AI invoice coding and exception review.

What AI invoice coding does in a practice

AI invoice coding turns a supplier bill into a proposed entry: supplier, date, amounts, account code, tax rate and tracking category, each with a short reason. Exception review is the half that matters. The system decides which proposals are safe to save as drafts and which need a person, and it says why.

Start by checking what your ledger already does. As listed on 19 October 2026, Xero offers JAX and AI Capture to digitise and categorise bills, with actions logged for review, and MYOB is progressively rolling out Smart Reconciliation. For simple books, those may be enough. A practice-level workflow earns its place when one team codes for 30 or more clients, each with its own chart of accounts, and the partners want one coding policy, one exception queue and evidence of review across every file. That is a document processing problem with a ledger at the end of it.

How the workflow runs

Inputs

Supplier bills from a shared inbox or supplier portal (PDF, email body, phone photos), the client’s chart of accounts and tracking categories, the last 12 months of reviewed coding for each supplier, and the practice’s written coding policy (capitalisation threshold, entertainment, private use).

Systems

Xero or MYOB through their APIs with permission to create draft bills only, the practice’s document store, a review queue, and a model API on terms that exclude training on client data.

Steps

Extract fields; check totals and the supplier’s GST number; match the supplier to its history; propose account, tax rate and tracking; score confidence; apply the policy rules; save as a draft or raise an exception with a reason.

Human review

A staff accountant clears the exception queue daily. A senior reviews new suppliers, bills over an agreed amount (say $5,000), anything that may be capital, and a weekly 5% sample of auto-drafted bills.

Outputs

Draft bills carrying the proposed code and a one-line reason, exception tickets, and a weekly correction report for each client.

Measures

Exception rate, correction rate on auto-drafted bills, review minutes per bill, and days from receipt to coded.

Risks

Wrong GST treatment, duplicates, credit notes read as bills, payment redirection fraud, and errors copied forward from old coding.

Illustrative example: a 12-person Hamilton practice

Illustrative example, synthetic numbers: a practice codes 2,400 supplier bills a month across 40 bookkeeping clients. Coding by hand takes about 2.5 minutes a bill. In a back test the team accepts a 70% auto-draft rate; exceptions are the harder bills and take 3 minutes each.

The fee line assumes 4,000 input tokens per bill (page image, supplier history, chart excerpt and instructions) and 400 output tokens, on Claude Sonnet 5.5 at US$2 per million input tokens and US$10 per million output tokens, as listed on 19 October 2026. That is about US$12 per 1,000 bills. Model fees are trivial next to review time, so spend effort on making exceptions quick to clear, not on shaving tokens.

NZ rules the workflow should check

• Taxable supply information. Since 1 April 2023, IRD’s rules set what records must show by value: over $200 needs the seller’s GST number; over $1,000 also needs the buyer’s name and an identifier such as an address or NZBN. The workflow flags bills that lack them before the GST claim relies on them.
• Record keeping. IRD expects records to be kept for at least 7 tax years, and records stored offshore, including in cloud services, need IRD approval held by you or your provider. Keep originals in your own document store and treat the model as a processor, not an archive.
• API limits. Xero allows 60 calls a minute and 5 concurrent calls per connected organisation, with daily caps of 1,000 calls on its starter tier and 5,000 above that (Xero Developer). Batch draft creation and do not poll. Our Xero integration guide covers the details.
• Privacy. Bills from sole traders contain personal information. The Privacy Commissioner’s guidance on AI and the information privacy principles is the baseline for where it is processed and who sees it.

What stays with a person

People keep every decision that moves money or changes the tax position: approving bills for payment, accepting a new supplier or changed bank account, capital versus expense calls, GST treatment of mixed, imported or second-hand supplies, and sign-off of the GST return. Clients under IRD review should be excluded from auto-drafting altogether until the review closes.

Where invoice coding goes wrong

• Old mistakes become policy. If last year’s bills were miscoded, the history lookup repeats the error confidently. Learn only from reviewed coding, and re-review a supplier when its proposal is corrected twice.
• Duplicates. The same bill arrives by email and through a supplier portal. Match on supplier, invoice number, date and amount before drafting.
• Credit notes and statements. A credit note read as a bill flips the sign; a statement read as an invoice double counts. Classify the document type first and send anything ambiguous to review.
• Bank detail changes. Fraudsters send convincing bills with new account numbers, and text inside a PDF can try to instruct the model. Never let a bill change payee details, and read our note on recognising prompt injection.
• Uncalibrated confidence. A model’s confidence score is not a probability. Check correction rates by confidence band each month and move the threshold, as described in our guide to human review queues.

Invoice coding is usually the first of three accounting workflows we map. The others are month end reporting preparation and client document collection; the accounting sector guide shows how they fit together.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.