Everyone is building a context graph
Enterprise AI vendors have settled on a name for what their agents need. Day AI calls its product “the context graph and agent control plane”. AlphaSense built its SuperAnalyst agent on what it calls a Context Graph. Google says Workspace Intelligence turns a company’s email, chat and files into a knowledge graph. Snowflake and Databricks build semantic models and catalogs for the same purpose.
They’re right. An AI agent is only as useful as the information it has about your company, and that information works best structured, linked to the sources it came from, and shared by the people and agents who use it.
What each vendor builds
AlphaSense launched SuperAnalyst, an always-on agent for research and market monitoring, in June. Its CTO, Raj Neervannan, described the Context Graph underneath it in July. AlphaSense builds it from documents, filings, broker research and expert calls, along with “enterprise content like investment memos, board decks, and meeting notes.” One model extracts facts and relationships from each source and a second model checks the extraction against the same material. Every claim cites its source, and AlphaSense timestamps relationships, so “what was true in Q2 is not assumed true in Q4.”
Google announced Workspace Intelligence in April, and says it “turns your scattered emails, chats, and files into a cohesive knowledge graph.” I wrote about how Google pairs that graph with a personal memory in Gemini Enterprise.
Day AI builds its graph from conversations. It calls the graph Customer Memory, “an LLM-optimized ontology built from automatically permissioned ingestion of your team’s email, calendar, meetings, and Slack.” Day AI says that “Every agent, on every team, reads and writes the same multiplayer context graph,” that each agent reports to one person, and that nothing goes out without that person’s approval.
So all three agree that a graph should be structured, linked to its sources and shared. They differ in what feeds the graph: AlphaSense reads documents, Google reads a company’s Workspace, and Day AI reads a go-to-market team’s conversations. Meta takes a fourth route, and has its business AI learn from a business’s chats with customers.
A change sent in one email
Imagine the accounts payable manager at Acme, one of your customers, writes to the account manager who looks after Acme: “From November 1, please invoice Acme Holdings instead of Acme Inc., and send invoices to ap@acmeholdings.com.” The account manager replies “Got it, thanks” and moves on.
A context graph built from that account manager’s email may well pick up the change, linked to the email it came from. On November 1, though, someone in finance sends invoices from the billing system the way they do every month. Nothing tells them to ask anyone whether Acme changed its billing details, and the invoice goes to Acme Inc. at the old address.
Facts change in AlphaSense’s sources too, but a company that changes its name or its auditor announces it in a public filing that anyone tracking the company can find. Acme announced its change in one email, to one person, and some of the people who have to act on it may never see the thread. Someone can query the graph and find what Acme’s accounts payable manager said and who they said it to, but only once they think to ask, and only if the graph has what one person learned.
Known to one person
Panopto’s survey of 1,001 US adults found that 42% of institutional knowledge is unique to the individual who holds it. After that email, Acme’s new billing details are an example of that kind of knowledge: the account manager knows, but finance doesn’t.
A graph that agents and people query solves the problem of finding the answer. Acme’s change needs something that reads the message on the team’s behalf, notices that it asks for a change to how Acme gets invoiced, and proposes the updates in the systems the team works from: a task for finance before the November run, and an update to Acme’s page in a shared brain with the new billing details, each linked to the email. Whoever owns the account confirms them, and from then on the graph, the billing run and the people who send invoices have the same answer.
That’s the layer we’re building at this+that. DoBox reads the messages a company already gets across Gmail, Outlook, Slack, Microsoft Teams, and Google Chat, turns them into tasks, and routes them to a workflow or a person. Contacts keeps the people in those messages current on its own. Our brain holds what the company knows, kept current today by workflows that write back to it as they run, with a person approving or editing each update. We’re building toward a brain that keeps most of itself current from the message stream through one maintenance workflow, but that part isn’t finished.
Try it with last month’s messages. Pick a change a customer sent in an email, like a new billing contact or a new delivery address, and check whether everyone who acts on it found out without having to ask. If not, see how DoBox can help.