V7 Bets That Agents Need a Graph, Not a Bigger Context Window

OpenAI published a customer story on V7, a London-founded company building agentic document workflows for finance, insurance, and real estate. The piece describes V7 Go, a platform that ingests a company's files into what V7 calls a Context Graph — a structured record of entities, relationships, and source-linked evidence that agents query instead of re-deriving context on every request. It is a customer profile written by a model vendor, so the architecture is more interesting than the endorsements.

The problem being described

The framing in the post is that current models reason well but arrive at each task with no knowledge of the business around it. Which version of a fund report is current. Whether the entity named three different ways across three systems is the same entity. That information sits in data rooms, spreadsheets, emails, and internal tools, unresolved and invisible to an agent.

The conventional answer is retrieval: search the documents, stuff the results into context, reason. OpenAI's account of V7's position is that this pays the same discovery cost on every request — dozens of searches burning time and tokens, with relationships between facts often missed because they were never represented anywhere a search could find them.

V7 was founded in 2018 by Alberto Rizzoli and Simon Edwardsson, who had previously built a computer vision accessibility app together.

How the Context Graph is put together

V7 Go connects to repositories such as SharePoint and Google Drive and scans them for entities, relationships, facts, attributes, and metrics. When a new file arrives, the system identifies the companies, funds, people, or other entity types defined in an ontology, attaches each extracted fact to a new or existing record, and preserves a citation back to the source document. If a query cannot be answered from the graph, V7 Go falls back to retrieval over the underlying documents.

Two properties of that design are worth separating from the marketing.

The first is that the graph is a persistent artifact rather than a per-request construction. V7 describes it as an order of magnitude cheaper and faster to traverse than long-context approaches — a vendor claim without published methodology, but a directionally believable one. Resolving an entity once and storing the result is cheaper than resolving it on every call, assuming the resolution is correct and stays correct.

The second is that every fact carries provenance. In deal screening and underwriting, an answer without a citation is not usable output, it is a lead. Building the citation link at extraction time rather than reconstructing it afterward is the difference between an auditable system and a plausible one.

For long-running agents, V7 Go keeps recent exchanges in the model's active context and pushes older material into the graph for retrieval when needed — a working set in context, a durable store behind it.

Model routing as an engineering decision

V7's model usage is tiered, and the post is unusually specific about it. GPT-5.6 Luna handles structured extraction and other high-volume work. GPT-5.6 Terra and Sol handle chat, the Go Agent path, and steps requiring more reasoning or tool use. GPT-6 Astra is being introduced on the most demanding Context Graph queries, including financial analysis spanning thousands of documents. AI-generated workflows map each step to fast, medium, and smart tiers.

V7 maintains a benchmark suite covering citation accuracy, extraction quality across hundreds of document types, answer correctness, instruction following, latency, cost, and real-world workflows, and tests new models against it. Edwardsson is quoted saying V7 chose OpenAI as its default because it performs best on the multi-step tool workflows the product depends on, and that in V7's Context Graph benchmark, GPT-5.6 Sol cut the tool-call error rate from 2.7% with GPT-5.5 to 0.2%.

Several reported efficiency results are more concrete than the usual case-study fare. V7 measured a 78% lower cost per document with GPT-5.6 Luna compared with GPT-5.4 mini. Moving document-heavy workloads from the Chat Completions API to the Responses API reduced token use by roughly 5% for some PDF-heavy workflows and improved caching reliability. Key workflows with several external calls now finish up to 50% faster in V7's latest harness.

There is also a telling detail about evaluation hygiene: V7 says GPT-5.6 Sol had saturated many of its existing benchmarks, so it built a harder set of graph-query questions on messier real-world data across thousands of documents, spanning four difficulty levels. On the very-hard tier, V7 reports Sol at 78% and GPT-6 Astra at 89%, with both near 100% on easy, medium, and hard. A team that notices its benchmarks have stopped discriminating and builds new ones is doing the work.

Where scepticism belongs

The outcome numbers in the post are V7's, reported through OpenAI, with no independent verification and no denominators.

V7 says agents complete 50-to-100-step workflows in minutes at 99.9% accuracy. Accuracy of what, measured against which ground truth, is not stated, and a figure that precise across workflows that long invites the question. On HERB, a benchmark for finding and connecting information spread across enterprise systems, V7 reports its retrieval-only system outperformed the official baseline by 69% and reduced hallucinations on unanswerable queries by 38% — both relative improvements against an unspecified base rate.

Customer results are similarly framed: asset managers screening deals 21 times faster, reducing a full day to about 15 minutes; a financial services team cutting review time from more than 100 hours to under 10, saving $12,000 in expert costs per task; insurance teams reducing claims processing errors by 13.5% against a manual baseline after giving agents history on prior claims and policies. The insurance figure in particular is refreshingly modest. But these are single-customer anecdotes without sample sizes.

What teams building agents can take from this

The transferable idea is not the product, it is the shape. If an agent has to rediscover the same organizational facts on every invocation, that cost compounds with every step in the workflow, and the errors compound with it. Materializing entity resolution and fact extraction once, with citations attached, converts a recurring inference problem into a storage and freshness problem — which is a problem software engineering already knows how to attack.

Freshness is where this gets hard, and V7's roadmap concedes as much. The stated goal is workflows that trigger when facts in the Context Graph change, flag inconsistencies, and surface analyses that need review — a restated fund report prompting a flag on work that still uses the old figures. That is cache invalidation with regulatory consequences. Until it exists, a structured memory of the business is also a structured record of what the business used to be true about.

The distribution choice is the other portable lesson. V7 Go exposes Context Graph querying and ingestion through its MCP server, so customers can reach it from ChatGPT and other compatible clients, and can create V7 Go workflows through MCP in Codex. V7 says that, together with simpler workflow design, this cut the time to build a medium-length workflow from around an hour to roughly 20 minutes. Shipping context as a protocol endpoint rather than only as a UI is increasingly what separates infrastructure from an application.

What to watch

• Whether the change-triggered workflows land, and how V7 handles restatements and retractions — invalidation is the unsolved half of durable agent memory.
• Independent evaluation on HERB or a comparable enterprise retrieval benchmark, rather than vendor-reported deltas against unpublished baselines.
• Whether MCP-exposed context graphs become a standard integration point for enterprise agents, or stay a per-vendor feature.
• How graph construction cost and quality hold up as ontologies grow across finance, insurance, and real estate simultaneously.
• Whether tiered model routing stays worthwhile as frontier models get cheaper, or collapses back into using one model for everything.