AI agent developers in New Zealand.

An AI agent is software in which a language model decides the next step and calls tools to act: it reads a record, drafts a reply, books a job or updates a CRM. Because it changes real records, a builder’s approach to permissions and testing matters more than its demo. Here are the New Zealand providers that say they build or set up agents, and what to ask them.

If your process follows the same steps every time, see our guide to AI automation agencies in New Zealand instead, and automation versus AI agents for where the line sits.

New Zealand AI agent developers compared

As listed on each provider’s own website on 20 October 2026. Prices exclude GST.

Provider profiles

Levy Street

Levy Street is a Wellington AI engineering company that builds AI agents connected to systems such as Xero, Microsoft 365 and HubSpot, with permissions set per action, human approval before anything irreversible and an evaluation suite run before every change. Its engineers built the Property Management Agent for APL Property, and Levy Street engineers work with Blackpearl Group’s AI team on Bebop, an AI sales agent owned and sold by Blackpearl Group (NZX/ASX: BPG).

API Connects

API Connects is an Auckland technology consultancy for enterprises that offers governed agentic workflows: multi-step agents running processes across systems, with human approval on high-impact actions. It says it evaluates each workflow against expected outputs before go-live.

Autohive

Autohive is a New Zealand company (Autohive AI Limited) whose no-code platform lets staff build AI agents and run them on a schedule, such as a weekly cash flow summary from Xero. It has a free plan, and paid plans from US$20 a month.

Automate AI

Automate AI is a Wellington agency, founded in 2024, that builds and hosts voice agents which answer calls, book into Google or Microsoft calendars and update CRMs. Its Packaging Products case study describes two agents for a Lower Hutt supplier, one managing Google Ads and one writing social posts.

Avtrix

Avtrix Software Solutions is a Taupō company that builds custom AI agents on a fixed price, connected to Xero, MYOB, HubSpot, Salesforce and Microsoft 365, with approval steps and audit logs. It says it tests agents on real enquiries before launch and lists NZ$8,000 to NZ$35,000 for most agent projects.

Changeable

Changeable is a Taranaki consultancy founded by Steve Wilson that builds governed AI agents for research, documents, triage and knowledge retrieval, each tied to a business process and human review. It starts with process analysis and says most SME engagements cost NZ$3,000 to NZ$20,000.

Datacom

Datacom is a New Zealand IT services company founded in Christchurch in 1965. Its AI practice serves enterprise and government clients in New Zealand and Australia, and its productivity offering includes personal AI assistants and intelligent agents.

Enlighten Designs

Enlighten Designs is a Hamilton software company founded in 1998 that describes itself as a Microsoft partner. Its AI development page offers agentic AI that lets applications make decisions and perform tasks, built on Azure AI and Vertex AI.

Ez-AI

Ez-AI is an Auckland business run by engineer Skene Bennellick. It builds assistants and agents that draft replies, book calls and hand over to a person, with every action logged; a listed holiday park assistant was put through 525 test conversations. Its typical price for agents is NZ$3,000 to NZ$12,000.

Harkness AI

Harkness AI is a Christchurch consultancy founded by Blake Harkness that connects Claude to business tools, such as the Site App Pro and PayHero connectors it built for Trap and Trigger. Its embedded AI engineer service places an engineer in your team one to three days a week, for businesses of 20 to 100 staff.

OpenClaws NZ

OpenClaws NZ is a Christchurch service run by AI trainer Caelan Huntress that installs OpenClaw, an open-source agent, on a dedicated Mac mini or virtual server, connected to Claude or ChatGPT. The listed package is NZ$3,900 plus GST, plus NZ$1,100 for the Mac mini.

How to choose an AI agent developer

Ask every provider these questions, including us. Good answers name your systems and come with an example.

• Which actions can the agent take without a person approving?
• Does it act under its own service account, or a staff member’s login?
• What happens when an email contains hidden instructions (prompt injection)?
• How many of our real past cases will you test it on, and what counts as a pass?
• Who owns the code, prompts, test cases and logs if we part ways?

If the job is a scheduled report or a draft that staff review, a platform such as Autohive or Copilot Studio may be enough. For a second opinion on any proposal, ours included, book an assessment.

How to test an agent before launch

MBIE’s Responsible AI Guidance for Businesses says developers or deployers “should be satisfied that it performs adequately, is reliable, and safe” before release, and should retest because models change.

1. Collect real cases. Gather 50 to 100 past cases, awkward ones included, with agreed correct outcomes.
2. Run read-only first. Score its conclusions on historical cases before it can change anything.
3. Score whole tasks. Count wrong or unnecessary actions; one wrong payment outweighs ten good drafts.
4. Attack it. Feed it documents with hidden instructions; MBIE lists red teaming among safety checks.

Our AI agent acceptance checklist turns these into a sign-off sheet.

Where agent projects go wrong

• A demo instead of a test. Ask for results on your own cases, not ten good runs.
• Broad access for convenience. An agent using a manager’s login can do anything that manager can.
• Writes nobody confirms. Xero says the V1 tools on its remote MCP server “do not elicit user confirmation”, so approvals must be designed around the connection.
• No owner after launch. Agents can quietly stop picking up work. Watch throughput, not just errors.

New Zealand rules that apply

• You stay responsible. The Privacy Commissioner says that if a third party processes personal information on your behalf, you remain responsible for it. That includes your agent builder and model provider.
• IPP 3A. Since 1 May 2026, an agent that collects personal information about people from other sources, such as researching leads, generally means taking reasonable steps to make them aware as soon as reasonably practicable.
• A person in the loop. MBIE’s quickstart says a human-in-the-loop is needed where important decisions and actions are made or informed by AI.
• Xero costs. Xero says real customer organisations need its Core developer plan or above to use the remote MCP server.

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