Sphere, a global indirect tax compliance startup, has published a detailed technical account of TRAM (Tax Review and Assessment Model), its AI-native system for automating tax research and determination across more than 100 jurisdictions. The announcement coincides with a $21 million Series A led by Andreessen Horowitz. TRAM is designed to replace the "content team" model that incumbents like Avalara, Vertex, and Thomson Reuters ONESOURCE have relied on for decades — armies of human tax researchers who manually ingest legislation, codify rules, and push periodic updates into proprietary databases. Vertex's content database alone claims 900 million effective tax rules built over 40 years, while Avalara employed more than 100 tax research professionals to maintain its 900,000-plus rules across 12,000 U.S. jurisdictions. Those figures come from Sphere's own technical writeup; Agent Wars has not independently verified them against Vertex's or Avalara's primary sources.

A proprietary web automation tool called WARP continuously crawls official government and tax authority websites across all supported jurisdictions, ingesting statutes, regulations, administrative bulletins, case law, and private letter rulings. Raw documents — web pages, PDFs, and spreadsheets — are normalized, split into <a href="/news/2026-03-14-captain-yc-w26-launches-automated-rag-platform-for-enterprise-ai-agents">semantically meaningful sections</a>, and enriched with metadata: jurisdiction, authority type, effective dates, and document hierarchy. Each section is indexed in both dense and sparse vector form. Dense vectors handle semantic interpretation of guidance; sparse vectors handle keyword-precise retrieval of exact numerical rules. At inference time, OpenAI reasoning models fine-tuned via Reinforcement Fine-Tuning on expert-labeled corrections perform multi-step reasoning over <a href="/news/2026-03-14-rag-document-poisoning-attack">retrieved materials</a> to generate product taxonomies, taxability determinations, and applicable rates.

All TRAM outputs flow into a review portal where Sphere's internal tax professionals approve or modify determinations before they reach the live production engine. Any correction comes with an explanation that feeds back into TRAM as a training signal — each flagged error sharpening the model's future outputs. The target is the combinatorial explosion that Sphere identifies as the structural bottleneck for incumbent vendors: product types multiplied by jurisdictions multiplied by time, a problem that has historically required headcount to solve. Those incumbents have not stood still. Vertex launched Indirect Tax Intelligence features, and Thomson Reuters repositioned ONESOURCE as an "Intelligent Compliance Network" following its acquisitions of Pagero Group and SafeSend. But both moves layer AI onto existing content-team infrastructure rather than replacing it. Sphere's thesis, if TRAM holds up at production scale, is that the research pipeline can run on compute instead of headcount — turning a labor-scaling problem into a cost the company can manage in a data center.