Only about one in three agentic AI projects reaches production. A survey of 300 data, AI and technology executives, published by MIT Technology Review Insights on October 5, 2026, put the figure at 34 percent. The report names the same blockers in the same order: legacy data systems, security and privacy concerns, and a lack of knowledge and context. Context graphs have moved to the centre of the response. The term now sits in Gartner’s Hype Cycle for Agentic AI, 2026, and vendors and data platforms have started to build around it. The reason is plain. Agents do not fail only because they cannot find a file. They fail because they cannot tell which file is current, who owns it, and why the last decision went the way it did.
What context graphs add beyond search
A knowledge graph records what things are and how they relate. A context graph records how decisions happen. In Gartner’s definition, summarised in a May 26, 2026 GlobeNewswire release, context graphs extend knowledge graphs by capturing operational context, decision traces, governance metadata and semantic meaning, while preserving temporal and causal links that show why a decision was made and how the reasoning unfolded.
That is a different record from search. Search returns the policy document. A context graph can show the policy, the exception granted last quarter, the person who approved it, and the contract that forced the change. Gartner also notes, in the same summary, that this kind of structured memory connects people, systems, documents, policies and events, so agents can reason across workflows and surface dependencies that a document list hides.
The commercial bet behind the term is large. A summary of Gartner’s context graph research by Atlan reports two Gartner predictions: more than 50 percent of AI agent systems will use context graphs by 2028, and context engineering features will be built into 80 percent of software tools used to build AI applications by 2028, with accuracy gains of at least 30 percent through that period. Those are forecasts, not results. They still tell you where enterprise budgets expect the problem to sit: not in the model, but in the record the model reads.
The production gap is a knowledge gap
The MIT Technology Review Insights survey makes the same point from the buyer side. A small group of production leaders, where an average of 61 percent of agentic projects advance beyond pilot, report stronger knowledge capabilities than the rest, with the clearest gap in semantics: shared definitions an agent can rely on. Data fragmentation was the most cited challenge to expanding agent access to knowledge, named by 55 percent of respondents. Read the full report summary and the pattern holds across its findings. Firms expect the biggest gain to come from strengthening the structural link between company data and agents, and they plan to invest in ingestion pipelines, AI-ready APIs, retrieval and knowledge graphs to do it.
Length does not solve this. Chroma’s context rot research tested 18 leading models and found that performance degrades as input length grows, even on simple tasks. Giving an agent more documents can make its answer worse if the record does not rank, date and connect them. A graph is one way to impose that order before the model sees anything. It decides what is authoritative, what is superseded, and what belongs together, instead of asking the model to guess from a pile of text.
A context graph is not a shortcut
The counter-argument deserves space. A graph built from stale or wrong sources will produce confident, well-linked errors. Gartner is explicit, in the Atlan summary linked above, that context graphs do not replace knowledge graphs. They augment them. The semantic work still has to happen: definitions, ownership and permissions.
Security is the harder limit. Among the production leaders in the MIT survey, 72 percent cited security and privacy as a major concern, a higher share than in the wider group. A record that connects decisions across teams also connects access. If an agent reads the graph with broader rights than the person who asked, the graph becomes a leak with structure. Permission checks have to run at query time, for the person asking, on every answer. That rule is not optional decoration. It is the condition that makes a shared decision record safe to build.
Start narrow. Pick one workflow where agent errors cost money, capture the decisions in that workflow first, and measure answer accuracy and policy compliance before you widen the scope. Gartner’s own guidance, as summarised by Atlan, treats this as an incremental, domain-by-domain effort, not a company-wide transformation project.
Search finds files. The graph keeps the why.
Companies already store outcomes. CRMs, ERPs and warehouses record what happened. They rarely record the reasoning a person would need to repeat the decision, or to challenge it. That gap is why the same questions return to the same senior people, and why agents trained on documents alone repeat the document and miss the judgment.
Jilu is built on the same distinction. Its context map keeps the path behind each decision, not only the file at the end of it. For the full account of how that record is collected, reviewed and kept current, read How Jilu Works. The related question of what agents should store at all is covered in What a Company Knows: AI Agents Need It Written Down, Not Just Memory. A company that keeps its why in a form people and agents can both read does not just answer faster. It stops re-deciding what it already decided.