Automation Policies and Materiality: Deciding What May Skip the Queue

There are two lazy answers to "how much of my bookkeeping should the AI handle on its own?" One is everything — which is negligence with a dashboard. The other is nothing — which buries a reviewer in routine confirmations until their attention, the scarcest resource in the whole system, is spent on the transactions that least need it.

The honest answer is: it depends on the account, and someone accountable should decide, explicitly, in advance. That is what automation policies are in our platform — not a global autopilot switch, but a per-account agreement about what may skip the review queue and under exactly what conditions.

Three modes, per account

Every account in your chart of accounts runs in one of three modes:

Manual. Every draft touching this account waits for a person, regardless of how confident the model is or how small the amount. This is the right mode for anything sensitive — subcontractor costs, director transactions, anything tax-inflected or judgement-heavy — and it is the default for every account until you and your accountant agree otherwise.

Accept-with-edit. The middle mode, for accounts where drafts are usually right but details often deserve a touch — a memo clarified, a split adjusted. Entries are presented expecting quick confirmation; the reviewer amends what needs amending and confirms in one step, and every amendment is recorded against the original draft. The person is still in the loop; the loop is simply shaped for speed on work that has earned it.

Automatic, within limits. For accounts with a long record of boring correctness — the same monthly charge from the same supplier landing in the same account, month after month — drafts may post without queueing, but only when they clear both gates described below, and always on the record. Auto-posted entries remain visible like any other entry, carry their full audit trail, and can be reopened, corrected, or reversed by a person at any time.

The two gates: confidence × materiality

Automatic posting is never a single-threshold decision. A draft must clear two independent tests:

The confidence gate. The model's confidence score for this draft must exceed the threshold set for this account. Confidence is the model's own estimate of its work — useful for ordering and gating, never proof of correctness — so thresholds for automation are set deliberately high, and per account rather than globally, because the same score means different things on different kinds of transaction.

The materiality gate. The transaction's value must fall below the materiality limit agreed for this account. Materiality is the accountant's concept doing the work here: how large would an error have to be, in this account, to matter to the accounts and the tax position? Above that line, no confidence score is sufficient — a person looks, full stop.

The two axes are deliberately independent, and both must clear. A 99-scoring draft for an unusually large amount queues, because size is risk regardless of certainty. A £3 draft the model is hesitant about also queues, because doubt is doubt regardless of size. Automation happens only in the quadrant where the model is sure and the stakes are small — and only on accounts where you have opted in.

Why small-value does not mean no-review by default

It is tempting to assume materiality alone should do all the work: below some de minimis line, why look at all? Three reasons, each learned the hard way by the accounting profession long before AI arrived:

Small errors compound. A miscategorisation pattern worth £4 a time is invisible individually and significant at four hundred occurrences — quietly distorting cost lines, margins, and anything downstream that trusts them. Reviewing small transactions is often how you find the systematic error while it is still small.

Small transactions are where problems hide on purpose. Fraud and abuse are materiality-aware; padding and duplicates gravitate to just below whatever threshold nobody examines. A review regime that publicly ignores everything under £N is an instruction manual for staying under £N.

New patterns arrive small. The first transaction with a new supplier, a new payment channel, a new expense type is frequently low-value — and it is precisely the transaction that sets the precedent the AI will imitate on the next fifty. Getting the first one right, with human eyes, is cheap insurance on all its descendants.

So the platform's default is manual everywhere, including trivial amounts. Auto mode is something an account earns — proposed on evidence from the correction history, agreed with you, and bounded by thresholds someone signed their name to.

Policies are agreements, not settings

A word on governance, because it is the difference between a policy and a preference. Automation policies in our platform are recorded and versioned: what mode each account runs in, what the thresholds are, who agreed them, and when they changed. When an auto-posted entry is questioned — by you, by an accountant, by HMRC — the trail shows exactly which policy version, threshold, and score permitted it. The system also never re-tunes itself: thresholds move only when a person moves them, and that move is itself a logged event.

This is the shape of the whole platform in miniature. AI drafts. AgentLedger validates. People approve — and where approval is delegated to a policy, the policy is explicit, bounded, evidenced, and owned by a human being who can be asked about it. Automation you cannot explain is risk. Automation you can reconstruct, decision by decision, is just well-managed work.

Check eligibility / Ask for Details