## Overview Orbital Witness developed Orbital Copilot, described as the first domain-specific AI agent for real estate legal work, to revolutionize how lawyers conduct due diligence and prepare lease reports. The case study chronicles the evolution from classical machine learning approaches through the LLM era to the current agentic paradigm, positioning the product as more than "just a wrapper around GPT-4" or "ChatGPT for lawyers." The company partnered with Bryan Cave Leighton Paisner (BCLP) and other prestigious law firms to develop and deploy this system, claiming time savings of up to 70% on lease reporting tasks. The genesis of the project traces back to mid-2022 when Orbital's data scientists began exploratory work with generative AI using Google's BERT and T5 models. The landscape shifted dramatically with ChatGPT's release in November 2022 and GPT-4's subsequent launch, which accelerated their R&D efforts. By June 2023, they presented "Generative AI: Opportunities and Risks for Property Transactions" to law firms, which generated immediate demand for access to their tools. The company rapidly progressed from closed alpha testing with select early adopters to launching paying customers by the end of 2023, with broader availability planned for 2024. ## Architectural Evolution and AI Agent Design The case study provides valuable context on the evolution of AI approaches in legal tech. Orbital initially built "some of the industry's most accurate ML models for classifying real estate legal text" using classical supervised learning with extensive labeled datasets. The advent of LLMs in 2023 reduced dependency on large labeled datasets, enabling systems built with LLM APIs (GPT-3.5, Claude, Gemini Pro) that process legal text as "context" with specific "prompts." They note that RAG techniques enhance these systems' ability to manage extensive context from multiple lengthy documents, though they characterize such approaches as suitable only for "simplistic tasks" with "several limitations" for typical legal due diligence work. The company positions AI agents as the next evolutionary step, drawing on Lilian Weng's definition of "LLM Powered Autonomous Agents" with three critical components: Planning, Memory, and Tool Use. In their architecture, the LLM functions as the "brain" while specialized tools handle intricate real estate legal tasks like querying land registries or determining how provisions might be varied by other documents. The system can reason about its outputs and decide whether to continue searching documents, request additional materials, or present a complete answer. This agentic approach allows Orbital Copilot to handle complex workflows that mimic lawyer reasoning. For example, when asked "What is the rent for this property?", the system can identify relevant details across multiple documents (leases, deeds of variation), follow definitional trails within documents, and logically deduce answers. The transparency component is particularly important for legal applications—the system reveals its thought process and reasoning, allowing lawyers to understand how conclusions were reached rather than treating it as a "black box." ## Production Capabilities and Features Orbital Copilot's production capabilities demonstrate sophisticated document understanding and analysis. The system can digest hundreds of pages of legal text across numerous PDF documents, handling both typed text and handwritten manuscripts through OCR capabilities. It resolves diverse questions ranging from straightforward queries like "What is the date of the lease?" to complex inquiries such as "How has the service charge varied?" The system maintains contextual understanding by tracking definitions within documents, follows information trails across one or several documents, and performs supplementary research by accessing proprietary legal knowledge bases or data from HM Land Registry. The system provides targeted summarization capabilities for entire documents or specific sections across multiple documents, and can rephrase complex legal jargon into layman's terms for client comprehension. Critically for legal applications, it provides trusted referencing by indicating specific parts of PDF documents it consulted, facilitating direct navigation to primary evidence. This citation capability allows lawyers to immediately verify source material rather than blindly trusting AI outputs. The case study provides concrete examples of the system in action. In one scenario, two documents are uploaded (a lease dated 06-06-2008 and a deed of variation dated 31-03-2016), and when asked "What is the rent and how has it been varied?", Orbital Copilot understands the question context, searches and reads pertinent sections in both documents, analyzes findings, and formulates a response. Another example shows processing of a short-form lease report with 10 targeted questions, demonstrating that the system includes pre-configured report templates ranging from basic to highly detailed inquiries. ## LLMOps Challenges in Production The case study candidly discusses several engineering challenges that provide insight into real-world LLMOps considerations for legal applications. Cost management emerges as a critical concern—utilizing state-of-the-art LLMs like GPT-4 for