Goldman Deploys Claude Agents as Lloyds Targets £100M in Value From 2026 AI Plan
Financial deployments are making multi-step AI workflows more concrete, while raising the need for systems that can identify, coordinate and constrain software acting across services.
Listen to this story
The audio brief
Story brief
3 key pointsFinancial institutions are moving agentic AI from demos into bounded workflows, but the evidence spans very different levels of commitment. Goldman has deployed Anthropic’s Claude for trade accounting and client onboarding; Lloyds’ £100 million figure remains an expected benefit from a planned 2026 enterprise rollout. Claims automation can cut processing from days to minutes while keeping humans involved. For...
- 01
Lloyds expects £100 million from fraud detection and back-office automation, but the target is not reported realized value.
- 02
The World Economic Forum and Accenture study covered more than 150 senior leaders across 100 financial institutions.
- 03
a16z’s OpenRouter analysis found agentic inference the fastest-growing behavior across 100 trillion tokens; it is not an industry benchmark.
Goldman Sachs is deploying autonomous agents powered by Anthropic’s Claude for trade accounting and client onboarding. Lloyds Banking Group has committed to an enterprise-wide agentic AI deployment in 2026, expecting £100 million in value from fraud detection and back-office automation.
The difference matters: Goldman’s cited uses are a deployment, while Lloyds’ rollout and its value figure are plans and an expectation for 2026. A World Economic Forum report developed with Accenture, drawing on more than 150 senior leaders at 100 financial institutions, found agentic systems already initiating payments, automating claims and managing client onboarding on behalf of businesses.
Autonomy is not uniform. Allianz Partners reduced claims-processing time from days to minutes with an AI claims tool while retaining human oversight. That example separates workflow speed from a decision to remove people from the process.
When a request becomes a chain of actions
PYMNTS describes AI agents as a fast-growing class of API consumers. Instead of making one request for a person to assess, agents can query endpoints, process responses and take actions through extended sequences without a human in the loop. That puts pressure on API arrangements built around human developers and users.
An a16z analysis of more than 100 trillion tokens of real-world OpenRouter usage calls agentic inference the fastest-growing behavior. It describes models planning, retrieving context from APIs, revising outputs and iterating across multiple turns. OpenRouter processed more than one trillion tokens a day as of late 2025.
The analysis argues that the competitive test is expanding beyond model accuracy and benchmarks to orchestration, control and reliable operation across extended workflows that touch multiple APIs. That is a view from the analysis, not an industry-wide performance standard.
Lloyds expects this value from fraud detection and back-office automation under its planned enterprise-wide agentic AI deployment in 2026.
The services proposed around autonomous software
MIT Media Lab associate professor Ramesh Raskar argues that the larger opportunity is not limited to task-specific agents. In his view, marketplaces, protocols and services that let agents operate at scale will be essential infrastructure.
- Identity and discovery systems, so agents can be identified and found.
- Trust and reputation services, plus insurance and legal services for mistakes.
- Stablecoin-based micropayment infrastructure for machine-speed transactions.
Raskar leads MIT’s Project NANDA, which is working to keep the agent web open before corporate consolidation. His framework identifies a proposed direction for the market, rather than a settled technical or legal standard.
Sources
- pymnts.comAI Agents Become the API Economy’s Biggest New Customers | PYMNTS.com
