Financial Technology Weekly: CBA puts a AUD200 million value on AI benefits, APRA penalises Bendigo Bank over cyber accountability failures, RBI demands AI model inventories from banks and IBM partners with OpenAI on enterprise deployment.
Commonwealth Bank of Australia (CBA) reported around AUD 200 million ($130 million) in gross benefits from AI use cases in FY2026 and expects the figure to double in FY2027. The disclosure provides a clearer measure of AI’s economic contribution, although CBA has not disclosed its AI-specific expenditure. ASX reported rising operating costs as it modernises critical trading, clearing and settlement infrastructure and addresses regulatory concerns.
Reserve Bank of India (RBI) Governor Sanjay Malhotra called on banks to maintain complete inventories of production AI models and establish board-approved governance policies. IBM and OpenAI targeted the integration of frontier models into core enterprise workflows, while Adyen developed payment infrastructure for agentic commerce. The developments place AI economics, governance and production controls at the centre of financial technology investment.
Read more on the week’s key developments:
1. CBA starts putting an economic value on AI
Commonwealth Bank of Australia reported around AUD 200 million ($130 million) in gross benefits from AI use cases in FY2026, including capacity reinvested elsewhere in the bank, of which around AUD 100 million was incremental in-year. Chief financial officer Alan Docherty said CBA had invested more than the benefits realised during FY2025 and FY2026, but expects gross benefits to double and exceed investment in FY2027. He said the bank had chosen to reinvest productivity gains, including faster code deployment, to deliver more technology change instead of immediately reducing resources.
CBA reported more than 20% growth in technology changes deployed and a 60% improvement in restoration time for critical incidents, although it did not attribute these outcomes solely to AI. CBA also does not disclose AI-specific expenditure within its AUD 2.4 billion ($1.56 billion) technology investment, preventing a standalone return calculation. Its FY2027 projection gives investors a clearer test of whether measured AI benefits can grow faster than the cost of deployment.
2. ASX’s technology rebuild keeps pressure on operating costs
ASX released its FY2026 full-year results on August 13 as it continued a multi-year programme to modernise critical technology, strengthen operational resilience and address regulatory concerns. Operating revenue increased 13.3%, while total expenses rose 21.1% to AUD 557.4 million ($362 million), contributing to a 180-basis-point decline in its EBITDA margin to 61.0%. Technology expenses increased 20.4% to AUD 93.3 million ($61 million), while ASX incurred AUD 30.8 million ($20 million) in costs associated with ASIC’s review of its governance, capability and risk-management practices.
Technology modernisation and regulatory remediation contributed to weaker operating leverage during the year. Costs associated with ASIC’s review are separately identifiable, but higher cloud usage, licensing fees and depreciation may remain as modernised systems enter operation. ASX expects expenses to rise by a further 18% to 21% in FY2027, with capital expenditure of AUD 180 million to AUD 200 million ($117 million to $130 million). The guidance points to an underlying cost base that will remain elevated beyond the inquiry.
3. APRA links Bendigo cyber controls to executive accountability
The Australian Prudential Regulation Authority commenced civil penalty proceedings against Bendigo and Adelaide Bank after it admitted breaching obligations under the former Banking Executive Accountability Regime in connection with a March 2023 cyberattack on its Alliance Bank business. The admissions concern customer authentication controls, security testing, information-security governance and the allocation of responsibility for the Alliance Bank IT system. APRA and the bank have proposed an AUD 8 million ($5.2 million) penalty, subject to Federal Court approval. APRA said the historical control weaknesses had since been remediated and it had no current concerns about the adequacy of the bank’s information-security controls.
The case shows how cybersecurity weaknesses can become accountability breaches when responsibility for an IT system is not clearly assigned and exercised. BEAR was replaced by the Financial Accountability Regime in March 2024, but the requirement to allocate responsibility to identifiable accountable persons remains. For banks with subsidiary or distributed operating models, the proceedings reinforce the need for accountability arrangements to cover the systems through which services are delivered.
4. RBI asks banks to maintain complete inventories of AI models
Reserve Bank of India Governor Sanjay Malhotra told banks at FIBAC 2026 to maintain a complete inventory of AI models in production and establish board-approved governance policies with clear accountability for outcomes. He also called for explainability in decisions that materially affect customers, red-teaming and stress-testing of AI systems and meaningful human oversight where errors could cause material harm. Malhotra said banks retain ultimate responsibility for decisions made using AI.
A model inventory gives banks and supervisors a practical starting point for identifying where AI operates, which decisions it influences and who owns the associated risk. It also supports validation, change control, incident escalation and assessment of bias, explainability, cybersecurity and third-party dependencies. Together with RBI’s June draft model-risk principles, Malhotra’s intervention places boards at the centre of this control structure and increases pressure on banks to expand their governance capacity as the number and materiality of AI models grow.
