## Company Overview and Use Case Tabs is a vertical AI company founded in December 2022 by CEO Ali and CTO Deepo, who previously worked together at Latch, a smart access company in New York City. The company focuses specifically on revenue intelligence for B2B sellers, addressing what they identify as the last major area of innovation in general ledger management after payroll and expense management have been largely solved. The core problem Tabs addresses is that most B2B companies operate with very lean revenue teams - typically a $50 million revenue B2B company might have just one VP of finance and perhaps an offshore billing specialist. Tabs aims to help these companies "get the most money into your bank account as quickly as possible" while enabling accurate reporting against revenue operations. ## Technical Architecture and Data Model The foundation of Tabs' LLMOps implementation is what they call the "commercial graph" - a knowledge graph that captures all possible information about merchant-customer relationships. This concept parallels other vertical graph approaches like Rippling's "employee graph" or Ramp's "vendor graph," but focuses specifically on commercial relationships in B2B sales contexts. The platform begins by taking sales contracts from a sales-led motion and extracting all key information to run revenue operations. While document information extraction was initially a core differentiator, the founders recognized early that this would become commoditized and focused instead on building intelligent workflows on top of the extracted data. The commercial graph is hydrated with diverse data sources including: - Contract terms and modifications - Customer communication history and sentiment analysis - Usage data and behavioral patterns - Payment history and collection patterns - Customer support interactions ## LLMOps Evolution: From Guided to Ambient AI Tabs' LLMOps journey illustrates a clear evolution in AI implementation strategies. The company started with traditional software engineering for the first year and a half, then layered on what they call "guided AI experiences" - essentially co-pilot functionality where the AI provides insights but operators remain responsible for completing tasks. However, their current focus is on "leapfrogging" the guided experience phase to move directly into fully agentic experiences. This represents a more proactive approach where agents either execute tasks autonomously or request permission to act, rather than simply providing recommendations. The company has deliberately avoided implementing a chat interface for their agentic experiences, instead focusing entirely on ambient functionality. This design decision reflects their belief that the future lies in agents that operate in the background without requiring explicit human prompting for each action. ## Ambient Agent Implementation Tabs' ambient agents represent a sophisticated implementation of background AI workers that monitor communication channels and trigger actions based on events rather than explicit commands. For example, their agents monitor email communications between the company and its customers, automatically processing relevant information and initiating appropriate responses or workflows. The agents are designed to integrate seamlessly with existing communication modes, particularly Slack and email. When an agent needs human approval, it sends draft communications through these channels just as a human team member would. For instance, when responding to customer emails, an agent might send a Slack message saying "we received this email, this is what we'd like to send out, does it look okay?" - mimicking natural human collaboration patterns. This approach addresses what the team identifies as a fundamental shift from background agents (which require human initiation) to truly ambient agents (which are triggered by events). The key distinction is that ambient agents listen continuously to communication streams and organizational processes, acting autonomously when appropriate triggers occur. ## Technical Challenges and LLMOps Considerations The primary technical challenge Tabs faces is information retrieval and relevance determination within their knowledge graph. As Deepo explains, "you can hydrate your knowledge graph as much as you want but it's really about knowing what is the most important stuff to pull." This challenge becomes particularly acute in enterprise B2B contexts where vast amounts of unstructured data exist across multiple dimensions. The company struggles with determining what information is relevant to specific tasks, particularly when dealing with customer behaviors, sentiment analysis, usage patterns, contract modifications, and communication history. They've identified that every vertical AI company will eventually face this information retrieval problem as their knowledge graphs become more comprehensive. Memory and learning represent another significant challenge. The agents need to remember previous interactions and decisions to avoid repeatedly asking for the same approvals or making the same mistakes. However, rather than implementing traditional memory systems, Tabs is exploring approaches that feed learned information back into the commercial graph, making it available for future retrieval. ## User Experience and Interface Design Tabs has made deliberate design choices about how humans interact with their ambient agents. They've rejected traditional UI patterns in favor of integration with existing communication tools. The audit trail and task management happen through what they're building as an "agent inbox" - essentially bringing developer workflow concepts like Kanban boards and ticket systems into the finance domain. This reflects their broader philosophy of bringing "developer processes to other areas of an organization," including concepts like running sprints and using prioritization systems for financial operations. For example, they might prioritize between a $10,000 invoice that might get paid versus a $100 invoice that will definitely get paid with outreach. However, their vision extends beyond prioritization to true parallelization - having "endless number of parallel threads" that can handle all tasks simultaneously, with gates only at points of ambiguity or dependency (such as month-end close processes where certain tasks must complete before others can begin). ## Industry-Specific Considerations The finance industry presents unique challenges for LLMOps implementation. As Deepo notes, "when the numbers are wrong they [accountants and finance people] are left to blame," which creates high demands for explainability and audit trails. This has influenced their interface design to include explicit approval mechanisms where users can "click a button and they can say oh like yes or no like explicitly." The company believes that memory and learned behaviors will become a significant moat for enterprise applications, as "every enterprise that we work with has a different way in which they do business and being able to do it on an enterprise by enterprise basis is important." This suggests that successful LLMOps in finance requires not just technical sophistication but deep customization to organizational patterns. ## Future Vision and Technical Roadmap Tabs envisions an end state of "beautiful operational software that no one ever has to go into" - essentially creating a fully ambient system where agents handle tactical operations while humans focus on strategic work. Their target organization structure includes a lean internal finance team (CFO, VP of finance) supported by multiple agents running ambiently in the background. The technical roadmap involves two main phases: first, hydrating their commercial graph with comprehensive data even if some proves irrelevant, and second, solving the information retrieval problem to ensure agents get the right context for their tasks. They're also exploring how to implement effective memory systems by feeding learned behaviors back into the graph structure. This case study illustrates the evolution of LLMOps from simple document processing to sophisticated ambient agent systems that integrate deeply with organizational workflows and communication patterns. While still early in implementation, Tabs' approach demonstrates how vertical AI companies can build defensible moats through domain-specific knowledge graphs and learned organizational behaviors.