The work rarely makes headlines, but it eats hours: an operations analyst toggling between a custodian portal and a portfolio accounting system, copying numbers by hand because the platforms won't talk to each other. Another clicks through a fund administrator's web interface, reconciling month-end reports one screen at a time. The software world calls this "swivel-chair work," and it survives in finance precisely because legacy systems never bothered with clean APIs.
Zomma, a San Francisco startup backed by Y Combinator, thinks it has found a way in. The company has built AI agents that automate these tasks not by integrating with the underlying systems, but by controlling them the same way humans do: watching screens, moving cursors, clicking buttons. The agents handle reconciliation between custodian portals and portfolio accounting platforms, process month-end packets, manage investor reporting, and navigate fund administrator interfaces across systems like Advent Geneva, Bloomberg Terminal, and Schwab Advisor Center, according to the company's website.
It's an approach that targets a stubborn reality in financial operations. Custodians, transfer agents, and carriers often provide web portals but no programmatic access. Operations teams end up manually shuttling data across dozens of these interfaces, a process that's tedious, error-prone, and expensive.
Zomma's agents read what's on screen, reason through the task using what the company describes as frontier language models, execute actions via simulated mouse and keyboard, then verify the results by reading the output again. The workflow mirrors how a human analyst would approach the same job, albeit faster and without the fatigue.
The company's approach carries an obvious risk: handing an AI agent control over sensitive financial systems. Zomma says it addresses this with mandatory human approval gates. "Every action stops for human approval before anything submits," according to the company's Y Combinator profile. When two-factor authentication prompts appear, the agents halt and wait for a person to clear them. "We never bypass MFA," the company states on its use-case page.
The system targets a range of finance operations work. Reconciliation between custodians and portfolio accounting systems, 1099 reconciliation, net asset value tie-outs, corporate action processing, regulatory filing preparation. Client onboarding across CRM and custodian portals. Periodic know-your-customer reviews. Compliance logging with append-only audit trails, the company says.
A blog post Zomma published in June detailed how the agents handle fund administrator and transfer agent portal automation for hedge funds and family offices. Another from May covered insurance carrier portal data entry, contrasting the startup's method with traditional robotic process automation and API integrations.
Building Trust in Automated Clicks
The security architecture reflects the sensitivity of what these agents touch. Zomma runs single-tenant deployments in its AWS us-east-2 region with TLS 1.2 or higher encryption in transit and AES-256 at rest, according to the company's security documentation. Customer data doesn't get used for model training. The system stores scoped credentials in a dedicated vault, integrates with 1Password, offers single sign-on through WorkOS, and maintains what the company describes as an immutable audit trail. SOC 2 Type II certification is pending.
In a June blog post, Zomma outlined what it calls the Agentic Trust Framework, describing four layers of control: input assurance, governed reasoning, policy-mediated action, and isolated execution. The architecture separates an orchestrator layer that makes policy decisions from ephemeral runtime instances that perform the actual work—a design meant to contain damage if something goes wrong.
Jihyun Kim co-founded Zomma and serves as CEO. Varun Arumugam and Cashel Fitzgerald are co-founders, according to the company's Y Combinator listing. The team has grown to five employees. Y Combinator partner Brad Flora works with the startup, which went through the accelerator's Summer 2026 batch.
"We already work with some of the most ambitious funds and fintechs," Kim wrote in a LinkedIn post last month announcing the Y Combinator backing. The company declined to identify specific customers. Y Combinator has scheduled Demo Day for the batch on September 10.
A Crowded Field, With Nuances
Zomma lands in a market where established players are adapting and new ones keep arriving. Incumbents like UiPath, Automation Anywhere, and SS&C Blue Prism have added what they call agentic features to existing robotic process automation platforms. Automation Anywhere launched pre-built "Autonomous Finance" solutions in May; SmartStream announced "Smart Agents" for banking back offices in June, according to press releases from both companies.
Newer entrants include Dispatch, which orchestrates account opening across platforms like Schwab, Fidelity, and Pershing, and Automatic.co, which deploys back-office agents for accounts payable and reconciliation. Skyvern, a Y Combinator alum from the S23 batch, builds browser automation using language models and computer vision.
Zomma's pitch emphasizes purpose-built finance workflows with those human approval checkpoints, rather than the citizen-developer scripting tools that some RPA vendors favor. The company offers both forward-deployed enterprise engagements and a developer API it describes as "Give your agent a computer," though access currently requires joining a waitlist.
Whether finance operations teams will trust AI agents to click through their most sensitive portals remains an open question. The technology works, at least in principle. The harder sell may be convincing compliance officers and risk managers that a machine watching a screen is safer than the analysts they've relied on for years—even if those analysts are drowning in swivel-chair work.
