The Opportunity and the Bottleneck
BCG’s research with OpenAI finds that agentic AI could increase retail banks’ profitability by 30% and cut costs by 30% to 40% by 2030. The bottleneck isn’t the front-end technology. Banks have spent decades digitizing customer-facing channels, but translating outputs from already-trusted systems, identity checks, fraud signals, credit data, still depends on humans manually reconciling and summarizing that information before a decision gets made. That gap between validated data and actionable decisions is where the report argues the real cost and delay sits, not in a lack of automation itself.
Why This Matters More in the Back Office Than the Front
The report’s central claim is that the back office, not the customer-facing layer, is where agentic AI delivers the most immediate impact. In credit onboarding, an AI agent can evaluate identity verification, sanctions screening, fraud signals, and credit bureau data to produce a structured risk summary for human underwriters, operating within existing risk and compliance frameworks rather than replacing them. For collections operations handling exception-heavy, document-dependent cases daily, this reframes the priority: the visible customer conversation matters less than whether the underlying case data can be reconciled and summarized reliably before a human ever needs to intervene.
The Nuance Most Deployments Skip: Evaluation Has Four Distinct Dimensions
BCG’s framework for evaluation-driven development breaks agent reliability into four measurable categories: LLM-level factuality, faithfulness, recall, and precision; agent-level tool call accuracy, planning quality, and escalation accuracy; business-level return users, productivity increase, and completion rate; and system-level latency, error rate, and time to first token. The report’s underwriting example is the clearest illustration of why this matters: an agent can execute every tool call correctly, verify income, pull the right documents, and still fail by never explicitly assessing credit history, a planning failure invisible to anyone only checking whether individual steps ran successfully. Testing whether an agent completes tasks is not the same as testing whether it completes the right sequence of tasks, and most deployments only check the former.
What This Means for Collections and BFSI Operations
BCG’s middleware argument is the most transferable point for collections teams building their own agentic capability: a single control layer should sit between AI applications and the bank’s core systems, GitHub, identity providers, model APIs, standardizing logging, permissions, and audit trails so every AI workload is governed consistently rather than each team building bespoke integrations. The report’s “do/don’t” guidance reinforces this: build on how employees already use AI rather than banning it into shadow use, pair ambition with incremental steps rather than only small siloed pilots, secure real C-suite fluency rather than delegating AI entirely to IT, and treat model risk management as a design input from day one rather than an afterthought bolted on before launch. For a collections operation evaluating its own AI rollout, the sequence BCG describes, controlled evaluation first, centralized governance second, scaled deployment third, is the difference between an agent that reduces cost and one that quietly accumulates risk no dashboard is built to catch.