V7’s Context Graph: bind provenance at write time

A confidential information memorandum sits on the desk: financials, deal terms, management background, risk clauses. A human needs most of a day to work through it. V7’s agent produced the screening note in fifteen minutes. The multiple is not the interesting part. The interesting part is that for those fifteen minutes, the model was not reading files.

What V7 is actually selling

V7 is an agentic platform for finance, insurance and real estate, and its core product is called the Context Graph. According to a customer story published with OpenAI, when files arrive V7 Go scans repositories such as SharePoint and Google Drive for entities, relationships, facts, attributes and metrics, links each fact to a record in a graph, and preserves cited evidence back to the original source. It falls back to document search only when the graph does not hold the answer.

That story also carries a set of result numbers. Asset managers screen deals 21 times faster than before, compressing a full-day process into fifteen minutes. One financial services team cut review time from more than 100 hours to under 10, saving about 12,000 dollars of expert cost per task. An insurance team reduced claims processing errors by 13.5% against a manual baseline. The same material repeats a V7 claim that 50 to 100 step workflows finish in minutes at 99.9% accuracy. (OpenAI customer story: How V7 gives AI agents institutional memory)

Attribution matters here, so let us be exact. Every number above is vendor or customer reported, with no third-party replication. Two other figures in the material are different in kind: V7 states that on HERB, a benchmark for finding and connecting information across enterprise systems, its retrieval-only system beat the official baseline by 69% and cut hallucinations on unanswerable queries by 38%. Those are V7’s own tests. And the 99.9% claim never defines which task, which distribution, or what the denominator is.

The structural shift: context became a product

The numbers are not the point. The point is that V7 pulled memory out of the workflow and shipped it as a product of its own, and exposed both querying and ingestion over an MCP server. Customers can query the firm’s memory from ChatGPT or another compatible client, and can create V7 Go workflows from Codex.

That split runs against the route most teams took. For the past year the default has been to stuff context into the model window, add retrieval when the window fills, add reranking when retrieval misses, and keep all of it inside one application, where context is an implementation detail. The graph route turns context into a component that can be bought, swapped and audited on its own. The side effect quoted in the same story, a medium-length workflow going from about an hour to about twenty minutes, is what that architecture buys you.

For a buyer this changes the question. Yesterday the question was whether your agent can do deal diligence. The question now is how your memory layer gets built, how it updates, and how I verify it. The first is about capability. The second is about an asset.

Why this works now

Turning documents into a graph is not a new idea. People built enterprise knowledge graphs five years ago. What changed is that three conditions arrived at once.

Models can now hold a long instruction chain together: the fifty to one hundred step workflows quoted in the story depend on not drifting after dozens of tool calls. Extraction quality crossed the useful line, so pulling entities out of contracts and metrics out of quarterly reports no longer needs hand-written rules and human proofreading. And the interface standardised: MCP lets any client read and write the memory layer, so a customer no longer has to move into one vendor’s interface to use it.

Cost moved in the same direction. The story claims a 78% lower cost per document than the previous generation of smaller models. That is a vendor number, and the direction matches the industry: extraction is repetitive, structured work, which is exactly where unit costs fall fastest. Remove any one of the three conditions and the graph stays inside a consulting deck.

Provenance belongs to write time

This is the judgment we take away, and it is the sentence we keep repeating to clients: provenance is a write-time action, not a query-time action.

The reasoning is not complicated. Assembling evidence at query time means the system has to look backwards at the moment it answers, asking where this sentence came from. Once a file has been edited, deleted, or copied three times, that step depends on the current state of storage rather than on the data itself. Bind the source when the fact enters the system, which file, which passage, which date, and provenance becomes a property of the fact and travels with it. Any later query can re-verify it.

The V7 material describes exactly that when it says facts preserve cited evidence back to the original source and each run carries an auditable trace. One caution follows from it: on that path the system producing the evidence is the system being audited. A vendor-generated audit trail is genuinely useful for internal review, and it is not a substitute for independent verification. On outcome metrics such as an error rate, who measured matters more than what number came out.

We do a smaller version of this ourselves. Every cited document gets a source record, and its origin lives in the same queryable store as the content, so any claim can be traced back to its row. It is unglamorous, and its value shows up later: when two versions of a fact disagree, you do not need to remember, you need to look.

