Rulebase

Leaderboard, task breakdowns, and recorded trajectories are live at bankarena.ai/. If you're training computer-use agents or building finance models, we'd love you to evaluate on Bank Arena - dm or [email protected]!

On the tasks all five models ran, full completion was: Gemini 4 Argon: 30.1% Claude Opus 5.5: 23.8% GPT-6 Astra: 15.0% GPT-6.1 Sol: 11.6% Grok 4.7: 5.4%

In 33 of 52 tasks, new information arrives after they've acted, forcing them to revisit earlier decisions. In one task: a customer's €85k transfer shows as canceled, so they send another. Then the first settles, the duplicate comes back, and a fraud signal means the first needs to be recalled. Some easy ways agents get this wrong: credit twice, recall the wrong transfer, or close the case early.

We built a banking environment, simulating real-world banking applications with exact UI and workflow fidelity: a real ledger, policy and evidence vault, work tracker, ticketing tool, and ops desk. Agents get screenshots and URLs, then click and type across payments, policies, work tracking, and customer tickets.

We're launching Bank Arena: 52 long-running banking tasks across several banking apps. We gave frontier computer-use agents a bank to run, from recovering transfers to settling merchant payouts. The median task takes experts ~75 minutes to complete, and the highest-scoring models completed <31% of tasks. Here's what we found ↓

Here's one of many stories of how some of the fastest-growing fintechs use @RulebaseHQ to scale their customer operations. A year ago, Rho reviewed only a tiny fraction of customer interactions by hand and caught issues weeks late. Today, Rulebase reviews 100% of interactions across phone, email, and chat in real time, then acts on what it finds. When an interaction shows a customer at risk, a repeat contact, or an SLA about to slip, a Rulebase agent picks it up, loops in the owning team, and follows through to resolution. The results: 4x more issues caught, 3x fewer escalations, and ~90% less manual review.

New website: rulebase.co. Thanks to all our customers and partners.

We launched Revenue on @RulebaseHQ today to help fast-growing fintechs onboard more customers and merchants. Our agents connect with your existing tools to run end-to-end onboarding: collecting and following up on missing documents, reactivating dormant accounts, and much more across voice, email, SMS, WhatsApp and more!

Excited to share that we raised $2.1M to build the future of CX and back-office operations in financial services. For fintechs and banks, customer service back-office work remains a maze of manual processes, including quality assurance, dispute investigations, and fraud and compliance checks. At @RulebaseHQ, we built Coworker to work alongside CX and ops teams to: - Review 100% of interactions for compliance and quality assurance - Automatically file and track fraud and dispute cases end-to-end across CX, case management, and back-office platforms - Proactively follow up with teammates and third-party partners to keep SLAs on track - Flag compliance risks in real time and generate audit-ready reports. Coworker bridges CX and back-office ops, automating the most manual tasks so nothing slips through the cracks and every customer issue is resolved quicker and more effectively. A massive thank you to our investors, customers, and the growing @RulebaseHQ team for believing in our vision and being part of this journey! Our round was led by @LorenStraub from @BoweryCapital, with participation from @ycombinator, @CommerceVC, @transposevc, and founders of t3 chat, Atrium, Cron (now Notion Calendar): @theo, @raphaelschaad, and more. Incredibly grateful to have the support of the @ycombinator community and mentors, @dflieb and @koomen. Thank you all for being incredible partners on this journey!

I'd love to speak with customer services/fraud teams at fintechs. We'll work with you closely to match the risk models and scoring to your specific needs and use cases. Contact me directly at cal.com/chidi-rulebaseco/30m….

After the call, we generate an in-depth report on all flagged calls for the fraud team to investigate further. In the video, I show an example of a malicious caller using a VoIP system to change a customer's details. But Radar works for many other scenarios involving attacks on customer services.

During the call, we run a real-time analysis of voice and behavioural characteristics: - Does the call match a known fraud script? - Is the caller using stolen info to take over a customer's account? - Does the voice match a sample previously added to a watchlist?