Westpac halved a data warehouse migration with AI: what NZ can use

Westpac used its full-year results on 2 November to put a dollar figure on AI-assisted legacy migration. A slide in the results showed the bank's St.George data warehouse move costing around A$35 million against an original A$70 million budget, according to iTnews. Chief executive Anthony Miller told analysts the migration finished more than two years early, with AI tools speeding up code conversion, testing and documentation, according to the Investing.com earnings call transcript.

The St.George data now sits on ADAPT, Westpac's Azure-based analytics platform, and the business rules have moved to Westpac Intelligence, The Asian Banker reported on 6 November. The project belongs to UNITE, the group's technology simplification programme. Westpac expects about A$15 million a year in direct benefits from FY27, which is also the year the old warehouse is due to be decommissioned (iTnews).

How far the figures go

All of the numbers come from Westpac: a results slide and the chief executive's comments. The bank hasn't published how it calculated the saving or which AI tools it used. It also hasn't said what share of the conversion work the tools did, and nobody outside the bank has checked the claim. The A$15 million a year is a forecast. None of the benefit has been booked yet, and the legacy warehouse is still switched on. Westpac first flagged the two-year time saving at its September data, digital and AI investor day, so the November slide adds a cost figure to a claim investors had already heard (iTnews).

The warehouse move is one small part of a much more expensive programme. Chief financial officer Nathan Goonan told analysts that UNITE made up 44% of roughly A$2.1 billion in FY26 investment spend, and he expects UNITE to cost between A$850 million and A$950 million in FY27 (Investing.com). Miller warned that the next 12 to 18 months of UNITE will be demanding. One migration finishing early says little about whether the wider programme will come in on time and on budget.

Why warehouse migrations suit AI assistance

A large share of the work in a data warehouse migration is translation. An older warehouse holds years of stored procedures, extract and load jobs, and reporting logic written for one database engine or proprietary tool. The new platform needs the same logic expressed in a different form. Converting that logic is repetitive and well bounded, and AI coding tools handle that kind of task reasonably well. Documentation is often the hidden cost. The people who wrote the original logic have usually moved on, and engineers spend weeks working out what a 2,000-line procedure does before anyone can rewrite it. An AI tool can draft that explanation for an engineer to review.

Migrations also come with a built-in way to check the AI's output. A team can run the old and new pipelines side by side and reconcile the results. Any difference shows up in testing before cutover, so errors in converted code get caught by a mechanical check before they reach a finance report. Work with looser feedback, such as redesigning a business process, gives a team far less evidence about whether the AI's contribution helped.

What it means for NZ organisations

Few NZ firms run a warehouse at the scale of a big Australian bank, yet many carry the same kind of technical debt. Common examples are an on-premises SQL Server reporting database from the 2000s, a pile of SSIS packages, Crystal Reports nobody wants to touch, or an ERP with a decade of custom reports. Estates like these often stall modernisation, because working out what the old code does takes up most of the migration estimate.

Westpac's figures suggest the cost of that discovery and conversion work is falling. Some organisations have shelved a data platform move or an ERP replacement because rebuilding the legacy reporting looked too expensive. Those estimates are worth redoing. Integrators will increasingly quote AI-assisted migration, and buyers should ask what the tools do and how the converted logic gets checked. If you're weighing up where a legacy move fits in a wider digital transformation plan, start with the conversion and testing effort, since that's where the savings will show up.

The Westpac result also has limits for smaller organisations. The bank already had a cloud target platform in place, a large engineering team and established test environments. A mid-sized NZ business often has none of these, and a halved migration cost depends on a settled destination. When the target architecture is still undecided, AI-converted code only moves the existing mess onto new infrastructure sooner. Data handling matters too. Sending proprietary code and sample data to a hosted AI service raises the question of where that material is processed and kept, and a company should settle that question with the provider before any code leaves its environment. Our AI consultancy page covers the governance side of choosing these tools.

Practical next steps

• Inventory your legacy reporting and integration code: count the stored procedures, scheduled jobs and reports, and record which ones still have active users. Retire reports with no users before migration starts.
• Pull out any shelved migration business case and work out how much of the estimate was discovery and code conversion. That portion is the part AI tools are most likely to shrink.
• Run a contained pilot on one subject area, such as accounts receivable reporting. Convert it with AI assistance and reconcile the new outputs against the old system across a full month-end close.
• Ask any vendor quoting AI-assisted migration to name the tools, say where your code and data will be processed, and explain how converted logic is verified before cutover.
• Put a firm decommissioning date in the business case. Westpac's forecast benefits depend on switching off the old warehouse in FY27, and the running costs of your old system only stop when you switch it off.