# Zomma: full text for language models Zomma is an applied AI lab that builds computer-use agents for the enterprise. Each agent has its own computer, reads the screen and works in the tools a team already uses, with no integration. Every agent is proven on a private benchmark built from the customer's own past cases before it goes live, and again after every change. Zomma starts where mistakes are expensive: compliance, reconciliation and risk work in financial services. Zomma is not a KYC identity-verification vendor, not a screening-list provider, not a transaction monitoring system, not a rules engine and not an RPA tool. It works on top of the lists, systems and tools a team already uses and does the reading and first write-up a person would otherwise do. It recommends; a person decides. Website: https://zommalabs.com Contact: hello@zommalabs.com Short index: https://zommalabs.com/llms.txt Pages, in site order: Home, Enterprise deployment, How computer use works, Security, Agent benchmarks, Company, Blog. Subprocessors: https://zommalabs.com/subprocessors --- ## Home (https://zommalabs.com/) AI agents for operations that can’t fail. Zomma builds computer-use agents for the enterprise. Each one is proven on a private benchmark from your own past cases before it goes live, and again after every change. Zomma is an applied AI lab. We turn frontier models into computer-use agents that enterprises can trust with real work. Built by a team from Goldman Sachs, J.P. Morgan, Apple, GovTech Singapore and Modern Intelligence. Backed by Y Combinator. ### How it works 1. Connect. Agents with their own computers. They read the screen, reason through each case and work in your existing tools, with no integration. 2. Evaluate. Builds a custom evaluation for your workflow from your past cases. 3. Run. A fleet works the live queue, one agent per case. 4. Review. The right person makes each decision your procedure requires. ### Agent Fleets One agent scopes the work and hands out the cases. The fleet grows with the queue. - One agent scopes, assigns, and tracks the work. - One case per agent. - Each on its own computer. - Every step recorded. Every case replayable. ### Custom benchmark for your real work. Public benchmarks don't measure your work. So we build every customer a private one from their own past cases and their reviewers' decisions. Every model update, agent change and procedure change reruns your evaluation before it goes live, including edge cases and actions the agent must refuse. ### Starting with financial services. Operations, compliance and risk work that runs on procedure, evidence and review. Agents add capacity to the teams already doing it. - Compliance. Alerts and screening hits investigated, with the rationale on file. - Reconciliation. Breaks traced across portals and spreadsheets, each one explained. - Risk. Periodic reviews that follow your policy step by step. ### Common questions Q: How do we know it gets the work right? A: You check before you trust it. We run it on work your team has already completed and compare its output with your team’s decisions, item by item. Then it runs on new work in the background until you are comfortable. Even live, it only recommends; your team makes every determination, and every step is recorded so the reviewer can see what was checked, what was found and why. Q: Does it work with the tools we already use? A: Yes. It works on screen, the way a person does, so your portals, case manager, registries and internal systems stay exactly as they are. There is no API integration to build and nothing to migrate. It signs in with its own credentials, kept in a vault, and every two-factor prompt is cleared by a person. Q: What can it read and work in? A: Anything your team works in on screen: bank and custodian portals, company registries, case managers, ticketing tools, PDFs, statements and spreadsheets, and your own internal systems, under credentials you control. For compliance teams that includes sanctions and PEP lists such as OFAC SDN, UK HMT, the EU and UN lists, registries such as Companies House and OpenCorporates, adverse media, court records and WHOIS. A source it cannot reach is marked unavailable in the case file rather than skipped. Q: Where does a person stay in the loop? A: Wherever your procedure requires one. Cases reach a reviewer with the reason they were routed: low confidence, a policy threshold or a possible match. A named reviewer approves each decision, and routine cases can be signed off in batches. Q: What can our security team review? A: Your security team can review the security page, the list of third-party services and the written data flow description. The SOC 2 Type 1 audit report is available, and SOC 2 Type 2 is pending. We also walk through the controls with your team and