V7 uses Context Graph to give enterprise agents lasting memory
V7 says its V7 Go platform gives AI agents a structured, source-linked memory of a company’s documents and data. In tests of its most difficult Context Graph queries, the startup reported 89% accuracy for GPT-6 Astra, compared with 78% for GPT-5.6 Sol on the very-hard tier. The benchmark used messier real-world information spread across thousands of documents.
The Europe and UK-based company, founded in 2018 by Alberto Rizzoli and Simon Edwardsson, is positioning V7 Go for mission-critical workflows in finance, insurance and real estate. Its premise is that models can reason over tasks but do not inherently know which fund report is current, how an entity is named in separate systems, or which evidence supports a specific decision.
From files to a queryable company record
V7 Go connects to repositories including SharePoint and Google Drive. It scans incoming files for entities, relationships, facts, attributes and metrics, then places those elements in the Context Graph. Each fact can be connected to a new or existing record and preserved with cited evidence pointing back to the original material.
GPT-5.6 Luna performs extraction across millions of files, while GPT-5.6 Terra and Sol are used for reasoning and tool use in multi-step workflows. The Context Graph is also available through an MCP server, enabling customers to query it from ChatGPT and other compatible clients, and to create V7 Go workflows through MCP in Codex.
V7 says the graph is cheaper and faster to traverse than long-context approaches, while retaining the ability to use retrieval-augmented generation when the graph lacks sufficient detail. On HERB, its benchmark for connecting information distributed across enterprise systems, V7 reported that its retrieval-only system outperformed the official baseline by 69% and reduced hallucinations on unanswerable questions by 38%.
Workflow results and model testing
The company says its agents complete workflows of 50 to 100 steps in minutes with 99.9% accuracy and an auditable decision trail. Reported customer outcomes include private-equity deal screening performed 21 times faster, a financial-services review reduced from more than 100 hours to under 10, and a 13.5% reduction in insurance claims-processing errors against a manual baseline.
V7 tests models on citation accuracy, extraction across hundreds of document types, answer correctness, instruction following, latency, cost and enterprise workflows. It reported that GPT-5.6 Sol reduced tool-call errors from 2.7% with GPT-5.5 to 0.2%. GPT-5.6 Luna also delivered a 78% lower cost per document than GPT-5.4 mini, while a move from the Chat Completions API to the Responses API reduced token use by roughly 5% for some PDF-heavy workflows and improved caching reliability.
What organisations should take from the approach
V7 is working towards workflows that react when facts in the Context Graph change, such as flagging analyses that rely on figures from a restated fund report. For organisations deploying agents in high-stakes document processes, the practical implication is to establish governed, evidence-linked context and auditable workflow traces before asking models to make decisions across long chains of work.