Short answer
AI document processing reads business documents such as invoices, forms, contracts and reports, extracts the fields you need, checks them against rules and existing records, and sends anything uncertain to a person. Levy Street builds document processing for New Zealand organisations that is measured field by field on your own documents, with a review queue designed around the errors that actually occur.
- Accuracy is measured per field on a held-out set of your real documents.
- Validation rules, such as totals and GST arithmetic, catch errors the model misses.
- If suppliers can send eInvoices, take the structured data and skip AI.
- IRD expects business records kept for at least 7 years, with approval for offshore storage.
What AI document processing handles well
AI document processing works best on documents that carry the same kinds of information in many different layouts: supplier invoices from hundreds of suppliers, application forms, delivery dockets. Modern vision models read layouts that older template-based OCR could not, without a template per supplier. It struggles with poor scans, messy handwriting, tables that run across pages, and anything that needs professional judgement rather than reading.
| Document | Typical fields | Difficulty | Notes |
|---|---|---|---|
| Supplier invoices and bills | Supplier, GST number, date, totals, line items | Low to medium | Line items are the hard part |
| Delivery dockets and proof of delivery | Consignment number, date, signature present | Medium | Phone photos, creases, handwriting |
| Application and intake forms | Names, addresses, ticked boxes, dates | Medium | Handwriting and scan quality |
| Contracts and leases | Parties, dates, terms, specific clauses | Medium to high | Needs citations to the clause |
| Inspection and technical reports | Findings, severity, recommended actions | High | Extraction is easy; interpretation is not |
Two worked workflows show the pattern in practice: supplier invoice processing and proof of delivery extraction.
Choosing between eInvoicing, OCR and vision models
If a supplier can send structured eInvoices, use them. Inland Revenue notes that taxable supply information can be exchanged directly between buyer and seller software, for example through PEPPOL eInvoicing, and structured data needs no AI to read. For the rest, the choice between traditional OCR and vision models depends on layout variety and volume; our engineering note on OCR versus vision models sets out the trade-off. For document-heavy workflows beyond extraction, see AI workflow automation.
How we measure extraction quality
We measure extraction quality field by field, on a held-out set of your own documents that was never used for tuning, and then add validation rules that catch what the model misses.
Start with representative documents
The most important input to the project is a realistic sample. We ask for 200 to 500 documents that reflect your real mix: every major supplier or form type, the worst scans, the multi-page ones, the credit notes. Staff who process them today label the correct values. The set is split into a development set and a held-out test set that is never used for tuning, so the accuracy we report is a fair estimate of what you will see in production.
Per-field accuracy and validation
Accuracy is reported per field, because “95% accurate” across a whole document hides the field that matters. Each field gets a rule for when it can be accepted automatically.
Illustrative targets for a supplier invoice project (not results from a client project):
| Field | Measured on held-out set | Auto-accept rule |
|---|---|---|
| Supplier GST number | Share correct | Valid format and matches the supplier record |
| Invoice date | Share correct | Within an expected range of today |
| Invoice total (GST inclusive) | Share correct | Line items and GST reconcile to the total |
| GST amount | Share correct | Equals 3/23 of the GST-inclusive amount for standard-rated supplies |
| Line items | Share fully correct | Reviewed when there are more than an agreed number of lines |
Validation rules do a lot of the work. Totals must reconcile, GST must match 15% of the exclusive amount on standard-rated lines, the supplier must exist, and duplicates (same supplier, number and amount) are blocked. A model can misread a digit; it is much less likely to misread it in a way that still passes every check.
The review workflow
- Flag with a reason. Items with a low-confidence field, a failed check or an unknown supplier go to the review queue with the reason shown.
- Review side by side. The reviewer sees the document and the extracted fields together, with the uncertain fields highlighted, and corrects only those.
- Learn from corrections. Every correction is logged and added to the evaluation set, so recurring errors are fixed and tested.
- Audit what went straight through. A weekly sample of auto-accepted items is checked, to catch errors the rules missed.
What you receive
- Document sample and labels
- Your representative set, labelled, split into development and held-out test sets.
- Extraction pipeline
- Reading, field extraction, validation rules and duplicate checks.
- Accuracy report
- Per-field results on the held-out set, rerun on every change.
- Review queue
- Side-by-side review with highlighted fields and reasons.
- System integration
- Drafts or records created in Xero, MYOB, your ERP or document system.
- Retention settings
- Where originals and extracted data are stored, and for how long.
How the project runs
- Discover: collect and label. Gather the representative sample, agree the fields and rules, and time the current manual process.
- Prove: measure on held-out documents. Build the pipeline, measure per-field accuracy, and set auto-accept rules from the evidence.
- Prove: parallel run. Process live documents alongside the existing manual process and compare.
- Scale: switch over and widen. Move to review-by-exception, then add document types one at a time.
What drives the cost
The drivers are document variety, volume, scan quality, the number of fields and rules, the target system, and the review interface. Model fees are usually modest. As an illustration with assumed numbers: 3,000 invoices a month at two pages each is 6,000 pages; at an assumed NZ$0.02 a page for reading and extraction, that is about NZ$120 a month. If 20% need a 90-second review, that is 15 hours of staff time, against roughly 150 hours if each invoice takes 3 minutes by hand. See what custom AI projects cost and model your own volumes in the AI ROI calculator.
Related work
- The Property Management Agent for APL Property handles routine administration for a regional NZ property firm, including email triage, records and orders, with staff keeping the decisions.
- 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, the same discipline we apply to contract and report extraction.
Limits
- Messy handwriting and poor photos. Accuracy drops sharply. Fix the form or the capture process if you can.
- Judgement disguised as extraction. Deciding whether a clause is unfavourable or a defect is serious is not reading. Keep a professional in the loop.
- Payments. Never let extraction trigger a payment without approval.
- Identity documents. Matching a face on an ID to a person is biometric processing with its own privacy code; do not add it casually.
- Small volume. A few dozen documents a month is cheaper to key by hand.
New Zealand rules built in
- Taxable supply information. IRD’s requirements vary by value: supplies of $200 or less, more than $200 up to $1,000, and more than $1,000, which also needs the buyer’s details if the buyer is GST registered. Validation rules check the right fields for each band.
- Record keeping. IRD says business records must be kept for at least 7 tax years, and if records are stored offshore, including in cloud services, you or your provider need IRD approval. We design retention and storage with that in mind.
- Data location. Storage can sit in Microsoft’s New Zealand North region in Auckland; model processing is offshore, with Amazon Bedrock’s AU profile keeping Claude processing in Australia and New Zealand.
- Privacy. Forms and contracts carry personal information, so security (IPP 5) and retention (IPP 9) under the Privacy Act 2020 apply alongside the other 12 information privacy principles.
Sources
- How taxable supply information for GST works, Inland Revenue, 2026-03-31
- Record keeping, Inland Revenue, accessed 2026-10-19
- Privacy Act 2020, New Zealand Legislation, version as at 1 October 2026
- Azure regions list, Microsoft Learn, accessed 2026-10-19
- Run generative AI inference with Amazon Bedrock in Asia Pacific (New Zealand), AWS, 2026-03-26
Written by the Levy Street team in Wellington. Last updated 19 October 2026.