MIT Technology Review Insights poll finds only 16% of organizations get transformative value from AI and 12% run agents in production

No model or product release is involved. The article promotes a downloadable report. Atlan's event page links it to "The AI Context Gap" study, available to Gartner IT Symposium registrants, with Atlan named as an example vendor (Atlan). Atlan sells a "context layer" product and says it delivers context to agents through MCP, APIs, SQL, search and vector stores (Atlan). No specific models or versions are named.

• Gartner's model: a context layer combines three parts: semantics (business meaning), operational state (real-time data) and provenance (auditability) (Mexico Business News).
• Atlan's model: a knowledge graph, an ontology and versioned "skills," plus context mining, a build/test/deploy lifecycle, learning loops that turn agent traces into shared knowledge, retrieval, and governance/drift monitoring (Atlan).
• Evaluation: Atlan AI Labs reports a text-to-SQL benchmark of 174 queries and 522 evaluations across 13 tables and 94 columns (Atlan). No public dataset, model identity or harness is cited, so the benchmark cannot be reproduced. Thirteen tables is far smaller than a typical enterprise schema.

• 16% transformative value; 12% agents in production: from a sponsor-funded poll with no published sample size or method (MIT Technology Review).
• 38% SQL-accuracy gain from context (p<0.0001): vendor-run and unreplicated. Atlan's guide says 174 queries and 522 evaluations, while a search summary describes it as "522 enterprise queries" (Atlan). Customer reports of "up to 5x" accuracy are anecdotal.
• Up to 80% accuracy gain and 60% cost cut: an analyst forecast, not a measured result. Coverage does not disclose the methodology (Yahoo Finance).
• Only one in five organizations get significant gen-AI value: a Gartner figure broadly consistent with the poll's 16% (Mexico Business News).

The "context gap" idea builds on MIT NANDA's 2025 finding that the main barrier is learning and context retention, not infrastructure or talent. That finding drew on 300+ public initiatives, 52 interviews and 153 survey responses (Virtualization Review). The concept also renames older semantic-layer and metadata work: vendors now present BI semantic layers as AI context infrastructure (Strategy). Alation argues that a context layer alone is still not enough for agents (Alation).

• Both MIT Technology Review Insights pieces on this theme are sponsored, by Atlan and Celonis. Each sponsor sells the remedy its piece recommends.
• Sankar's claim that AGI is "already here" for practical purposes is her opinion. It is not tested (MIT Technology Review).
• The headline "failure" numbers depend on definitions. Critics say NANDA's 95% figure counted success only as measurable P&L impact six months after deployment, which ignores productivity gains (Zirous).
• Gartner says no vendor offers a complete context layer out of the box (Mexico Business News).

• Budget: treat ontology, lineage and governed definitions as a prerequisite for agents rather than optional extras. Expect to combine bought tools with custom-built parts, per Gartner.
• Measure before buying: run your own A/B test of agents with and without curated context on your own schemas, tracking accuracy and token cost.
• Integration: ask vendors to show context delivery over MCP or APIs to the models you already use, along with drift monitoring and audit trails.
• Read the survey numbers with care: the figures are directional and sponsor-funded. They suggest most firms are still piloting, but they do not show that a context layer causes ROI.