NAB gives 40,000 staff an AI assistant: lessons for NZ firms

What NAB announced

NAB published its 2026 full-year result on 5 November. On the media call, chief executive Andrew Irvine said the bank had rolled out an internal AI assistant called Ruby. iTnews reported that Ruby runs on an Anthropic platform and that more than 40,000 NAB employees will have access. Irvine said the tool would be "broadly available to those colleagues for personal productivity." NAB hasn't said which Anthropic model Ruby uses, what the licences cost, or whether BNZ staff in New Zealand will get Ruby.

The assistant is only one part of NAB's programme. Over the past year the bank has set up around 50 business-led squads drawn from several divisions, and those squads work with Anthropic and AWS on use cases aimed at the bank's growth priorities in business banking, deposits and proprietary home lending, according to iTnews. Irvine gave one example: helping business bankers prepare credit memos faster.

The numbers, and what they cover

NAB released percentage gains and no dollar value for its AI spend. The Asian Banker listed the reported figures. Staff using AI cut business credit memo preparation time by 20% and cut the time spent finding corporate and institutional customer insights by 75%. Design-to-code time fell by 50%, cyber incident triage got 80% faster and technology cyber assessments took 90% less time. iTnews also reported a 20% cut in business loan approval time from agentic AI.

Those figures need careful reading. The Asian Banker points out that NAB's results cover particular tasks and user groups and don't measure productivity across the whole bank. It also notes that drafting a credit memo faster doesn't on its own prove that loans get approved faster. NAB's agents for complaints handling, fraud detection and financial crime are still in development.

The same results release confirmed that NAB will cut up to 2,500 roles by September 2027. The Finance Sector Union has said NAB's AI investment is contributing to those reductions, iTnews reported. Any organisation copying NAB's approach should expect staff to connect the two announcements, whether or not management does.

Westpac's migration result the same week

Westpac gave a different kind of AI result in its own November results. The Asian Banker reported that Westpac now expects moving the St.George data warehouse to cost AUD 35 million, half the original AUD 70 million budget. Engineers used AI to convert legacy code, test the results and document the changes. Chief executive Anthony Miller said the migration finished more than two years ahead of schedule. Westpac plans to retire the old warehouse in its 2027 financial year and expects benefits of AUD 15 million a year from then.

The Westpac result has a fixed scope and a budget to compare against, so the saving is easy to verify. NAB's time savings are harder to check because nobody outside the bank can see the baseline or how the measurements were taken.

Why NZ businesses should pay attention

Australia's two largest banks by headcount are past the pilot stage. They're reporting measured results at their annual results, in front of investors. NAB owns BNZ and Westpac owns Westpac NZ, so the same tools and working methods are likely to reach New Zealand customers and suppliers, although neither bank has published a timetable for that.

Smaller New Zealand firms are adopting AI at a different scale with the same underlying question. A Prospa survey of 501 NZ SME decision-makers, reported by IT Brief on 26 October, found 59% now use AI in some form, up from 50% six months earlier. Among the non-users, 34% said they're unsure of the return on investment and 32% said they don't know which tools suit them. Only 12% named upfront cost.

NAB's structure is a practical answer to both of those barriers. The bank chose one approved general assistant for everyone, which settles the tool question for routine work. Separately, it put small business-led teams onto named processes where time can be measured before and after. A 40-person Auckland distributor won't form 50 squads, but the same two-track model works with one sanctioned assistant and two processes worth measuring.

NAB's approach also shows a weakness that NZ firms can avoid. Task-level time savings are easy to report and hard to turn into a business case. Twenty percent off memo drafting is worth money only if the freed hours go somewhere useful, or if the faster memo shortens the time a customer waits for a decision. Owners who measure the end-to-end outcome alongside the task time will have a stronger case when the renewal invoice arrives.

Practical next steps

• Choose one approved assistant. Put staff on a business-tier plan that has admin controls and contractual data terms, and write down which data classes can and can't go into it. Staff already using personal accounts should move onto the approved plan.
• Time two document-heavy tasks before you change anything. Quotes, credit applications, tender responses and month-end commentary are good candidates. Without a baseline you won't be able to report a percentage that anyone believes.
• Give each use case a business owner. NAB's squads are business-led. The person who owns the process, not the IT team, should decide whether a change has worked.
• Measure the outcome the customer sees. Track turnaround time or error rate for the whole process as well as the time saved on the AI-assisted step.
• Look for a migration or clean-up job with a fixed budget. Westpac's result came from converting legacy code and documentation, which is bounded work with a clear cost to beat. An ageing reporting database or an unmaintained integration is a reasonable first target.
• Plan the staff conversation early. Decide what happens to the hours that AI saves and tell staff before the tool goes live, not after the first restructure rumour.

If you want help choosing an assistant, setting data rules or scoping the first measured use case, our AI consultancy team works with NZ and Australian organisations on adoption and governance. Process work that comes out of those use cases sits with our process and automation practice.