Financial services has the highest chatbot adoption of any sector yet one of the lowest resolution rates, and Parloa co-founder and chief executive Malte Kosub argues that gap is the direct result of a deliberate strategic choice, not a technology shortfall.
Speaking to The Fintech Times, Kosub cited findings from Parloa’s State of Agentic CX report: banking and insurance achieved 64.2 per cent chatbot adoption but resolved only 7.4 per cent of customer issues. His explanation is blunt. “The important distinction to make is that the measure of success was never resolution; it was containment,” he said. “That’s a fundamentally different problem from how do we help this customer get an answer?”
The deflection economy
The report found that 65.7 per cent of classifiable chatbots in financial services were rule-based systems, decision trees with no genuine reasoning capability. Kosub argues these were sold and bought as AI without meeting any meaningful definition of the term. The procurement environment enabled it: if a system deflected enough inbound calls to justify a cost-reduction case, the buying decision was made. Whether the customer actually resolved their issue was, in many cases, not tracked with equivalent rigour.
He points to a structural governance failure behind this pattern. For much of the first decade of chatbot deployment, the purchase was classified as an IT or digital transformation decision. Customer experience functions were consulted late or not at all, and the people measuring success were not the same people measuring customer satisfaction. Neither group was primarily accountable to revenue. “When the CFO is focused on headcount reduction, but the IT department is measured on system uptime, genuine customer resolution falls through the gap,” Kosub said.
The report also found that 96 per cent of dedicated support lines in financial services still use legacy IVR infrastructure, and that 89.9 per cent of escalation attempts failed: the customer asked to speak to a human and the system could not complete the handoff. Kosub frames this last figure as a direct regulatory exposure. In the current FCA environment, institutions that cannot demonstrate genuine access to human agents when customers need them will find that position increasingly difficult to defend.
The readiness gap
Kosub’s benchmark for what comes next is agent-to-agent readiness, a model in which a customer’s personal AI contacts an enterprise’s service infrastructure directly. The report found only 1 per cent of enterprises were ready for this. The barriers are compound: authentication frameworks that do not support non-human callers, APIs that cannot execute real transactions, and workflow logic locked inside legacy IVR trees that cannot be extended.
The broader competitive context matters here. Several enterprise software vendors, contact-centre platform providers and AI-native startups are competing in the same agentic customer-service space Parloa occupies. The distinguishing factor Kosub identifies in Parloa’s existing clients, which include Allianz, Swiss Life and Barmenia, is not the AI model selected but whether the underlying service systems could support real-time data access. Institutions that had modernised their core policy or account infrastructure before deploying an agent achieved genuine resolution capability. Those that had not found themselves training an agent to navigate around systems it could not connect to.
The sceptical case, which Kosub addresses directly, is that agentic AI is the same deflection promise with updated vocabulary. He offers a conditional acceptance: if enterprise resolution rates in banking and insurance have not risen substantially from 7.4 per cent to above 40 per cent for AI-handled interactions within the next two years, and if human escalation success rates have not improved from the 10 per cent currently observed, he would accept the verdict. His argument for why this cycle is different rests on infrastructure rather than intent: large language models can reason across context and ambiguity in a way that decision-tree systems cannot, and customers who use AI tools in their daily lives now have a precise calibration for what inadequate AI looks like. The tolerance for scripted non-answers has, in his assessment, dropped considerably.
For financial services leaders, the practical implication is sequencing. Technology replacement is the visible layer, but Kosub’s analysis suggests the prior question is organisational: who owns the agent’s scope, escalation rules and fallback behaviours, and are those people accountable for customer outcomes as well as cost lines.
The post Parloa CEO: Bank Chatbots Built to Deflect, Not Resolve appeared first on The Fintech Times.
