Studies on generative AI adoption indicate that very few pilots generate real value in production. The model is not the bottleneck. Three things hold it back: scattered context, ambiguous business meaning and governance that cannot keep up with the speed of agents. The few cases that succeed treat context as infrastructure, not as a detail.
The figure has become a catchphrase, and for good reason. Recent research (including the study The GenAI Divide, from MIT's NANDA initiative, in 2025) shows that the overwhelming majority of generative AI pilots in companies never generate measurable impact. A lot of investment, a lot of impressive demos, and almost nothing that survives contact with real operations.
The easy reading is to blame the model: "we needed a better LLM". But today's models are powerful enough for most corporate cases. The problem lies elsewhere.
The model is not the bottleneck. Context is.
Think about what happens when someone asks a generic AI: "why did gross margin fall last quarter?". The model has no access to your numbers, does not know what "gross margin" means inside your company, and does not know the rules about who can see what. It answers vaguely, or worse, makes something up.
The model is not being dumb. It is a lack of context. And the missing context splits into three fronts.
1. Distributed context
Your data and the knowledge about it live in different places: the CRM, the data warehouse, orphan spreadsheets, Slack messages and the heads of half a dozen people. No AI agent reaches all of that on its own. A pilot that works on a clean demo database falls apart when it has to cross the company's fragmented reality.
2. Ambiguous business meaning
"Active customer", "revenue", "TAM" mean different things on the internet and in your operation. Generic AI knows the world's thousand definitions and gets yours wrong. Without a layer that says which definition is canonical in the company, every answer becomes a bet.
3. Governance that cannot keep up
Access policies, PII classification and usage rules usually live in static documents. AI agents act in real time, along any path. When governance is not applied at the moment of access, the company is caught between two bad extremes: locking AI down out of fear, or opening up and being exposed.
What the few success cases do differently
The pilots that cross into production have one thing in common: they stop treating context as a detail of the prompt and start treating it as infrastructure. Instead of pasting pieces of data into the prompt and hoping, they build a structured foundation that brings together, in one place:
- the data and how it relates,
- the business meaning (glossary and ontology),
- governance applied at the point of access,
- and lineage, the traceability of every number.
That is the idea of a context layer: what goes in raw at the bottom arrives understandable at the top, ready for any agent to reason safely. With that, the same LLM starts delivering a completely different result.
How to cross from pilot to production
A practical path, in three moves:
- Unify: connect the sources and turn each asset into a node in a single context graph, with end-to-end lineage.
- Collaborate: humans and AI co-create the context (glossary, ontology and policies), all versioned. The analyst becomes a curator, not a SQL writer.
- Activate: expose that context to every agent in a standardized way, via MCP, with governance applied at runtime.
When context becomes infrastructure, the pilot stops being a fragile demo and becomes something operations can rely on every day.
Conclusion
Generative AI's "valley of death" is not a model problem. It is a context problem. The companies that understand this stop searching for the perfect LLM and start building the foundation that makes any LLM understand the business. That is where the pilot becomes production.