OpenAI Publishes V7 Context Graph Case Study

OpenAI published a case study on the startup V7 on September 21. It covers V7 Go, which turns files scattered across a company into context that agents can query, and it came with a number: GPT-6 Astra hit 89% accuracy on the hardest set of graph queries. GPT-5.6 Sol scored 78% on the very hard tier of the same test.

The announcement frames the problem this way. Today's models can reason through complex tasks, but they do not automatically understand the business context those tasks sit in. Which fund report is current, and how the same company is named across three different systems. That context is scattered and unresolved across documents, data rooms, spreadsheets, emails and internal tools, and it stays invisible to agents.

V7 was founded in 2018 by Alberto Rizzoli and Simon Edwardsson. The two had previously built a widely used computer vision accessibility app together, and they started V7 to help companies teach AI systems how their businesses actually work. Its main product now is V7 Go, an agentic platform for building mission critical workflows.

The architecture has two layers. GPT-5.6 Luna extracts information from millions of files and organizes it into the Context Graph, which connects entities, relationships and cited evidence to power MCP search and repeatable workflows. A single workflow can span hundreds of steps. Multi-step instructions that need more reasoning and tool use run on GPT-5.6 Terra and Sol, and the company has started putting GPT-6 Astra on its most demanding graph queries, such as financial analysis across thousands of documents.

With context, models and tools working together, the company said agents finish 50 to 100 step workflows in minutes, reach 99.9% accuracy and leave an auditable trail of every decision. Alberto Rizzoli, Co-Founder and CEO at V7, said that "to solve hard enterprise use cases across finance and insurance, AI needs to learn how your business operates just as well as it learned from the Internet."

From an engineering view the waste this design targets is clear. Without a graph, agents have to rediscover context on every request, burning time and tokens across dozens of searches while still missing key information buried in relationships. V7 Go connects to repositories such as SharePoint and Google Drive, scans them for entities, relationships, facts, attributes and metrics to populate the graph, and the announcement said that graph is an order of magnitude cheaper and faster to traverse than long-context approaches.

When a new file arrives, V7 Go identifies the companies, funds, people or any other entity in an ontology, connects each fact to a new or existing record, and preserves cited evidence back to the original source. If the graph does not hold enough information, it can still search the underlying documents with RAG. For long-running agents, recent exchanges stay in the model's active context while older material moves into the graph to be retrieved when needed.

V7 also measured how much that structure matters. On the HERB benchmark, which tests the ability to find and connect information spread across enterprise systems, the company said its retrieval-only system outperformed the official baseline by 69% and reduced hallucinations on unanswerable queries by 38%. METAL confirmed that HERB is a benchmark released by Salesforce AI Research in June 2025, posing multi-hop questions over 39,190 enterprise artifacts drawn from documents, meeting transcripts, Slack messages, GitHub and URLs. The paper reported that even the best-performing agentic RAG methods averaged only 32.96, and named retrieval as the bottleneck.

Customer numbers came with it. Asset managers screen deals 21 times faster than before, cutting a full-day process to 15 minutes, and one financial services team brought review time down from more than 100 hours to under 10, saving $12,000 in expert costs per task. An insurance team reduced errors in claims processing by 13.5% against a manual baseline after giving its agents historical knowledge of all previous claims and existing policies.

The demo in the announcement is private equity deal screening. A workflow pulls key financials, deal terms, management details and cited risk fields out of a deal document called a Confidential Information Memorandum, then produces a screening note. Insurance underwriting runs the same way, the company said.

Model selection is split into tiers. In AI-generated workflows, V7 Go maps each step to fast, medium and smart tiers. GPT-5.6 Luna handles structured extraction and high-volume work, while GPT-5.6 Terra or Sol take chat, the Go Agent path, and steps that need more reasoning or tool use.

Simon Edwardsson, Co-Founder and CTO at V7, explained the choice with performance figures. "We chose OpenAI as our default because it performs best on the multi-step tool workflows V7 Go depends on," he said, adding that in the company's own Context Graph benchmark the tool-call error rate fell from 2.7% with GPT-5.5 to 0.2% with GPT-5.6 Sol. He also said Terra let the team remove many intermediate workflow stages that existed only to simplify the task for the model, which alone saved days of delivery work.

The rest of the efficiency numbers sit in the plumbing. Key workflows with several external calls now finish up to 50% faster in V7's latest harness, and GPT-5.6 Luna cost 78% less per document than GPT-5.4 mini. V7 moved its document-heavy V7 Go workloads from the Chat Completions API to the Responses API, and in its own testing token use fell by roughly 5% for some PDF-heavy workflows while caching reliability improved. The announcement also noted that OpenAI approved and implemented a capacity increase within hours, where other providers took weeks.

The 89% figure exists because the benchmarks were saturated. GPT-5.6 Sol had filled most of the existing tests close to the ceiling, so V7 built a new set of hard graph queries from messier real-world data across thousands of documents. The dataset spans four difficulty levels, and both models scored close to 100% on easy, medium and hard, splitting only at very hard into 78% and 89%. METAL reported at the launch of GPT-6 Astra that alignment gains and a critical capability rating arrived together.

The way out is MCP. V7 Go exposes Context Graph querying and ingestion through its own MCP server, so customers can use it straight from ChatGPT and other compatible clients, and they can create V7 Go workflows through MCP inside Codex. Combined with simpler workflow design, that cut the time to build a medium-length workflow from about an hour to roughly 20 minutes.

Attempts to attach agents to company data have run all year. METAL reported that OpenAI shipped an agent that answers questions about company data, and there too the deciding factor was not model performance but the path to internal material. V7's own product page points at the same spot. Even the best LLMs max out at a 1M token context window while a firm holds 15 years of deal history, and the model forgets all of it the moment the session ends. The same page lists more than 300 integrations and more than 120 agents.

The next goal is a graph that moves first. The V7 team is working toward workflows that start when facts in the Context Graph change, flag inconsistencies, and show people which analyses need another look. A restated fund report, for example, would prompt V7 Go to point at the work still resting on the old figures.

What this case shows is that the axis of agent competition is shifting from models to memory. If the same model produces different results depending on how a company's material is structured, then the asset that gets priced is not model access but organized context. Rizzoli said that "finance firms getting real value from AI will not be the ones with the most agents. They will be the ones with the best context."