Evidence Register
A public register of AI-agent deployments in finance, each graded announced, piloted, live or at scale, with outcomes where they are published.
What works.
Agentic Finance Evidence Programme
Most claims about AI agents in finance are announcements, not evidence. Few deployments publish outcomes, failures go unrecorded, and regulators’ own evaluations are still to come: the FCA plans to publish its evaluation of AI Live Testing between January and March 2027. City Agentic Finance Lab (CAFiL) runs this programme to answer five questions, across all four themes: what works, what fails, how value is created, how to govern, and how to explore safely.
What we produce
Failure Taxonomy
Why agentic projects stall, are rolled back or are cancelled, built from public records and from interviews under the Chatham House Rule. Gartner forecasts that over 40% of agentic AI projects will be cancelled by the end of 2027.
What fails.
Value Benchmarks
Common measures of where agents create measurable value: cost to serve, time to decision, error and complaint rates, customer outcomes.
How value is created.
Governance Playbook
Practical controls for accountable deployment under the Senior Managers & Certification Regime, the Consumer Duty and the EU AI Act, including “the Autonomy Ladder”.
How to govern.
Agent Sandbox
An open multi-agent simulation testbed with synthetic customers and markets, building on our members’ long-standing work in agent-based simulation. No live money.
How to explore safely.
Agentic Readiness Diagnostic
A 20-question self-assessment for boards and leadership teams, also available as a facilitated diagnostic.
Where to start.
CAFiL publishes the first highlights in the Agentic Finance Outlook 2027 in December 2026; the public Evidence Register dashboard follows in 2027.
How we grade deployments
- Announced
- stated publicly; not yet in use.
- Piloted
- tested with real users or transactions in a limited, controlled setting.
- Live
- available to customers in normal operation.
- At scale
- live across a large share of the relevant customers or volume.
Each entry links to its sources and shows the date it was last checked.
Contribute to the Evidence Register
Are you deploying AI agents in a regulated financial firm? Tell us about it — we publish only what you approve.