answer your security questionnaires. ### Closing Bring your team to the frontier. Book a call: https://zommalabs.com/call --- ## Vocabulary Computer-use agent An AI agent that operates software the way a person does: it looks at the screen, moves the cursor, clicks and types. It works in the portals, registries, PDFs, case managers and internal tools a team already uses, so there is nothing to integrate. Case file What one agent hands back after one piece of work: every finding, a timestamped screenshot of every screen the agent read, the procedure step it applied, and a recommendation for the person who decides. One case file per subject. A source the agent could not reach is marked unavailable in the file rather than skipped. Screening The part of the product that builds the case file. Give it a person or a company and it checks the sanctions and PEP lists, the company register, the officers and owners, the website and adverse media, and returns one case file with a screenshot behind every finding. Your cases The part of the product that applies your written procedure to a case and recommends the outcome, with the reasoning under it. Teams start it on cases they have already closed, compare its recommendations with their own decisions, then let it run on new cases. Recommendation The last line of every case file: clear, escalate, or request information, with the evidence and the reasoning under it. The agent recommends. A person on your team makes the decision, and the decision is recorded. Replay The step-by-step record of a run: what was opened, what was searched, when each screenshot was taken and what went into the case file. Your auditor can open it later and see what the agent saw, what it did, and why. KYB and KYC Know Your Business and Know Your Customer checks on a new or existing customer: who they are, who owns and runs them, and whether they appear on a sanctions or PEP list or in adverse media. Zomma builds one case file per subject. Transaction monitoring alert A payment flagged by your monitoring system for a person to look at. Zomma checks the counterparty and the beneficiary, applies the context of the payment corridor, and recommends a disposition with the evidence under it. Periodic review A refresh of the file on an existing customer, on the schedule your procedure sets: re-screening against sanctions, PEP and adverse media and updating what has changed, so a growing book does not mean a growing backlog. --- ## Enterprise deployment (https://zommalabs.com/enterprise) Computer-use agents, deployed with your team. We sit with your team, write your procedure into the agent and prove it on your past cases before it goes live. It runs on your screens, under your control. Starting with financial services. ### What it runs today The first workflows we deploy, in the order most teams start. - Screening: KYB and KYC. Registry, officers, UBOs, website operator versus stated. One case file per business, child case files per UBO. - Reviews: Periodic reviews. Re-screens existing customers against sanctions, PEP and adverse media. Carried forward within 90 days, never silently re-run. - Alerts: Transaction monitoring. Counterparty and beneficiary checks on a transaction alert, corridor context applied, disposition recommended with the evidence behind it. ### Agentic Trust Framework The trust layer for agents doing real work. An agent doing real work in a regulated firm needs an identity, scoped permissions, a delegation chain, human approval before anything submits, and an audit trail you can hand an examiner. You can put it into your systems because every action is provable. And the firm keeps the knowledge: every correction becomes permanent firm memory. The delegation chain: you, agent identity, scoped task, audit log. ### Security & control Built for financial services from day one. - SOC 2 Type 1 audited. Type 2 pending. Independent audit, part of our security program. Reports shared with qualifying firms once issued. - Zero data retention. Nothing the agent touches trains a model, ours or anyone's. You keep full ownership. - SEC-ready audit trail. Every action and approval in an append-only log built for SEC Rule 204-2, kept seven years. ### Deployment Live in days, not quarters. 1. Map the procedure. We sit with your analysts and write your SOP into the procedure, exceptions included, across every screen it touches. 2. Run it while you watch. It runs your closed cases until its recommendations match your team's decisions, then works new cases quietly in the background. Your analysts correct it, and every correction becomes firm memory. 