thorough analysis of hundreds of pages of legal documents is "crucial for achieving the accuracy our lawyer clients depend on" but "incurs significant costs." The company acknowledges needing to optimize LLM usage to balance cost-effectiveness with high-quality output, though they anticipate decreasing expenses as GPU production ramps up and LLM efficiency improves. Resource availability presents another significant challenge. The global shortage of Nvidia GPUs and soaring demand for LLM functionalities has forced providers to impose rate limits on tokens processed per minute through their APIs. This directly affects Orbital's capacity to onboard new customers and influences task execution speed within Orbital Copilot. The company frames this as a significant short-term constraint requiring careful management, with expectations that the issue will diminish as GPU availability increases. Reliability concerns also surface, with the observation that many LLM providers, "despite their technological prowess, are relatively new to managing complex, fault-tolerant services on a global scale." This inexperience manifests as occasional service fragility, uptime issues, and performance degradation that directly impact operations. The company notes this necessitates "continuous vigilance and adaptability to maintain uninterrupted service quality." These production challenges are particularly noteworthy because they represent real operational constraints for LLMOps at scale rather than theoretical concerns. The rate limiting issue especially highlights how external API dependencies can create bottlenecks for customer onboarding and user experience, while cost considerations become paramount when processing large documents repeatedly for multiple clients. ## Deployment and Customer Adoption Orbital Copilot progressed through a structured deployment pipeline from R&D concept to production service. The timeline shows exploratory work beginning mid-2022, internal tool development accelerating after ChatGPT's release, a presentation generating customer demand in June 2023, followed by a Closed Alpha with select early adopters providing feedback, then a private closed beta in Q4 2023, and finally the transition to paying customers by early 2024. The customer cohort includes prestigious law firms that represent significant validation for the technology: BCLP (31 offices worldwide, clients representing 35% of Fortune 500), Clifford Chance (one of the world's largest law firms), Charles Russell Speechlys (international firm with offices across UK, Europe, Asia, Middle East), Macfarlanes (London-based), Ropes and Gray (13 offices on three continents, named "Law Firm of the Year" 2022), Walker Morris, Thomson Snell and Passmore (oldest law firm tracing back to late 16th century), Thompson Smith and Puxon, and Able UK (land developer and port operator). Customer testimonials provide qualitative validation. Matt Taylor from Clifford Chance describes it as "next generation legal technology" helping focus lawyers on areas most valuable to clients. Samant Narula from BCLP emphasizes "marked benefits to clients by accelerating lease reporting and enhancing real estate due diligence" and notes many lawyers being "deeply engaged in the development of leading-edge technology." Amy Shuttleworth from Charles Russell Speechlys calls it "the kind of tool that every lawyer wishes they had," highlighting the ability to "review documents quickly and efficiently, whilst also allowing you to check and verify the information." The partnership with BCLP represents particular strategic significance—described as a "global design partnership," it extends lease reporting capabilities initially developed for the UK market to BCLP's real estate practices in both UK and US. This marks Orbital's expansion from UK-centric operations to transatlantic presence, with existing clients expressing eagerness for global deployment. ## Business Impact and Validation The claimed business impact centers on the 70% time savings figure for lease reporting tasks. The case study contextualizes this by noting that comprehensive lease reports for a single property can take 2-10+ hours depending on complexity, and lawyers often bill in six-minute increments. The time savings therefore translate to "substantial financial savings per property for law firms and their clients," with cumulative efficiency gains given the regularity of lease reports in real estate law. However, the case study should be evaluated critically as marketing material from the vendor. The 70% time savings claim, while impressive, lacks detailed methodology or independent validation. The document describes "thorough testing with top-tier UK law firms, involving real client work" but doesn't provide sample sizes, statistical rigor, or information about which specific tasks achieved what levels of savings. The phrase "up to 70%" suggests maximum rather than average performance, which is a common marketing technique. The customer testimonials, while positive, are brief and lack specific quantitative details about productivity improvements or cost savings. The rapid progression from concept (mid-2022) to paying customers (early 2024) in approximately 18 months is notable but also raises questions about long-term reliability and whether all edge