Banks don't have an AI accuracy problem — they have a production-readiness problem

The definition gap

Invisible Technologies' guide for banks this week names the real blocker in bank AI, and it is not model quality. Banks often fail to deploy AI, the firm argues, because they enter vendor selection without a working definition of production-ready — and production-ready AI in banking is an operational standard, not a technical one. The argument is that POCs validate model accuracy in controlled conditions, which differ significantly from production environments; judging a platform by POC accuracy is measuring the wrong thing. m:…4963

The framework that follows is all operational plumbing: shadow-mode pilots running 30 to 90 days to capture edge cases and seasonal patterns, exit and transition planning in every vendor evaluation to mitigate vendor concentration risk, and independent validation of vendor methodology under SR 11-7 guidance. It is a vendor's guide, so read it as the standard its author wants to sell against. But the core claim survives scrutiny — in regulated banking, the go-live decision is a risk decision, and accuracy is only one input. m:…4963

The bank that did it

Itaú is the worked example of what the operational standard looks like when it clears. The Brazilian bank is deploying generative AI across its operations with a dual-platform approach: the ia.i conversational interface for customers and IARA, a governance and orchestration platform managing access to different AI models internally. The rollout numbers, per OpenAI's account: ia.i reached approximately 300,000 customers in its first phase, with a stated aim of reaching 100% of the customer base by the end of 2026. The bank describes its strategy as 'AI First' for both customers and employees. m:…8129

The internal numbers are the more instructive ones. Itaú's legal solution runs on a repository of more than 90 million documents and delivered efficiency gains of 20% to 80% across its processes. That range — 20 to 80 — is wide enough to be honest about variance across workflows, and it is the governance layer, IARA, that makes the variance manageable: controlled model access inside the bank's architecture rather than AI bolted onto the side. Note the source: this is OpenAI telling Itaú's story, so the framing favors the platform. The numbers are what they are — company-reported, directional, but concrete. m:…8129

The plumbing underneath

And beneath the governance platforms sits the data layer, which got its own production-grade upgrade this week. Databricks announced general availability of multi-statement transactions on its lakehouse — atomic, multi-step workflows with row-level concurrency, running natively across both Lake and Apache Iceberg tables and respecting Unity Catalog attribute-based access controls, including row-level security and column masking. The production proof points are other companies': Fivetran commits SCD Type 2 record updates as single atomic units with it, and Accolade processes approximately 1,700 transactions per day in production using multi-statement, multi-table transactions. m:…5655

This matters for bank AI more than it looks. Agent workflows that write — updating records, committing decisions, reconciling accounts — need the same atomicity guarantees as the systems they automate. A lakehouse that cannot transact is a lakehouse your AI cannot safely write to. Production readiness, again, turns out to be an operational standard. m:…5655

Why it matters

Stack the week and banking AI's real story is the distance between demo and deployment — measured in definitions, governance layers, and transactional guarantees. Invisible says define production-ready operationally before you buy. Itaú shows what the buying looks like at scale: a customer interface, a governance platform, and a dated commitment to the full base. Databricks ships the atomicity the write-heavy AI workflows will need underneath. For bankers evaluating AI vendors, the question list writes itself: show me the shadow pilot plan, the exit plan, the governance layer, and the transaction guarantees. Accuracy is table stakes. m:…4963 m:…8129 m:…5655