3. Let it recommend. It works new cases and recommends a disposition, with human approval before anything final. A dedicated engineer stays with you. 4. Add the next case type. Each case type you add is faster than the last, because the firm memory compounds. --- ## How computer use works (https://zommalabs.com/computer-use) An agent that uses software the way your team does. It watches the screen, moves the cursor, clicks, types, and switches between apps, just as your team does today. No API to build, no script to maintain. ### What computer use is Computer use is an AI model's ability to operate a computer like a person: observing the screen, controlling the cursor, clicking, typing, and switching between applications. It is the difference between an assistant that can talk about the work and one that can actually sit down and do it. Traditional automation needs an API integration an engineer has to build, or an RPA script that breaks when the screen changes. A computer-use agent perceives the interface visually instead. You give it a goal and it operates the software your team already uses: sanctions lists, registries, case managers, legacy desktop apps, web portals, and internal tools. ### How it works Perceive, reason, act, verify, on repeat. 1. Perceive. It reads the screen as a person would: pixels, layout, text, and window state, with no API contracts or brittle selectors. 2. Reason. A frontier model decides the next action from the goal and the current screen: which field to fill, which button to click, when to reassess. 3. Act. It moves the cursor, clicks, types, scrolls, and switches apps with ordinary human inputs. The underlying software is never modified. 4. Verify. It re-reads the screen after each step to confirm the action worked, catches errors and pop-ups, and corrects course before moving on. ### How Zomma compares Built for workflows, not just code. Each capability is rated for Zomma, coding agents and browser agents: full, partial, or not designed for it. - Operates real apps like a person, no API. Zomma: full. Coding agents: not designed for it. Browser agents: partial. - Learns a workflow by watching once. Zomma: full. Coding agents: not designed for it. Browser agents: not designed for it. - Improves from human corrections (firm memory). Zomma: full. Coding agents: partial. Browser agents: not designed for it. - Re-plans and recovers when steps break. Zomma: full. Coding agents: partial. Browser agents: partial. - Proven on your own past cases before it goes live. Zomma: full. Coding agents: not designed for it. Browser agents: not designed for it. ### No integration project. No brittle scripts. Because the agent works at the screen level, it does not need long integration timelines and it survives UI redesigns. It adapts to the exceptions that break scripts: a missing field, an unexpected pop-up, a mismatched record. It handles them or escalates to a person. ### What it runs today The first workflows we deploy, in the order most teams start. - Screening: KYB and KYC. Registry, officers, UBOs, website operator versus stated. One case file per business, child case files per UBO. - Reviews: Periodic reviews. Re-screens existing customers against sanctions, PEP and adverse media. Carried forward within 90 days, never silently re-run. - Alerts: Transaction monitoring. Counterparty and beneficiary checks on a transaction alert, corridor context applied, disposition recommended with the evidence behind it. --- ## Security (https://zommalabs.com/security) Security built in from day one. Each customer’s data is kept separate. Agents work on a cloud desktop that is wiped after every session, and every step is recorded. ### How it works The left hand shouldn't know what the right hand is doing. 1. You ask Zomma in plain language. Talk to Zomma like a teammate. Zomma works out what you want, but doesn’t act yet. 2. An orchestrator decides how to respond and act. Your policies, settings, and credentials live in a separate control layer. It has the final say on what the agent can and can't do. 3. The runtime does the work. Nothing else. Runs in isolation. No access to your policies, settings or credentials. Only the instructions it's been cleared to carry out. 4. Anything risky waits for you. Moving money? Sending a batch email to current clients? It stops and waits for your explicit approval before taking action. Orchestrator: Plans the steps, and keeps a record. Holds your policies, credentials, and context. Logs down its decisions. Runtime: Does the actual work. Works in its own locked-down desktop, one task at a time. It can suggest, but never decide. ### Security & trust at a glance The short version, for your security folks. - Hosting & residency: Amazon Web Services (AWS), region us-east-2 (Ohio). Case files, screenshots and logs are stored in the United States, inside Zomma’s virtual private cloud (VPC). - Encryption: Encrypted in transit (TLS 1.2+) and at rest (AES-256). - Model training: Your data is not used to train Zomma’s or any provider’s foundation models. - Credentials: Application logins are entered in a secure prompt, stored encrypted in a dedicated vault, never shown in chat, and never sent to an AI model. 