ANZ's AI control plane: what NZ firms can copy at a smaller scale

What ANZ announced

ANZ Group chief executive Nuno Matos used the bank's full year results briefing on 9 November 2026 to say the bank is building an Enterprise AI Control Plane. According to his remarks, the control plane will give ANZ visibility, governance and technical control over users, models, agents, tools, cost and platforms, and it'll record how decisions were reached. ANZ says the system will put defined limits on what its agents can do and stop activity that falls outside those limits. Matos put the principle in one line: "As we increasingly use AI agents, humans must remain in control."

ANZ isn't starting from nothing. The bank already runs several AI gateways and a governance council, and iTnews reports that the control plane builds on both. ANZ hasn't said what software sits underneath the control plane or whether outside vendors helped build it. iTnews also notes that a blog post from ANZ CIO Donald Patra gave no architectural detail.

ANZ is the second large Australian organisation to describe this kind of structure. Telstra has stood up its own AI control plane with Accenture, and Accenture's case study describes a single view of Telstra's AI applications, models and infrastructure that tracks performance, risk and cost. The same case study puts Telstra at more than 100 live AI use cases, which gives a sense of the volume that pushes a company towards central tooling.

Why a bank needs a control plane for agents

A chatbot that drafts an email can't do much harm on its own, because a person reads the draft before anything happens. An AI agent calls tools, reads systems and takes actions, sometimes many in a row, so the risk sits in what the agent has access to and how far it can go before anyone notices. Most organisations that have run AI pilots for a year or two end up with agents spread across different products. Some come built into SaaS platforms, some were built by internal teams, and some were switched on by staff who bought a licence themselves. Each one has its own permissions, its own logs (if there are any) and its own bill.

ANZ's list of what the control plane covers reads like a checklist of the gaps that open up in that situation. "Users" and "agents" are about identity: who or what is acting. "Tools" covers what each agent is allowed to touch. "Cost" matters because agents that loop or retry can run up model charges quickly. "Traceability of decisions" means being able to reconstruct afterwards why an agent did what it did, which is the first thing a regulator, auditor or customer will ask for when something goes wrong.

The control plane also fits a pattern ANZ described elsewhere in the same briefing. Matos said the bank is prioritising AI use cases with clear business value and steering away from "AI everywhere". A central register of agents, with costs attached, is how a business finds out which use cases are earning their keep.

What it means for NZ businesses

ANZ is New Zealand's largest bank by customer numbers, so its group-level governance choices tend to reach the NZ operation and, eventually, the expectations placed on its suppliers and business customers. Firms that integrate with ANZ, supply it with services or compete for the same staff should expect questions about how their own AI tools are controlled.

Most NZ organisations won't build a bespoke control plane, and they don't need to. The ideas carry over without a bank's budget, though. A 200-person distributor running a Copilot agent in Microsoft 365, an AI triage agent in its service desk and a document agent in accounts payable faces the same questions ANZ does at a smaller scale: what each agent can reach, who approved that access, what each agent costs per month, and how quickly someone can switch one off.

The weak point we see most often is ownership. Agents get switched on inside a platform by whoever administers that platform, so nobody holds the full list. When a staff member leaves or a vendor changes its pricing, the gaps only come to light after something breaks. ANZ's governance council plus gateway plus control plane structure is an answer to that ownership problem as much as to a technical one.

There's a cost angle too. Vendors are moving AI features to per-action or per-document billing, and an agent with broad permissions and no spending limit can produce an unpleasant invoice. Putting cost on the same dashboard as permissions, as ANZ and Telstra both describe, lets finance and IT look at the same figures.

Practical next steps

These steps assume a mid-sized NZ organisation with a handful of AI agents in production or pilot. They're roughly in order of effort.

• Build an agent register. List every AI agent in use, including ones built into SaaS products. For each one, record the business owner, the systems it can read or write, the account it runs under and the monthly cost. A spreadsheet is fine to start with.
• Give each agent its own identity. Agents that run under a staff member's login inherit everything that person can do and leave muddy audit trails. Use service accounts or the agent identities your platform provides, with the narrowest permissions that still let the agent do its job.
• Write down the limits for each agent. Set the maximum transaction value, which records it can change, and which actions need human approval. ANZ's "defined limits" are this step in practice. Put the limits into platform configuration where you can, so they don't live only in a policy document.
• Route model calls through one gateway where practical. A single AI gateway, whether that's a cloud provider's service or a self-hosted proxy, gives you one place for logging, cost caps and blocking a model or tool quickly. Teams handling sensitive data may prefer a private AI server for some workloads.
• Keep decision logs you can actually read. Check that each agent records its inputs, the tools it called and its output, and that those records are kept long enough to satisfy your auditors and the Privacy Act 2020.
• Test the off switch. Pick one agent and time how long it takes to disable the agent, revoke its credentials and confirm it has stopped. Repeat the test after major platform updates.
• Give one group the final say. A small AI governance group with IT, risk and operations represented can approve new agents against the register. The group doesn't need to be large; it needs authority.

ANZ hasn't published how its control plane performs in practice, and the bank's results don't yet show any measured effect from it. The structure is still worth noting, because it shows where large Australian organisations are putting their effort now that agents are moving from pilots into daily work. If you're working out how to govern AI agents across your own systems, our AI consultancy team can help you set up the register and limits before the number of agents grows.