cases have been adequately addressed in such a compressed timeline. ## Technical Approach and Differentiation The case study emphasizes differentiation from simpler LLM applications, arguing that systems which "summarise documents or answer specific one-off questions" represent "simplistic tasks" where "the technology to perform these types of tasks is becoming well understood." They position the "real long-term value" as "a dynamic AI assistant built on the AI Agent architectural pattern." This framing serves marketing purposes but also reflects genuine architectural differences between basic RAG systems and more sophisticated agentic approaches. The multi-document reasoning capability represents a key technical achievement—the system must not only extract information from individual documents but follow logical connections across documents (e.g., understanding how a deed of variation modifies a lease). The tool-use capability allowing access to external data sources like HM Land Registry demonstrates integration beyond pure document analysis. The transparent reasoning and citation features address the critical trust requirements for legal applications where lawyers must verify and stand behind the AI's outputs. The reliance on GPT-4 specifically (mentioned as essential for meeting "high standards required for legal document analysis") indicates they're using the most capable models rather than cost-optimized alternatives. This design choice aligns with the legal domain's low tolerance for errors but contributes to the cost challenges discussed. The system appears to use RAG as a foundational technique (retrieving relevant document sections as context) but extends it with the agentic components for planning, reasoning, and tool use. ## Future Trajectory and Market Position Orbital positions itself as building "the world's premier AI Agent tailored for real estate legal work" and claims Orbital Copilot is "the first product of its kind in the industry." The global expansion plans from UK to US markets with BCLP, along with interest from existing clients for worldwide deployment, suggest ambitions beyond a niche UK product. The reference to developing "groundbreaking features, slated for release in the first and second quarters of 2024" indicates ongoing product development, though specifics aren't provided. The case study contextualizes Orbital Copilot within broader industry trends, citing Sam Altman's DevDay announcement of customizable "GPTs" and the GPT Store, and Ethan Mollick's exploration of emerging agent capabilities. This positioning alongside OpenAI's strategic direction and prominent AI commentators serves to validate the agentic approach while also suggesting Orbital is riding a broader wave of AI agent adoption rather than pioneering entirely novel territory. The company describes itself as "product-centric" and "deeply invested in understanding and addressing our customers' needs," with a "customer-first approach" driving product development. This emphasis on user research and iterative development with law firm partners represents sound LLMOps practice, particularly for specialized domains like legal work where domain expertise is critical for building useful tools. ## Balanced Assessment As an LLMOps case study, this document provides valuable insights into deploying sophisticated LLM applications in a high-stakes professional domain. The candid discussion of cost management, rate limiting, and reliability challenges reflects real production constraints that many organizations face but don't always publicly acknowledge. The emphasis on transparency, citation, and reasoning capabilities addresses genuine requirements for legal applications where "black box" AI is insufficient. However, the source material is fundamentally marketing content ("Let's build the future together: Limited spots on our early adopters waitlist are available") and should be evaluated accordingly. The performance claims lack independent validation, the technical details remain high-level without implementation specifics, and the rapid timeline raises questions about long-term robustness. The positioning as "the first product of its kind" and claims of building "the world's premier AI Agent" represent marketing language rather than objectively verifiable statements. The customer adoption by prestigious law firms provides meaningful validation that the product delivers sufficient value for firms to pay for it and integrate it into client work. The partnership model with BCLP for global expansion suggests deeper engagement than typical vendor-customer relationships. The progression from closed alpha through beta to paying customers represents a structured deployment approach appropriate for high-stakes applications. From an LLMOps perspective, the case study illustrates important considerations: domain-specific applications require more than generic LLM wrappers; cost and rate limits are real operational constraints; reliability and trust mechanisms (citations, reasoning transparency) are essential for professional adoption; and iterative development with domain experts is critical for success. The agentic architecture represents a more sophisticated approach than basic RAG implementations, though the actual implementation details remain proprietary.