1Password integration is currently supported. - Agent actions: Actions run through policy checks, with human authorization for sensitive steps and full audit logs. Read- and draft-only scope available. - Access control: Per-organization tenant isolation, SSO/SAML via WorkOS, role-based access, immutable audit trail. - Compliance: GDPR/CCPA-aligned data handling. SOC 2 Type 1 audited. Type 2 pending. ### Where your data goes Your organization is the boundary. - The agent: Works inside your workspace, in a cloud desktop that is wiped after every session. Nothing is shared across customers. Each customer runs single-tenant. - Screenshots and case files: Stored per organization, in Zomma’s AWS environment in us-east-2, with a retention period you set. - Credentials: Stay in the isolated vault. They are never shown in chat and never sent to a model. 1Password is supported. - Model calls: Go to Anthropic and OpenAI under zero data retention and no training on your data, agreed in writing. - Before you sign: A written data-flow description is available before any agreement is signed. Ask for it on the walkthrough. ### Go deeper Our rigorous approach to agentic trust and security is our competitive advantage. Blog: Agentic Trust Framework. We present the Agentic Trust Framework (ATF): how we screen what comes in, keep the model's reasoning in check, approve every action before it happens, and run each task in isolation. https://zommalabs.com/blog/agentic-trust-framework ### Your IT department asked Straight answers to the common questions. Q: Is data access read-only, or can it change things? A: Zomma can both read and make changes, but it cannot change anything on its own. Every change is checked against your rules, anything that leaves your systems needs a person to approve it, and all of it is logged. If you prefer, we can set it up read-only or draft-only, so it never makes a change at all. Q: Where is the data stored? A: Case files, screenshots and logs are stored within Zomma’s private AWS environment in us-east-2 (US East, Ohio), with encrypted storage at rest and in transit. Sensitive credentials are kept in an isolated vault, and the cloud desktops are wiped clean after every session. Q: Does the agent decide anything on its own? A: No. It recommends clear, escalate, or request more information, with the evidence under the recommendation. A person makes the decision, and a person clears every two-factor prompt. Both are logged. Q: Which third parties see our data? A: The third-party services that touch your data are listed on our subprocessors page. They include AWS for hosting, Anthropic and OpenAI as model providers under zero data retention and no training, 1Password for the credential vault, and Vercel, Google and Apollo.io for the website only. A written data-flow description is available before anything is signed. Q: What can our security team review before we sign? A: A data processing agreement if required, the list of third-party services, the SOC 2 Type 1 audit report (Type 2 pending), a technical walkthrough of how it works, and answers to your security questionnaires. ### For your security team Bring your reviewer. We'll answer everything. A 30-minute walkthrough of our controls, architecture, and audit trail, with your security lead in the room. - Data Processing Agreement (DPA), if required - SOC 2 Type 1 audited. Type 2 pending. - A technical walkthrough of how it works - Answers to your security questionnaires The third-party services that process customer data are listed at https://zommalabs.com/subprocessors --- ## Agent benchmarks (https://zommalabs.com/agent-benchmarks) Test AI on real work benchmarks. Building AI agents for operations? You need custom benchmarks that test them for accuracy, reliability, and safety. ### What to check before an agent does real work. Whether you're building your own agents or evaluating a vendor's. - Test on your own tasks, not public benchmarks. - Test the whole setup: model, instructions, tools, and permissions. - Check what it does when information is missing or an action is not allowed. - Retest whenever the model, tools, or procedures change. ### Recreate the task as a test. We recreate tasks like payment reconciliation and compliance reviews in a test environment and define what a correct result looks like. Example benchmark: Reconcile payments. Starts with: Payment records and a ledger. A passing run: Matches the right records. Flags discrepancies. Saves the result. ### Compare models and harnesses. Run the same tests on the complete agent setup: the model, instructions, tools, permissions, memory, and