AI integration services.

What AI integration covers

AI integration is the work of giving an AI system the access it needs, and no more, to the software your organisation already runs. It covers four things: read paths (what the model can see, such as invoices, emails or job notes), write paths (what it can change, such as drafts, notes or staging records), identity (whose permissions it acts under), and events (what triggers it, such as a new email or a changed record). It is not dropping a chat widget onto a website. Many stalled AI pilots are not short of model quality; they cannot reach the data that matters.

Who it suits

• You use AI tools that cannot see your data. Staff copy and paste between ChatGPT or Copilot and your systems all day.
• You are building an AI system and need it wired into accounting, CRM, documents or job management. See custom AI development.
• You have several AI tools and no control. Each vendor connected itself to your Microsoft 365 or Google Workspace with broad permissions, and nobody has reviewed them.
• Your core system is old. Valuable data sits in an on-premises database or file share with no modern API.

Systems of record and how we connect to them

Our integration guides cover the native AI features and API options for each system, starting with Xero and Microsoft 365.

Access patterns: which to use when

• App permissions with a service identity. For background systems that act on behalf of the organisation. Scope to specific record types and log every call.
• Delegated permissions. For assistants that act on behalf of a person. The AI inherits that person’s access, so a search never returns documents the user could not open.
• MCP servers. When several AI clients or agents need the same tools. Our guide to MCP for business systems explains the trade-offs.
• Events over polling. Webhooks react within seconds and avoid rate limits; polling is the fallback when a vendor offers no events.
• Outbound connectors for legacy. A small agent inside your network that dials out, so no inbound ports are opened.
• Never: shared admin logins, personal staff credentials or screen scraping, unless there is truly no alternative and the risk is accepted in writing.

Who maintains what after launch

What you receive

Integration design

Each connection’s read and write scope, identity, trigger and failure behaviour.

Connectors

API clients, MCP servers or outbound connectors, with retries, rate limiting and logging.

Staging and approval

Where AI writes land before a person approves them.

Monitoring

Failure, volume and latency alerts for every connection.

Maintenance register

The table above, filled in with names.

How it runs

1. Discover: inventory. List the systems involved, their APIs, existing app connections and who administers each.
2. Prove: read first. Build read-only connections and confirm the AI gets the right data with the right permissions on real records.
3. Prove: write to staging. Add write paths that create drafts or staging records only, and test with the people who will approve them.
4. Scale: harden and hand over. Monitoring, credential rotation, documentation and the maintenance register signed off by owners.

What drives the cost

The drivers are the number of systems, the quality of their APIs, authentication complexity, whether you need write-back, legacy connectors, volume against vendor rate limits, and monitoring. A clean cloud API with good documentation is quick; an on-premises database behind a firewall is not. We scope each connection before pricing. See what custom AI projects cost and estimate the payoff with the AI ROI calculator.

Systems we have connected

• The Property Management Agent for APL Property is an internal AI agent for a regional NZ property firm that works across email, records and orders, handling triage and follow-up with staff keeping the decisions.
• freshOps is commercial cleaning operations software from FreshOps Group in Australia. Levy Street worked on the platform.
• SPAWN D is Levy Street’s open-source, invite-only, pre-alpha control plane for running CLI coding agents on machines you own from a browser or phone. Its daemon dials out rather than opening inbound ports, the same pattern we use for on-premises connectors.

When not to integrate

• The vendor’s own AI already does it. If a feature built into Xero, Microsoft 365 or your helpdesk covers the need, use it.
• No API and no stable export. Screen scraping breaks on every vendor update.
• The process is about to change. Integrating a system you are replacing next year is rarely worth it.

New Zealand specifics

• Older systems are common. In MYOB’s April 2026 mid-market research, reported by IT Brief, only 36% of NZ mid-sized businesses said they run a modern cloud-based ERP, so many integrations need legacy patterns.
• Data location. Microsoft’s New Zealand North Azure region and the AWS New Zealand region let integration data be stored in Auckland. Model processing is still offshore: the closest option is Amazon Bedrock’s AU profile, which keeps Claude processing within Australia and New Zealand.
• Privacy. Under the Privacy Act 2020, a provider that only processes data for you leaves you fully responsible for it. Integration scopes are reviewed against the 14 information privacy principles, especially security (IPP 5) and retention (IPP 9).

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