How to read the three numbers

First, read 21x and 10x as direction, not as a promise. They come from a single vendor’s customer cases, with no published measurement method, sample size or baseline. The direction is believable, because deal diligence and insurance review are document-dense, structurally repetitive and checkable, which makes them the best early targets for agent work.

Second, the precondition attached to the claims figure is worth more than the conclusion. The result followed granting agents knowledge of all previous claims and existing policies. In procurement terms, that 13.5% was bought with data integration, not with a better model. If your firm cannot get its historical cases into a queryable structure, the same product will not reproduce the same result.

Third, the jump from 78% to 89% on V7’s own benchmark is a story about benchmark saturation. V7 says existing models had saturated its old test set, so it built a harder set of graph-query questions across four difficulty levels. On easy, medium and hard, both models scored close to perfect. Only the very-hard tier separated them, by eleven points. Two conclusions follow. A vendor score means nothing without knowing the test set. And once a test set approaches saturation, it stops distinguishing models and becomes useful only for validating a specific workflow.

Our position

Three claims.

First, when buying agents, look at the memory layer before the protocol layer. Tool calling and orchestration frameworks are commoditized and replaceable in a fortnight. A memory layer binds your files, your entity names and your historical conclusions, and replacing it is measured in quarters.

Second, never evaluate a memory layer from a demo. Write the exam on your own corpus. Demo questions are prepared for demos. The questions that decide the outcome sound like: how many versions of that restated quarterly report does the system hold?

Third, provenance has to be separable from the system. An audit trail that only the audited party can generate and read is self-certification. A buyer needs the ability to export the underlying records with their sources, and to re-verify the same fact through an independent query.

A four-question acceptance test

This test needs no new tooling. It needs your own documents, and it fits in half a day. It filters out most systems that demo beautifully and land badly.

Build or buy

The second question a buyer faces is the path: build retrieval and reranking in-house, or buy a memory layer.

Building is controllable and cheap at the start. Put the files in full-text search, bolt on a vector store, and you have a demo in two weeks. The cost arrives in phase two, when the business asks whether this policy and last year’s claim belong to the same insured party. Vector search cannot answer that structurally, so you build entity resolution, relationship extraction and provenance handling yourself. That layer is roughly the work of buying a product, minus the vendor who maintains it.

Buying has harder preconditions. Your files need to be digital and reachable, and entity naming needs to be reasonably stable. Meet both and a memory layer buys back data engineering time. Miss them and you have bought an expensive demo environment.

There is a third path, and it is usually our first recommendation: get the documents into machine-readable shape, unify the naming, kill the scans, and collapse the scattered versions. That step produces no demo, and it decides how far either of the other two can go.

Where this does not apply

Stated boundaries beat inflated scope.

• The payoff concentrates in document-heavy, repetitive, checkable work: financial services, insurance, real estate, legal. For a marketing team with few documents and shifting definitions, the graph never pays for itself.
• The memory layer is only as good as the stability of your entity model. If your business definitions get rebuilt every quarter, fix the definitions before buying memory.
• A graph is not free. Extracting entities, relationships and metrics from files is a data engineering project, and the precondition on that 13.5% figure is exactly that work.
• Every number in the customer story is single-sourced. Treat self-reported results as direction with unverified magnitudes, and retest on one or two of your real processes.
• For smaller firms, getting documents into something full-text searchable beats buying a graph. A memory layer assumes files are already digital and reachable.

The decision in front of you

Back to the memorandum. Fifteen minutes does not prove the model got smarter. It proves the agent did not have to spend time working out what this fund manager is called in the previous system. That work used to belong to people, it stayed in their heads, and it left with them.

So there is one question worth answering: does the context your company depends on live in documents, in the memory of your long-serving staff, or in a structure that can be queried, verified and handed to the next model without retraining it?

References: OpenAI customer story, How V7 gives AI agents institutional memory (21 September 2026): openai.com/index/v7— the 21x, 100-hours-to-10, 12,000 dollars, 13.5%, 69%, 38% and 78%/89% figures all come from that story and are vendor or customer self-reported, not independently replicated. V7 website: v7labs.com.