← All posts
AI   Nov 19, 2026 · 7 min read · by , Co-Founder

How AI Agents Use a Work Graph

In short

AI agents use a work graph to trace any task up to the goal it serves and to find blocked or neglected items, with the activity behind each answer attached. A work graph is a model of a company's plan and activity: goals, KPIs, projects and tasks as objects, with links that say which serves which.

When Google introduced its Knowledge Graph in May 2012, Amit Singhal summed up the change in three words: things, not strings. A search engine that matched character strings could find pages containing "Taj Mahal".

One that knew the Taj Mahal as a thing, with a location and an architect and a relationship to other things, could answer questions about it. The same shift is now under way inside companies, and for the same reason. An AI agent that only has your documents has strings. To reason about your plan it needs the things.

A work graph is the name several vendors, Asana among them, use for a model of this kind applied to work. The idea is older than the current interest in agents, but agents are what make it matter, because a graph is a structure a machine can walk.

What does a work graph contain?

The Vindaris Work Graph, whose schema is public, has four parts.

  1. Objects. The plan items are strategies, goals, KPIs, projects and tasks. People and teams give them context, as do meetings with a transcript.
  2. Links. A task contributes to a project, which contributes to a goal or a KPI, and so on up to the strategy. A project can be blocked by another project. Each plan link records whether a person drew it or the graph added it from work activity, and a person can undo the second kind.
  3. Signals. Each signal is one observed piece of work activity attached to the objects it concerns, for example a Jira ticket that changed status or a Slack thread in which someone said a project is blocked.
  4. Alerts. When the signals around an object add up to a problem, the graph raises an alert to named people, with the reasons in plain language and the signals behind each reason.

One detail deserves attention. Every plan item carries two separate values: the status a person set by hand, and the state Vindaris derives from the signals. The two often disagree, and the disagreement is informative. A project its owner marked green three weeks ago, in which no task has moved since, is a different thing from a green project with ten completions this week. You can find the longer definition in our glossary entry on the Work Graph.

What can an AI agent do with a work graph?

What does an agent do with this that it could not do with a folder of documents? Four things, each of which comes down to walking the graph in some direction.

The first is walking up. Given a task, the agent follows the links to the project, the goal and the strategy above it. It can then answer the question every employee has on a Monday morning, namely whether the thing in front of them serves anything leadership cares about. A task with no path upward is unaligned effort, and the agent can say so before the person spends a day on it.

The second is walking down. Given a goal, the agent collects everything under it and asks which parts are blocked or have gone quiet. This produces the board update, and it produces it from tickets and threads. Nobody on the team writes a status report to feed it.

The third is asking why. An alert in the graph carries its reasons and the signals that support them. So when the agent reports that a project is blocked, it can point to the thread and the date. However confident a model sounds, an answer you can check is worth more than one you cannot, a point we made at length in our post on AI and task-level data.

The fourth is noticing absence. Documents record what exists. A graph also shows what is missing, such as a goal with no project under it or a priority that a meeting named and no plan item matches.

It shows, too, when a person's recent activity has moved away from what the rest of their team is working on. Search cannot return a document that nobody wrote. A graph makes the gap visible as a missing link, and Vindaris flags it weeks before a quarterly review would.

Why does a work graph have to maintain itself?

Interestingly, none of this is new in principle. Consultants have drawn goal trees on whiteboards for decades, and we drew a great many of them over seven years of running strategy execution programs in more than 300 organizations. The trees were accurate on the day of the workshop. They decayed afterward because people had to maintain them by hand, and people always stop doing that eventually.

The difference now is that the upkeep can come from work activity. Vindaris analyzes that activity across the tools your teams use, such as Jira, Slack, Microsoft 365 and meeting transcripts, places new tasks in the graph, and updates the state of every item. An agent connected over MCP queries the result. That upkeep turns the Work Graph into an AI context layer.

Singhal's line was about search, and search changed more than anyone expected once it knew about things. What changes in a company when every agent, and through the agents every employee, can walk from any task to the strategy it serves?

FAQ

What is a work graph? A work graph is a model of a company's plan and activity. Strategies, goals, KPIs, projects and tasks are objects, and links between them record which serves which. Signals from work tools keep the state of each object current.

How is a work graph different from a knowledge graph? A knowledge graph records facts and relationships in general. A work graph is specific to the plan and the activity around it, and it carries a current state for every item, derived from that activity.

Is the Vindaris Work Graph schema public? Yes. The object types, link types, signal kinds and alert definitions are published on GitHub at github.com/Vindaris/work-graph-schema under a CC BY 4.0 license.

How does an agent query the Work Graph? Through the Vindaris MCP server. Tools such as get_goal and get_work_graph_alerts return a goal with its links and the open alerts with their reasons.