafts an entry

The AI reads each document and proposes a proper double-entry transaction: date, counterparty, amounts, and the accounts to debit and credit. Alongside the draft it records a confidence score — its own estimate of how sure it is. Drafting is all the AI does here; it holds no power to post, pay, or file anything.

Step 3 — AgentLedger validates

Every draft must pass AgentLedger, our plain-text double-entry ledger kernel. The checks are mechanical and deterministic: postings must balance to zero, accounts must exist in your chart of accounts, dates must be sane, the entry must parse cleanly. Failures are rejected with a reason — never quietly repaired — so nothing structurally unsound ever reaches a reviewer, and reviewer attention is spent on judgement rather than arithmetic.

Step 4 — A person decides

Validated drafts queue for your accountant, ordered doubtful-first so the freshest attention lands on the entries the model was least sure about. The reviewer has three verbs, each recorded with who and when:

• Approve — correct as drafted; it posts.
• Correct — amend, then post; the original draft is preserved alongside the fix.
• Query — genuinely unclear; it is held and a question comes to you.

Where you have agreed a per-account automation policy, routine entries can post automatically — but only below an agreed materiality limit and above an agreed confidence threshold, always logged, always reversible. The default for every account is manual.

Step 5 — Reports you can stand behind

Profit and loss, balance sheet, trial balance, and self-assessment summaries are computed directly from the approved ledger — the same validated text, so reviewed figures and reported figures cannot quietly diverge. When Making Tax Digital submissions are due, they are prepared from those approved books and sent to HMRC only with your explicit authorisation. Nothing files itself.

A worked example: the £86.40 receipt

Tuesday, 9:40am: you photograph a hardware-shop receipt — £86.40 — and send it over WhatsApp.

The AI reads it cleanly: supplier, date, and total all extract with high confidence. But the line items are terse ("fixings, misc consumables, blade set"), and the right account is genuinely ambiguous — job materials if this was bought for a specific customer's work, tools and equipment if it is general kit. The model's confidence on the categorisation is low.

So the draft is held, not posted. It passes AgentLedger's checks — the entry balances and is well-formed — but low confidence on a judgement field sends it to the front of the review queue. Your accountant looks, agrees the receipt could honestly go either way, and sends you a one-line query: "Howdens £86.40 on the 8th — for the Harrogate job, or general kit?"

You reply "Harrogate job". The entry is corrected to job materials, approved, and posted. The full story — the photo, the AI's draft and its score, the query, your answer, the correction, the approver, the timestamps — stays attached to the entry permanently. If anyone ever asks why that £86.40 sits where it sits, the answer is on the record.

That is the system working as designed: not the AI being right, but the uncertainty being caught, routed to a person, and resolved on the record.

Why this shape

Every design choice above follows from one belief: your books deserve both the speed of machines and the accountability of people, and neither substitutes for the other. The AI removes the typing. The kernel removes the arithmetic errors. The human owns the judgement. And the audit trail means no one ever has to take any of it on faith.

Check eligibility / Ask for Details