Banks and insurers turn to partners for AI agents, with $450bn on the table by 2028

Finance firms lean on suppliers for AI agents - and it's working

Most banks and insurers aren't building AI agents alone. Two-thirds of more than 1,000 respondents in Capgemini Research Institute's World Cloud Report in Financial Services 2026 say they partner with suppliers for AI expertise and cloud platforms. Only 33% are going fully proprietary, 49% are mixing in-house with vendor tech, and 15% are buying agents off the shelf.

Real moves back this up. Vitality deepened its work with Google to launch an AI platform that helps people make better decisions about health and lifestyle. MUFG, Japan's largest bank, is pushing to become an AI-native company using agentic AI, rethinking data, and partnering with OpenAI and Sakana AI.

Growth is on the table - if firms can scale

The report estimates AI agents could add up to $450bn in economic value by 2028. Yet only 10% of banks and insurers have deployed agents at scale. 80% are still in idea or pilot stages. The signal: demand is there, value is clear, execution is the gap.

As Capgemini's Ravi Khokhar put it, the agentic era opens doors - but substance has to beat hype. That means clear use cases, strong governance, and humans working alongside agents.

Where AI agents deliver value today

Banks are already using cloud-native agents across core workflows:

• Customer service: 75%
• Fraud detection: 64%
• Loan processing: 64%
• Customer onboarding: 59%

Insurers show similar traction:

• Customer service: 70%
• Underwriting: 68%
• Claims processing: 65%
• Customer onboarding: 58%

Euroclear Sweden's CIO, Jörgen Olofsson, highlights the upside: agents automate repetitive tasks, improve support functions, and scale operations. Leaders at Liberty Mutual and Capital One report faster processing, more intuitive interactions, and shorter onboarding cycles - especially as policies and governance mature.

Why the partner-first, hybrid model wins

Speed, skills, and stack choices drive the split. External partners bring proven tooling, orchestration, security controls, and model access. Internal teams keep domain logic, data strategy, and risk controls close to the vest. The result: faster time-to-value without giving up strategic control.

This is also a hedge against vendor lock-in. A hybrid approach lets you swap components (models, vector stores, agent frameworks) as they evolve, while standardizing interfaces and observability in-house.

Governance first: keep humans in the loop

Nearly half of firms are creating AI supervisor and coordinator roles, and 92% of executives admit there's a leadership skills gap. 48% are standing up supervisory roles; 46% are reskilling and reallocating staff. Markel's CTO Lloyd Scholz is direct: in regulated settings, human oversight, explainability, and governance are non-negotiable.

What to do next: a practical playbook for CIOs, CTOs, and product leads

• Prioritize the right use cases: Pick 2-3 workflows with repetitive steps, clear guardrails, and measurable outcomes (e.g., onboarding cycle time, claim touch-time, fraud loss rate).
• Go hybrid by default: Partner for platform and safety tooling; build proprietary flows where your data and IP create edge. Bake in model-agnostic architecture from day one.
• Data readiness beats model obsession: Define golden sources, retrieval patterns, and lineage. Establish PII policies, masking, and data minimization for agent contexts.
• Codify governance: Create policies for agent scopes, escalation thresholds, audit trails, and bias testing. Stand up an AI risk review with compliance and model risk functions.
• Staff the "AI babysitters": Appoint agent supervisors to review actions, tune prompts/tools, and monitor drift. Train product owners and ops teams on agent behavior and failure modes.
• Instrumentation from the start: Track containment, handoff rates, task success, latency, customer satisfaction, and model costs. Set business targets per use case.
• Security and access control: Enforce least privilege for tools/actions. Isolate agent sandboxes. Run vendor risk assessments and contract for data residency and audit rights.
• Scale with guardrails: Move from pilot to production with phased rollouts, canary traffic, and A/B testing. Automate evaluations with synthetic and real scenarios.
• Vendor contracts that age well: Negotiate IP ownership, model-switch rights, rate limits, and SLA penalties. Require observability hooks and export of logs/metrics.

Reference and upskilling resources

For the full industry view, see the Capgemini Research Institute's analysis of financial services and AI agents here. Building internal capability is just as critical - explore finance-focused AI tools and training options here.

The bottom line

Partnerships are the default. Hybrid is the strategy. Governance is the moat. With only a small fraction operating at scale today, firms that combine vendor speed with disciplined in-house control will set the standard - and capture outsized value by 2028.