Arva AI, a New York startup that builds AI models and agents for banks, raised $20m in a Series A round led by Arthur Ventures. The company announced the round on 17 November 2026.
Y Combinator and Gradient, both earlier backers, also took part. Arva did not disclose a valuation or how much it has raised in total.
“This year our customers started handing real critical decisions to Arva.”Rhim Shah, Co-Founder and CEO, Arva AI
What Arva AI does
Arva builds software that takes over decisions in banks' back offices, starting with financial-crime compliance, where analysts check the businesses and people a bank deals with. Its first in-house model, Arva Intel, collects information on companies and individuals from online sources. Arva said that in its own testing the model was 13% more accurate than leading general-purpose models. It has not published independent results.
Arva Intel works with AgentCore, a continual-learning system that adjusts the agents to each bank's policies and past decisions. The company argues that banks need specialised AI whose conclusions can be explained to auditors and regulators.
Arva named a top-10 US bank, an S&P 500 payments company and the fintech Wise among its recent deployments.
Who is behind it
Arva was founded in 2024 by Rhim Shah, the chief executive, who previously led the financial-crime product at Revolut Business, and Oliver Wales, the chief technology officer, who was formerly chief product engineer at Seal. The company went through Y Combinator's Summer 2024 batch. Its Y Combinator profile lists New York as its home and a team of 20, and its own website says it works from London and New York.
Patrick Meenan, a general partner at Arthur Ventures, said banks need AI they can rely on for their most consequential decisions. He pointed to the team's combination of banking knowledge, technical skill and focus on accuracy and auditability.
What the money is for
Arva said the money will go first to Arva Research, the model-development group it launched in September 2026, to build more models and extend AgentCore. Shah said that how much work banks hand to agents depends on accuracy, and that this is why research gets the funding first.
The company also plans to extend its agents to more back-office decisions, including disputes, payment operations and lending. Shah said banks still do most of that back-office work by hand.
Meenan said Arthur Ventures would work with Shah and his team as they build models and infrastructure for critical banking operations.
Why Arva says accuracy matters
Arva's case rests on accuracy and auditability rather than speed alone. A bank that lets software make compliance or lending decisions still has to explain each outcome to auditors and regulators. Arva argues that general-purpose models are not built for that, and says its domain-specific models, together with AgentCore's tuning to each bank's own policies, are designed to fill the gap.
Shah described the company's aim as building the models that let banks hand their back-office work to agents and trust the result. Pulse 2.0 and FF News both cited a Boston Consulting Group estimate that AI agents could cut banks' costs by 30% to 40% by 2030.
Company website: arva.ai
