Short answer
AI workflow automation connects the steps of a business process, such as reading an email, checking a record and updating a system, and uses an AI model only where judgement on messy input is needed. Levy Street builds these automations for New Zealand teams with fixed rules wherever possible, an explicit exception queue, and a measured before-and-after cost so you know what the automation actually saves.
- Use plain rules for every step you can write down; use a model only for messy input.
- Every item must end as done, queued for a person with a reason, or rejected.
- Time the current process before building, so savings are measured rather than claimed.
- Exception rate is the number that decides whether an automation pays.
What AI workflow automation is, and what it is not
AI workflow automation is ordinary process automation with a model added at the steps that need reading or judgement. The trigger, routing, system updates and notifications stay as fixed rules because rules are cheap, fast and predictable. The model reads the free-text email, extracts fields from the attachment, classifies the request or drafts the reply.
It is not an autonomous agent deciding what to do next. If the steps are the same every time, an agent adds cost and risk without benefit; our comparison of automation versus AI agents explains where the line sits. When the path genuinely varies case by case, see AI agent development.
Who it suits
- Shared inboxes. Job requests, supplier queries or applications arrive by email and someone reads, sorts and re-types every one.
- Re-keying between systems. The same details are typed into a CRM, a job system and an accounting package.
- Document chasing. Staff spend hours a week asking clients or suppliers for missing paperwork and checking what came back.
- Approvals. Requests wait in someone’s inbox because nobody assembled the information needed to approve them.
Before and after: a worked example
Illustrative example: a 25-person Wellington trades business receives about 400 job requests a month by email and enters them into ServiceM8 by hand.
| Step | Before (per request) | After (per request) |
|---|---|---|
| Read and triage the email | Office staff, about 3 minutes | Model classifies job type, urgency and suburb |
| Create the job | Re-keyed into ServiceM8, about 4 minutes | Draft job created through the ServiceM8 API |
| Reply to the customer | Typed reply, about 3 minutes | Reply drafted from templates and the job details |
| Human time | About 10 minutes | 1.5 minutes to check and approve; 10 minutes for exceptions |
The arithmetic, with stated assumptions. Before: 400 × 10 minutes is about 67 hours a month. After: if 85% of requests are routine (340 × 1.5 minutes = 8.5 hours) and 15% are exceptions handled the old way (60 × 10 minutes = 10 hours), plus 2 hours a month of monitoring, the total is about 20.5 hours. That frees roughly 46 hours a month. At an assumed loaded staff cost of NZ$45 an hour, that is about NZ$2,080 a month, before running costs. If platform, hosting and model fees came to an assumed NZ$300 a month, the net is about NZ$1,780. Change the exception rate to 40% and the hours saved fall by about 30%, which is why we measure it first.
Exception handling is the design
The automation is only as good as what it does with items it cannot handle. Our rule: every item ends in exactly one of three states. Done, with a log entry. Queued for a person, with the reason shown, such as a missing address, an unknown customer, low model confidence or an angry tone. Rejected, with a reason sent back to the sender. Nothing is silently dropped.
Common exceptions are designed in from the start: emergencies that need a phone call now, duplicates, attachments the model cannot read, and requests outside your service area. The exception rate is reported weekly, and each new rule added to handle a recurring exception should push it down. Our guide to human review queues covers queue design.
What you receive
- Process maps
- The current workflow with real timings, and the automated version with every decision point and exception path.
- The automation
- Built on n8n, Make, Zapier or custom code depending on volume, data sensitivity and who will maintain it.
- Exception queue
- A place staff work exceptions with the source item, extracted data and reason side by side.
- Operating cost sheet
- Platform, hosting and model fees per month at your volume, plus review time.
- Measures dashboard
- Volume, straight-through rate, exception rate, time per item and errors found in sampling.
- Runbook
- How to pause, change and restart the automation, and who owns it.
How the work runs
- Discover: time the real process. Two weeks of actual volume, timed, with exceptions categorised. This gives the baseline every later claim is measured against.
- Prove: shadow mode. The automation runs alongside staff without acting. We compare its output with what people did and fix the gaps.
- Prove: supervised. The automation acts, a person approves each item, and we measure review time and corrections.
- Scale: sampling and the next workflow. Once the numbers hold, review moves to a sample, and the same connections are reused for the next process.
What drives the cost
The main drivers are how many systems are involved and whether they have usable APIs, monthly volume, the exception rate, the review interface, and whether the platform is hosted or self-hosted for data reasons. Model fees are usually minor next to staff time. Our page on AI automation pricing in New Zealand works through scenarios in NZD, what custom AI projects cost covers larger builds, and the AI ROI calculator lets you plug in your own volumes. We price each automation as a fixed scope after the timing study.
Related work
- The Property Management Agent for APL Property handles routine administration for a regional NZ property firm: email triage, records, orders and follow-up, with staff keeping the decisions.
- MOGGED Digital is Levy Street’s AI-native Wellington digital studio for websites, SEO and ads, which says it has served 100+ businesses, mostly in New Zealand. MOGGED describes AI running through its process to research, draft and test, with every decision made by a person.
Limits: when not to automate
- Low volume. A few dozen items a month rarely repays the build and maintenance.
- A process in flux. Automating a workflow that changes every quarter means rebuilding every quarter. Stabilise it first.
- No API. Screen-scraping a system without an API breaks whenever the screen changes.
- Decisions with legal effect on people. MBIE’s quickstart for businesses using AI tools says a human-in-the-loop is needed where important decisions are made or informed by AI. Keep those decisions with people.
- Errors that cost more than the time saved. Payments and anything irreversible stay behind an approval.
New Zealand specifics
- Employment. Where automation could affect roles, the Employment Relations Act 2000 requires employers to share relevant information and give staff a chance to comment before deciding, and Employment New Zealand says workplace changes need genuine reasons, consultation and a fair process.
- Privacy. Automation platforms often process data offshore. Under the Privacy Act 2020, using a provider purely on your behalf leaves you fully responsible for what happens to the data, so check its settings and terms against the 14 information privacy principles.
- Local software. The most common NZ connections are ServiceM8, Simpro, Xero, MYOB, Cin7 and Microsoft 365. Our n8n guide covers self-hosting where data must stay under your control.
Sources
- The basics of workplace changes, Employment New Zealand, 2025-11-13
- Employment Relations Act 2000, New Zealand Legislation, version as at 10 July 2026
- Quickstart: Businesses using AI tools, Ministry of Business, Innovation and Employment, undated
- Working with third-party providers, Office of the Privacy Commissioner, 2024-11-26
- Privacy Act 2020, New Zealand Legislation, version as at 1 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.