AI month end reporting preparation.

What month end preparation looks like with AI

Month end preparation is the work between the last bank feed of the month and a management pack the client can read: reconciling, checking, adjusting, comparing and explaining. AI helps with two of those, checking and explaining, provided the system is designed so that software does the arithmetic and the language model only describes results it has been handed.

The pattern that works is a nightly job. Once the period closes, code pulls the trial balance, aged receivables and payables, and bank reconciliation status. It runs the practice’s close checklist as rules. When the checks pass, or the failures are logged, it calculates variances against the prior month, the same month last year and budget, then asks a model to draft commentary on anything above the materiality threshold. The accountant opens a pack that is mostly assembled, with the open questions at the top.

How the workflow runs

Inputs

Trial balance and comparatives from Xero or MYOB, bank reconciliation status, aged receivables and payables, budget, last month’s approved commentary, the client’s materiality threshold, and the practice’s close checklist.

Systems

The ledger API with read-only access, a small reporting database or spreadsheet engine for calculations, a model API, and the practice’s report template in Word, Excel or Power BI.

Steps

Pull data after the period closes; run close checks; calculate variances in code; select movements over threshold; draft commentary with account references; check every number in the draft against the report; assemble the pack.

Human review

The client accountant clears failed checks, posts any adjustments, rewrites commentary that is wrong or too confident, and approves the pack. A manager reviews packs for clients with lending covenants.

Outputs

A draft management pack, a close-check log, a list of questions for the client, and a record of who approved it.

Measures

Days from period end to pack sent, editing minutes per pack, share of commentary sentences changed, and close checks failing more than once.

Risks

Invented explanations, numbers copied from the wrong column, data pulled before feeds finished, and stale budget figures.

Illustrative example: a Nelson practice with 25 monthly packs

Illustrative example, synthetic numbers: a 9-person Nelson practice prepares monthly packs for 25 clients. Done by hand, each pack takes about 3 hours: 20 minutes pulling reports, 60 minutes on checks, 60 minutes writing commentary and 40 minutes formatting.

Model cost assumption: about 20,000 input tokens per pack (trial balance with comparatives, variance table, last month’s commentary and instructions) and 1,500 output tokens, drafted with Claude Opus 5.5 at US$4 per million input and US$20 per million output tokens, as listed on Anthropic’s pricing page on 19 October 2026. That is about US$0.11 per pack, or under US$3 for all 25 clients. Three redrafts per pack would still cost less than a coffee. The minutes in the table are what the practice should track, because they decide whether the workflow is worth keeping.

NZ timing and software details

• GST deadlines shape the close. A GST return is due by the 28th of the month after the taxable period ends, with exceptions for periods ending 31 March (due 7 May) and 30 November (due 15 January), and IRD does not grant extensions for filing. Schedule the month end checks so GST control account differences surface well before the 28th.
• Ledger features first. As listed on 19 October 2026, Xero’s JAX answers plain-language questions from live data, and MYOB’s AI Business Insights is in beta for eligible plans. Use them for client self-service; use a practice workflow for consistent packs and review records. Our Xero integration guide explains read-only access.
• Client data stays scoped. Each run should see one client’s file only. Mixing clients in one prompt is how one client’s figures end up in another’s commentary.

What the accountant still owns

The accountant decides which adjustments to post, whether an accrual or provision is needed, and what the client is told. Commentary about cash runway, covenant headroom, tax exposure or the owner’s drawings is judgement, and it should be written by the person accountable for the relationship. Our lesson on using AI to prepare management reports covers the editing habits that keep drafts honest.

Where month end automation breaks

• Plausible but wrong reasons. A model asked “why did revenue fall” will offer a reason. Require each explanation to point to accounts or transactions, and mark anything without evidence as a question for the client.
• Early data pulls. If the job runs before the last bank feed arrives, the pack reports a false cash position. Gate the run on reconciliation status, not the calendar.
• Number drift. Drafts sometimes round or transpose figures. A validation step that extracts every number from the commentary and matches it to the report catches this; see structured outputs and validation.
• Restated comparatives. If last month was adjusted after the pack went out, variance commentary is calculated against the wrong base. Snapshot each approved period.

Month end preparation depends on clean coding and complete documents, so it pairs with invoice coding and exception review and client document collection. For the bigger picture, see AI for accounting in New Zealand or our AI reporting service.

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