execution environment. - Correct results: Does it complete the task and save the right result? - Missing information: Does it ask for review when it cannot finish? - Working controls: Do permissions and approvals block unauthorized actions? Does stopping the agent stop its actions? ### Train the agent on your workflow. - Computer-use data: Each run produces task inputs, screenshots, and recorded actions. - Custom RL environments: Successful and failed runs both become training tasks. ### The best teams test agents the way they test software. They write tests from their own work, rerun them on every change, and train on both the successes and the failures. --- ## Company (https://zommalabs.com/company) Work worth doing. Zomma is an applied AI lab that builds computer-use agents for the enterprise. Each agent has its own computer, reads the screen and works in the tools a team already uses, with no integration. Every agent is proven on a private benchmark built from the customer's own past cases before it goes live, and again after every change. Zomma starts where mistakes are expensive: compliance, reconciliation and risk work in financial services. Founding team: Varun, Cash and Jihyun. Backed by Y Combinator. Based in San Francisco. A letter from Jihyun Kim, co-founder and CEO: For the past few hundred years, every leap in technology promised us the same: take the heavy, repeating work off our hands and give back our time, our judgment, our ambition. And it mostly did; we built more and reached further than our grandparents could picture. But somewhere along the way, work became a chore. The most capable people spend their days on work beneath their judgment, being the glue between legacy systems. I spent years at Goldman Sachs, then at a wealth tech startup, then launching a hedge fund. And everywhere, I saw the best operators, who were often the best systems thinkers, overwhelmed with mechanical work with less left for the work only they could do. With today's leap, we see the same promise but also fear what it will take away. Automation as lost jobs. AI coming for people's work. That story is corrosive because it treats people as a cost to cut, when they are the source of judgment, creativity and relationship no system yet has. Fear is a poor foundation for the most important tool of our time because it gives away human agency that will always be ours. The goal was never fewer people; it is people freed for more of what is human, new, and worth doing. This year, for the first time, computer use became good enough to make that real. AI agents can finally do the work people do, learn how systems run on screens, and hand it back multiplied. We are building the most ambitious version we can imagine. Not something that replaces people, but something that raises the bar of what every single team can become. Work should make us more ourselves, not less. And we are building it, instead of fearing it. If that's the future you believe in, come build it with us. Jihyun Kim, Co-founder & CEO, Zomma --- ## Writing (https://zommalabs.com/blog) Older posts were written for an earlier fund operations and advisory frame and should be read as dated. The current frame is the one above. - How to test AI agents for finance operations before production (2026-10-04) https://zommalabs.com/blog/test-ai-agents-finance-operations For operations leaders at fintechs and financial institutions: how to test AI agents on your own workflows before production, and what to own in a buy vs build decision. - FINRA guidance on AI agents, explained for operations teams (2026-09-29) https://zommalabs.com/blog/finra-ai-agents-guidance-operations What FINRA has said about AI agents through September 2026, which rules apply, and what operations teams at broker-dealers should do before deploying an agent. - AI Agents for Transaction Monitoring in Cross-Border Payments: A 2026 Guide (2026-08-17) https://zommalabs.com/blog/ai-agents-transaction-monitoring-cross-border-payments Why transaction monitoring queues at cross-border payment companies grow faster than the teams that work them, what an alert investigation actually involves, why regulators forbid the obvious shortcuts, and how AI agents add review capacity inside the systems you already run. - The trust gap: why 95% of fund managers use AI and only 7% run agents (2026-07-20) https://zommalabs.com/blog/trust-gap-ai-agents-fund-operations 95% of fund managers use AI. 7% run agents. A former ops operator on what the gap is made of, what closes it, and what agents still cannot do. - Why Zero Trust is the Baseline for Implementing Agentic Systems (2026-07-18, updated 2026-07-18) https://zommalabs.com/blog/zero-trust-agentic-systems My field notes on why Zero Trust is important to building trust in agentic systems, and