Finance Hits the Same AI Wall as Insurance—85% Want Agents, Only 25% Can Trust Them

Agentic AI is hitting the same wall in finance that we’ve seen in insurance: enterprises want autonomous agents, but they can’t trust them enough to deploy at scale. We reported recently on how insurance companies are pushing agentic AI despite only 40% of employees feeling prepared to use it, with 70% of scaling challenges being organizational rather than technical. Finance faces an identical problem—85% of businesses want to operate as agentic enterprises, but fewer than 25% have mature governance frameworks to make that happen safely.

The core issue is explainability. When an AI agent recommends a portfolio allocation or flags a compliance issue, auditors need to trace exactly how that conclusion was reached. If the reasoning trace is opaque, the recommendation is worthless regardless of accuracy—regulatory requirements and fiduciary responsibility demand transparent decision-making. This is the automation opacity problem: adding more agents often creates more complexity than value when you can’t verify their logic.

Sentient, an open-source AI lab, launched Arena today to address this directly. Arena is a production-grade stress-testing environment that deliberately feeds agents incomplete information, ambiguous instructions, and conflicting sources—replicating the messy reality of actual corporate workflows. Instead of just scoring whether agents produce correct outputs, Arena records full reasoning traces so engineering teams can debug failures and understand why agents make specific decisions.

Franklin Templeton, which oversees over $1.5 trillion in assets, is among the partners testing Arena alongside Founders Fund, Pantera, alphaXiv, Fireworks, Openhands, and OpenRouter. Julian Love, Managing Principal at Franklin Templeton Digital Assets, framed the shift clearly:

“The question is no longer whether these systems are powerful or if they can generate an answer, but whether they’re reliable in real workflows.”

That phrasing—”reliable in real workflows”—echoes exactly what we found in insurance deployment challenges. As we covered in our analysis of agentic AI in insurance, the technology delivers impressive pilot results (30% efficiency gains, 23-day faster processing, 65% fewer complaints), but only 7% of insurers have scaled beyond pilots. The pattern repeats: technology works in controlled environments, fails when exposed to messy reality, and organizational readiness lags technical capability.

The finance-specific twist is regulatory compliance. Insurance companies face regulatory requirements too, but finance workflows often involve fiduciary duties where unexplainable decisions create legal liability. When an agent recommends portfolio changes or flags suspicious transactions, “the AI said so” isn’t acceptable documentation. Auditors need complete reasoning traces showing exactly which data points influenced which intermediate conclusions leading to final recommendations.

Himanshu Tyagi, Co-Founder of Sentient, emphasizes this production reality:

“AI agents are no longer an experiment inside the enterprise; they’re being put into workflows that touch customers, money, and operational outcomes. It’s not enough for a system to be impressive in a demo. Enterprises need to know whether agents can reason reliably in production, where failures are expensive, and trust is fragile.”

The current state mirrors what we’ve documented across AI enterprise adoption: businesses run an average of twelve separate agents, frequently in silos, without coordinated governance. This fragmentation creates the same problems we identified in insurance—impressive individual agent capabilities that don’t translate to organization-wide impact because integration and oversight mechanisms don’t exist.

Arena’s approach—stress-testing with incomplete information and recording full logic traces—addresses the trust deficit directly. Rather than optimizing for demo performance, it replicates production complexity: ambiguous instructions, conflicting data sources, incomplete context. If agents can maintain explainable reasoning under those conditions, they might actually be deployable at scale.

The partnership roster suggests serious institutional interest. Franklin Templeton managing $1.5 trillion doesn’t partner with experimental platforms lightly. Asset managers, crypto funds (Pantera), and venture capital (Founders Fund) all face similar explainability requirements despite operating in different regulatory environments. They need agents that can justify recommendations, not just generate them.

Open-source development models like Sentient’s (which also architects frameworks like ROMA and the Dobby model) offer advantages here. Rather than relying on proprietary black boxes from major AI labs, financial institutions can adapt open-source agent capabilities to private internal data while maintaining full visibility into decision-making processes. This matters for compliance—regulators are more likely to accept AI-driven decisions when the institution can demonstrate complete understanding of how those decisions were generated.

The gap between ambition and execution remains large. Three-quarters of businesses plan to deploy autonomous agents, but only a quarter have governance frameworks mature enough to do so safely. This isn’t a technology problem—it’s the organizational readiness challenge we keep encountering across insurance, finance, and enterprise AI adoption generally. Technology advances faster than institutions can build oversight structures, training programs, and integration frameworks.

Arena represents an attempt to close that gap by providing infrastructure for systematic agent evaluation before production deployment. Whether it succeeds depends on whether financial institutions will invest as heavily in building trust mechanisms as they invest in deploying agents. Based on what we’ve seen in insurance, that organizational investment often lags behind technology spending—which is why 93% of insurance agents still haven’t scaled despite years of effort.

For discussions on AI in finance, enterprise agent deployment, and regulatory compliance, join our WhatsApp community where financial technologists discuss implementation challenges.

Related: Insurers Push Agentic AI While Employees Don’t Trust Basic AI—The Disconnect Deepens

Related: Insurance Executives Love AI—But Employees Don’t Trust It