Short answer
AI reporting uses language models to draft commentary, answer questions about your data and assemble recurring reports, while the numbers themselves come from your accounting, sales and operations systems. Levy Street builds AI reporting for New Zealand organisations in which every figure traces to a source query, reconciliation checks run before anything is drafted, and a person approves what goes to the board.
- Models write the words; queries and code compute every number.
- Each figure in a report links to the query that produced it.
- Reconciliation checks against the ledger run before commentary is drafted.
- A named person approves anything that reaches a board or a client.
Where AI helps in reporting, and where it must not touch the numbers
AI is good at the words around reports and bad at being the source of the numbers. The rule we build to: models write the words; queries and code compute every number. A language model asked to total a column or recall last month’s revenue can produce a plausible, wrong figure. The Privacy Commissioner’s guidance on generative AI warns that these tools “often produce very confident errors of fact or logic”. So the model calls defined queries, receives results, and writes about them. Code inserts the figures.
Where AI does help:
- Commentary drafts. Variance explanations, trend summaries and the “what happened this month” narrative, written from query results and operational notes.
- Questions in plain English. “Which customers’ orders fell more than 20% this quarter?” answered by generating or selecting a query, showing it, and running it.
- Assembly. Pulling figures, charts and commentary from several systems into your board pack format.
- Flags. Unusual movements, missing data and late postings highlighted for a person to investigate.
Who it suits
- Month-end takes days. A finance lead spends two or three days assembling the management pack from Xero or MYOB, a CRM and spreadsheets.
- The board wants more commentary. Directors want explanations, not just tables, and nobody has time to write them.
- People keep asking for numbers. Managers email finance for figures that a well-built question tool could answer safely.
- Client reporting. Agencies and advisers produce similar reports for many clients every month.
The reporting demo and reconciliation checks
Trust in AI reporting comes from two things: proving it on reports you have already published, and checking every figure against the ledger before any words are written.
The source-grounded reporting demo
Before a build, we take the last three months of one recurring report and regenerate it from source. The demo shows each regenerated figure beside the one you published, each sentence beside the query results it relies on, and every discrepancy explained. Sometimes the discrepancies are in the published report. Either way, you see exactly how the system behaves on your data before committing. Our lesson on using AI to prepare management reports covers the manual version of the same discipline.
Reconciliation checks
These run before any commentary is drafted. If a check fails, drafting stops and a person is told why.
- Revenue, expenses and profit for the period tie to the trial balance in Xero or MYOB
- Totals in every table equal the sum of their rows
- Figures in the commentary match the figures in the tables, by automated comparison
- The data was refreshed after the period closed, with the refresh time shown on the report
- Row counts and record totals are within expected ranges, so a failed sync is caught
- Period labels, currency and units are consistent throughout
- Prior-period comparatives match what was previously reported, or the restatement is flagged
- Any figure without a traceable source query blocks the report
How the system is built
| Layer | What it does | Typical tools |
|---|---|---|
| Sources | Systems of record | Xero, MYOB, CRM, Shopify, job systems, spreadsheets |
| Data layer | Combines sources and keeps history | A warehouse, or a Power BI semantic model |
| Metric definitions | One agreed definition of each measure | A documented metrics layer |
| Query layer | Produces every number in the report | Reviewed SQL or DAX, generated queries shown before running |
| Language layer | Drafts commentary from query results | A model chosen by evaluation |
| Review | A person approves before release | Approval step with traceability |
What you receive
- Metric definitions
- The agreed definition of each measure, in writing, so “gross margin” means one thing.
- Report pipeline
- Queries, checks, drafting and assembly for each recurring report.
- Traceability
- Every figure linked to its query and source, visible to the approver.
- Question tool (optional)
- Plain-English questions answered with the query shown, limited to approved data.
- Check log
- A record of every reconciliation run, pass or fail.
- Runbook
- How to add a metric, change a report or investigate a failed check.
How the work runs
- Discover: one report, its definitions. Pick the recurring report that costs the most time, and agree the definitions of every measure in it.
- Prove: the regeneration demo. Rebuild three past periods from source and review the differences.
- Prove: parallel month. Produce the next live report both ways, then compare effort and accuracy.
- Scale: more reports and questions. Add reports and, once definitions are stable, the question tool.
What drives the cost
The drivers are the number of source systems and how clean their data is, how long it takes to agree metric definitions, whether a data layer exists already, report frequency and the number of recipients, and whether you want the open question tool. Agreeing definitions is often the slowest step, and the most valuable. See what custom AI projects cost, and compare the hours saved at month-end using the AI ROI calculator. Integration with your systems is covered in AI integration services.
Related work
- Lucent, a Levy Street venture, is an agentic investment research platform for wealth, equity and private-markets teams covering NZX and ASX. It runs several frontier models together and grounds its outputs in citations.
- GTM-Bench is a 72-task benchmark for AI agents doing go-to-market research, published by Blackpearl Group in June 2026 and co-authored by Levy Street team members. Its scoring checks whether each claim is supported by evidence and treats low-quality records as harmful, the same standard we apply to report commentary.
Limits
- Invented causes. A model will explain a variance confidently with a reason it made up. Commentary on causes must come from the data or from a person.
- Ambiguous questions. “Sales last quarter” can mean invoiced, paid or booked. Without agreed definitions, the question tool gives inconsistent answers.
- Wrong joins. Generated queries can join tables incorrectly. Show the query, test common questions, and restrict the tool to reviewed views.
- Statutory reporting. Financial statements and regulated returns remain the work of your accountants and auditors.
- Small and simple. If your reporting is one spreadsheet that takes an hour, keep the spreadsheet.
New Zealand specifics
- Local systems and calendars. Most NZ reporting draws on Xero or MYOB, often with Power BI. Periods follow your GST filing cycle and, for many organisations, a 31 March balance date.
- Power BI Copilot. Microsoft’s documentation, as fetched on 19 October 2026, requires paid Fabric (F2 or higher) or Premium capacity, and disables Copilot by default for tenants outside the US and EU data boundary unless an admin allows processing outside your region. That is a data-location decision, not just a setting.
- Reports about people. Staff performance or customer-level reports involve personal information, and IPP 8 of the Privacy Act 2020 requires reasonable steps to check accuracy before use. The Privacy Commissioner’s 2026 survey found 67% of New Zealanders concerned about AI being used to make decisions about people from personal data. All 14 information privacy principles apply.
Sources
- Copilot for Power BI overview, Microsoft Learn, accessed 2026-10-19
- GTM-Bench: Evaluating Agentic AI Systems for Buyer/Seller Coherence in Go-to-Market Workflows, Blackpearl Group, 2026-06
- 2026 annual survey on privacy, Office of the Privacy Commissioner, 2026-05-11
- Privacy Act 2020, New Zealand Legislation, version as at 1 October 2026
- Generative artificial intelligence, Office of the Privacy Commissioner, 2023-06-15
Written by the Levy Street team in Wellington. Last updated 19 October 2026.