what we could build on top of it. - Zomma-ing towards SPIFFE: a Secure Production Identity Framework for Everyone (and their Agents) (2026-07-04, updated 2026-07-04) https://zommalabs.com/blog/zomma-spiffe-implementation Why Zomma assigns SPIFFE IDs to our ephemeral runtime instances the orchestrator spins up — short-lived X.509 certificates from our own trust domain make every agentic action attributable and accountable. - Can an AI agent replace data entry? (2026-07-02) https://zommalabs.com/blog/ai-agent-replace-data-entry An agent handles the rule-based part of data entry, reading from one system and typing into another, faster and without the errors that creep in by hour three. The judgment calls still need a person. - How to Automate Fund Admin and Transfer Agent Portal Work (2026-07-01) https://zommalabs.com/blog/automate-fund-admin-transfer-agent-portals Fund admin portals, prime broker sites, and transfer agents rarely have APIs. How hedge fund and family office ops teams automate the daily downloads, re-keying, and month-end reconciliation with computer-use agents. - Implementing the Agentic Trust Framework (ATF): Zomma's Playbook (2026-06-24, updated 2026-06-24) https://zommalabs.com/blog/agentic-trust-framework Zomma's playbook for implementing the Agentic Trust Framework: four concern bands, deterministic controls, human approval, and an examiner-ready audit trail. - Headless vs Headful: Why Computer Use Survives an Agent-Run Internet (2026-06-17, updated 2026-06-17) https://zommalabs.com/blog/headless-vs-headful-computer-use Headless vs headful, explained, and why an agent-dominated future still needs computer use. Most software has no API and never will, so an agent that can operate a screen can work where API-only agents can't. - Enterprise Computer Use: What It Is, How It Works, and How to Deploy It Safely (2026-06-10, updated 2026-07-02) https://zommalabs.com/blog/enterprise-computer-use Enterprise computer use is AI agents operating software the way a person does, for regulated work that lives behind portals with no API. How it works, how it differs from RPA and APIs, and why in regulated finance the screen is the control surface, not a fallback. - Computer-Use Agents vs RPA: Why a Decade of Enterprise Automation Kept Breaking (2026-05-27, updated 2026-07-02) https://zommalabs.com/blog/computer-use-agents-vs-rpa Why RPA bots keep breaking, how computer-use agents handle a screen that changes, where each one actually fits, and why financial back offices feel the shift first. - AI Agents for KYC & Customer-Profile Review: A 2026 Guide for Compliance Teams (2026-05-13, updated 2026-07-02) https://zommalabs.com/blog/ai-agents-for-kyc-customer-profile-review What a KYC customer-profile review requires, how often you actually have to do one (it's risk-based, not fixed-cadence), the rules by business type in 2026, and how AI agents clear the review backlog by working inside the compliance systems you already use. - Why your best operations person is your biggest risk (2026-04-29) https://zommalabs.com/blog/ops-knowledge-loss-financial-services Small operations teams run on unwritten judgment. When the person who knows the exceptions leaves, the firm loses more than capacity. - How to reduce manual data entry in operations (2026-04-15) https://zommalabs.com/blog/reduce-manual-data-entry-financial-services Operations teams move data between portals, spreadsheets, CRMs, and internal systems all day. Here is how to cut that work without a large integration project. - What does it cost to automate an operations workflow? (2026-04-08) https://zommalabs.com/blog/cost-automate-back-office RPA, offshore teams, internal engineering, and computer-use agents all have different economics. The right answer depends on workflow volume, risk, and maintenance burden. - Automation tools for the legacy systems teams still use (2026-04-01) https://zommalabs.com/blog/desktop-automation-legacy-software Many operations workflows still depend on old portals, desktop apps, and brittle exports. Here are the automation approaches that still work. - How long should an operations automation project take? (2026-03-25) https://zommalabs.com/blog/how-long-rpa-implementation Traditional RPA can take months. Computer-use agents can prove value in weeks because they start where the work already happens. - How to automate computer work without an API (2026-03-11) https://zommalabs.com/blog/automate-desktop-without-api Most automation tools need an API. Many portals and internal systems do not have one. Screen-based agents solve a different problem. - RPA vs AI agent: what operations teams should know (2026-03-04) https://zommalabs.com/blog/rpa-vs-ai-agent RPA follows scripts. Computer-use agents read screens and adapt. Here is what that difference means for real workflows.