Pre-deployment AI project cancellations rose 33% between late 2024 and late 2025, even as 59% of deployed AI initiatives are delivering tangible business value, Dennis Gada, Executive Vice President and Global Head of Banking & Financial Services at Infosys, told TNIE.
Gada said the increase in cancellations reflected stricter scrutiny of projects before they were deployed, with banks placing greater emphasis on whether AI initiatives could demonstrate measurable business outcomes.
“This surge in early-stage pruning is healthy; it demonstrates that institutions are failing fast and cheap rather than funding multi-million-dollar science experiments,” Gada said.
He said the gap in banking was no longer simply between institutions experimenting with AI and those putting it into production.
“The real divide across global banking is no longer experimentation versus production—it is between institutions that can make AI work inside a controlled sandbox and those equipped with the enterprise data architectures, governance frameworks, and operational discipline to run it dependably at scale,” Gada said.
According to the Infosys Bank Tech Index, 59% of deployed AI initiatives are already delivering tangible business value.
Gada said projects that fail to reach production often face three problems - the absence of measurable unit economics, fragmented data and treating AI as an IT project rather than changing the underlying workflow.
“Models cannot compensate for siloed, uncurated, or stale data,” he said.
At the same time, Gada said the nature of banks’ AI spending was changing, with investment moving away from model access, experimentation licences and standalone tools towards the systems needed to support AI deployment at scale.
“We are seeing a decisive capital rotation. The initial wave of spending was dominated by model access, experimentation licenses, and standalone front-end tools. Today, capital is flowing into the underlying plumbing required to make AI enterprise-grade: unified data platforms, hybrid cloud environments, zero-trust cybersecurity, and legacy core modernization,” Gada said.
Banks had also recognised that foundation models were becoming less of a differentiator, he said.
“The sustainable differentiator is not the base model; it is the bank’s proprietary data estate, the orchestration middleware, and the governance guardrails governing execution,” Gada said.
The projects moving fastest into production are largely those involving high-volume and rules-based work, including transaction surveillance, anti-money laundering triage, contract parsing and software testing, Gada said.
“These environments offer deterministic ground truths, continuous audit logging, and measurable operational savings without exposing the bank to black-box decision risk,” he said.
Banks, however, were taking a cautious approach to systems that carry out multi-step tasks across different systems. Such systems were being used for internal workflow triage, with human checks retained before financial commitments or ledger postings were completed.
Areas involving unsupervised customer-facing financial advice, algorithmic credit decisions and autonomous capital allocation remained restricted because of regulatory and explainability requirements, Gada said.
Customer service represented the highest single source of value creation from AI in banking at 22%, according to Gada, while software engineering and cybersecurity were among the areas producing the largest operational cost savings.
