AI compliance evidence collection.

What evidence collection does before an audit

Compliance evidence collection is the week before an assurance audit, done continuously instead. Someone takes the programme’s checklist, finds the record that proves each item, notices what is missing and chases it. On most farms and orchards those records are spread across a shared drive, email attachments, paper diaries and contractors’ inboxes. The AI workflow reads them, matches each to the checklist line it supports with a page reference, and produces a gap list weeks before the auditor arrives.

The rule that makes this safe is strict: the system finds and indexes records that exist. It never writes a record that should have been made at the time.

How evidence collection runs

Inputs

The current assurance checklist, spray diaries, operator certificates, sprayer calibration records, chemical store inventory, staff training and induction records, contractor details and declarations, water test results, and last audit’s findings and corrective actions.

Systems

A shared document store, an inbox for contractor documents, text extraction for scans and photos of paper records, language models for field extraction and checklist mapping, and a tracker for gaps and actions.

Steps

Load the checklist; extract names, dates, expiry, products and rates from each document; map documents to checklist items with page references; list gaps, weak evidence and expiring certificates; draft requests to staff and contractors; assemble an evidence index for the auditor.

Human review

The orchard or farm manager confirms each mapping and decides how every gap is handled. The owner signs any declaration. Doubtful contractor documents are checked at source.

Outputs

An evidence index linked to source documents, a gap list, an expiry calendar, drafted requests, and a corrective action log.

Measures

Checklist items with confirmed evidence two weeks before the audit, certificates expiring within 60 days, hours to assemble the pack, and audit findings compared with the previous year.

Risks

Backfilled or fabricated records, an outdated checklist, misread dates, and evidence that mentions an item without satisfying it.

Illustrative example: a 45 ha Hawke’s Bay apple orchard

Illustrative example, synthetic numbers: a 45 ha Hawke’s Bay apple orchard prepares for its annual assurance audit. Assume a checklist of 120 items and about 300 supporting documents, including spray records, contractor certificates and training sign-offs. Assembling the pack by hand takes the orchard manager about 30 hours.

Assumptions, at Anthropic prices listed on 19 October 2026: mapping each of the 120 items with Claude Sonnet 5.5 (US$2 per million input, US$10 per million output tokens), reading about 20,000 tokens of candidate documents and writing 500, costs about 4.5 US cents per item, or US$5.40. Extracting fields from 300 documents with Claude Haiku 5.5 (US$0.10 and US$0.50 per million) at 3,000 input and 300 output tokens adds about 14 US cents. The saving is in hours the manager would otherwise spend in a filing cabinet during picking season.

NZ programmes and records involved

• Horticulture assurance. NZGAP provides assurance for New Zealand horticulture, covering food safety issues including agrichemical residues, and publishes forms, checklists and guidelines for growers. Use the programme’s current documents as the checklist source.
• Operator certificates. Growsafe, run by the New Zealand Agrichemical Education Trust, certifies agrichemical users. Certificates expire, can be renewed online if less than six months expired, and spray contractors holding a Growsafe RCA must complete continuing professional development. Expiry dates are the most common fixable gap.
• Agrichemical management. NZS 8409:2021 gives practical guidance on the safe, responsible and effective management of agrichemicals. It is a useful structure for organising spray, storage and calibration evidence.
• Livestock movements. On cattle and deer farms, OSPRI states that movements between NAIT locations must be recorded within 48 hours, so movement confirmations are routine evidence.
• Staff records. Training files and certificates hold personal information, so the Privacy Act 2020 applies. Share only what the auditor needs.

What the grower signs off

The grower decides whether a record actually satisfies a checklist item, how each gap is explained and corrected, what is shown to the auditor, and what is declared. Those calls carry the farm’s name. Keep a log of what the system proposed and what the grower accepted; our guide to AI audit logs explains what to record.

Where evidence collection goes wrong

• Backfilling. Asked to “complete the pack”, a model will happily draft a spray diary entry for last February. Block record generation outright and label every model-written sentence as a draft note.
• Certificates checked on the wrong date. A contractor’s certificate valid today may have lapsed during a job in March. Check expiry against each job date.
• Last year’s checklist. Programmes revise their requirements. Mapping against an old version gives false confidence. Record the checklist version on the index.
• Date formats. Imported product documents may use month-first dates, so 03/04 can mean 4 March. Store dates in one format and flag ambiguous ones.
• Evidence that only mentions the item. A training sign-in sheet shows attendance, not what was taught. Mark weak matches separately from strong ones.

Evidence collection draws on farm record retrieval and the service history from equipment maintenance logs. The agriculture sector guide covers the wider picture, and our AI document processing service handles the extraction side.

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