5. Adyen positions payment infrastructure for agentic commerce
Adyen used its first-half 2026 results on August 13 to place agentic commerce more firmly within its payments architecture. Adyen Agentic is intended to let enterprise merchants process payments securely across different AI-agent protocols, reducing the complexity of connecting separately with multiple agent ecosystems. The company has highlighted merchant concerns around security, data privacy and the potential loss of direct customer relationships as agents assume a larger role in purchasing.
As product selection and payment intent migrate to third-party agents, Adyen is positioning payments as a control point where customers can be recognised and authentication and fraud rules applied. Its participation in Google’s Agent Payments Protocol and the x402 Foundation supports this cross-protocol approach. In agentic commerce, payment-provider differentiation may increasingly depend on interoperability, identity and merchant control alongside checkout conversion.
6. ABN AMRO expands production AI as financial impact remains unclear
ABN AMRO reported on August 12 that it had almost 50 AI use cases in production. During the second quarter, it launched a generative AI voice bot for customer enquiries and a knowledge assistant for KYC and AML analysts. A week earlier, the bank partnered with Mistral to explore AI applications that meet its security, privacy, transparency and regulatory requirements while reducing dependence on non-European technology providers.
Almost 50 production use cases demonstrate broader operational adoption, while the Mistral partnership adds a European option to ABN AMRO’s model-supplier strategy. However, task-level productivity may increase capacity or improve documentation without reducing the overall expense base. The bank reported broadly stable underlying costs and 253 fewer full-time employees during the quarter, but did not attribute either outcome to AI. Unlike CBA, ABN AMRO has yet to attach a financial value to its expanding AI portfolio.
7. Better Home & Finance tests whether AI-native lending can generate operating leverage
Better Home & Finance reported second-quarter loan volume of $1.67 billion, up 38% year on year. Partner volume generated through its Tinman AI Platform reached $912 million, or 55% of total volume, up from 50% in the previous quarter. Net revenue increased 28% to $54.7 million and net loss narrowed 16% to $30.6 million. Adjusted EBITDA loss narrowed 39% to $14 million, including a $6.5 million reserve release related to older loans.
Tinman’s growing share of lending provides a measurable test of Better’s AI-native model because platform penetration can be compared with volume, revenue and losses. The results show greater platform scale, but limited evidence of operating leverage. Technology expense increased to $8.8 million from $6.4 million year on year, while the reserve release contributed to the EBITDA improvement. Stronger unit economics would require origination costs and operating losses to decline as platform volumes grow.
8. IBM and OpenAI target AI deployment across core enterprise operations
IBM announced a strategic partnership with OpenAI on August 13 to help enterprises deploy AI across core operations. It will embed GPT-5.6, Codex and ChatGPT Work into IBM Consulting Advantage and develop industry solutions, including for financial services. A dedicated OpenAI practice will support clients in redesigning legacy workflows, modernising applications and managing cybersecurity and AI risk. The agreement expands the companies’ existing cyber collaboration.
The partnership addresses the integration work required to connect frontier models with fragmented processes and legacy systems. Banks must determine how agents access data, which actions they can execute and how their outputs are monitored. IBM describes its cybersecurity delivery platform as vendor-agnostic, but deeper integration with OpenAI raises questions about model portability, service continuity and data access. IBM is positioning consulting and security as the control layer, showing how enterprise AI spending extends beyond models and computing capacity to integration, process redesign and governance.
9. Oracle and AWS extend database integration for critical workloads
Oracle and Amazon Web Services expanded their long-term collaboration on August 13 as Oracle AI Database@AWS reached 22 AWS Regions. Oracle also made Exadata Database Service on Exascale Infrastructure generally available on the platform, extending Exadata performance through a pay-per-use model. The service allows Oracle database workloads to connect directly with AWS analytics and AI services, while Oracle said organisations in highly regulated sectors, including financial services, are migrating critical workloads to the platform.
The arrangement reduces migration friction for enterprises that want to retain Oracle database technology while using AWS applications, analytics and AI services. It offers latency as low as 165 microseconds between applications and databases and integrates with Amazon Bedrock and SageMaker. For banks, this can support latency-sensitive and transaction-intensive workloads without first re-architecting applications or duplicating data. The architecture also creates a linked dependency across two providers, requiring incident ownership, resilience testing and exit planning to account for both.
10. Rakuten allocates AI development costs as FinTech profit rises
Rakuten Group reported on August 10 that second-quarter FinTech revenue increased 27.0% year on year to JPY 295.4 billion ($1.85 billion), while non-GAAP operating income rose 60.1% to JPY 69.2 billion ($434 million). Growth was supported by customer expansion, higher transaction values and cost controls. From the first quarter of 2026, Rakuten began allocating a portion of AI-related development costs to individual businesses and retrospectively adjusted segment results from the first quarter of 2025.
The change brings part of Rakuten’s AI investment into the reported cost base of its banking, cards, securities and other FinTech operations. This gives a more complete view of segment profitability, but Rakuten does not disclose how much AI expenditure was allocated to FinTech. The increase in operating income therefore cannot be attributed to AI productivity or used to calculate a return on AI investment. The change improves cost attribution while leaving the scale and financial contribution of AI spending unresolved.