Study Finds Gaps in Banks’ AI Context

Study Finds Gaps in Banks’ AI Context
Wednesday, November 18, 2026
  • •IDC-commissioned study finds more than 8 in 10 organizations see room to improve AI data
  • •Only about 9% broadly embed context in AI workflows; around 17% broadly implement knowledge graphs
  • •Nearly three in four consumers using AI for financial tasks turn to ChatGPT, Claude and Gemini before bank chatbots

More than 8 in 10 organizations in banking, insurance and government say the data feeding their AI needs improvement, and 43% say it needs significant improvement or a complete overhaul, according to an IDC study commissioned by AI software company Quantexa and published Tuesday. IDC surveyed 366 technology leaders at large banks, insurers and government agencies. The study found that banks and insurers are deploying agents in workflows faster than they can explain the context behind the agents’ reasoning. Context includes the instructions, rules, knowledge and memory agents need to make decisions.

Agents are expected to take on nearly one-third of work within the next year, but only about 9% of companies have broadly embedded context in decision-making and AI workflows. IDC’s Stewart Bond said organizations have invested in context building blocks, but those capabilities remain in separate pockets. Only about half of organizations implement multiple context capabilities, mostly for specific use cases; around 17% report broad implementation of knowledge graphs, which map reasoning paths for AI.

Data silos also make it harder for banks to compete with third-party financial tools. Nearly three in four consumers using AI for financial advice and tasks turn to ChatGPT, Claude and Gemini before their bank’s chatbot, according to McKinsey’s Global Banking Annual Review 2026. Meta’s Muse agent reads consumer balances and transactions through Plaid, while PayPal added its wallet to ChatGPT checkout. The study says integrating siloed data, especially lineage and decision data, is a major challenge: without it, banks cannot explain how agents reached decisions. Bond said missing data and decision lineage, rather than model complexity, most limits trust in AI decisions.

The study recommends making data lineage and decision data design requirements and defining which decisions agents will own, with a human in the loop. At Microsoft Ignite on Tuesday, Salesforce and Microsoft launched Copilotforce to bring Salesforce customer context and AI capabilities into Microsoft’s software environment.

Subscribe to the AIB Newsletter

The fastest, most reliable AI news, straight to your inbox