$2.5 Trillion in AI Spending. 95% Zero ROI. The Problem Isn't the Models.

Here are two numbers that should not coexist.

$2.52 trillion. That’s what Gartner projects organizations worldwide will spend on AI in 2026 — up 44% from 2025, rising to $3.3 trillion by 2027.

95%. That’s the percentage of organizations that reported zero ROI on their generative AI projects in 2025, according to MIT research cited by 1BusinessWorld’s analysis of enterprise AI returns.

Two and a half trillion dollars flowing into AI. Ninety-five percent of the organizations spending it seeing nothing come back. That’s not a market correction waiting to happen. That’s a market correction happening right now — and most companies don’t realize they’re on the wrong side of it.

The CEO Reality Check

The MIT number could be dismissed as academic if it existed in isolation. It doesn’t.

PwC’s 2025 Global CEO Survey found that 56% of CEOs say AI produced neither increased revenue nor decreased costs over the prior twelve months. Not a majority of startups. A majority of CEOs — the people with the clearest visibility into whether investments are producing returns. Only 12% reported that AI delivered both revenue growth and cost reduction.

Twelve percent. After years of investment, after billions in infrastructure, after every major vendor promising that AI would redefine productivity — twelve percent of CEOs saw both sides of the ROI equation materialize.

The conversation in boardrooms is shifting. It’s no longer “how fast can we adopt AI?” It’s “why isn’t the AI we already adopted producing results?”

The Model Fallacy

The default answer to the ROI problem has been: we need better models. Bigger context windows. More parameters. Faster inference. The assumption is intuitive — if the output isn’t good enough, the model must not be good enough.

The assumption is wrong.

IEEE Spectrum reported in January 2026 that “over the course of 2025, most of the core models reached a quality plateau.” GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 — the major models converged to remarkably similar capability levels. The gap between the best model and the fifth-best model narrowed to the point of irrelevance for most enterprise use cases.

And yet ROI didn’t improve. Models got better. Returns didn’t follow.

Stack Overflow’s 2025 Developer Survey quantified the disconnect from the practitioner side: developer trust in AI-generated output dropped to 29%. Two years earlier, trust levels were above 70%. Sixty-six percent of developers reported that AI solutions are “close but miss the mark” — not wrong, not useless, but not quite right. As Stack Overflow’s analysis put it: the trust gap isn’t about model capability. It’s about the gap between what the model can do in theory and what it delivers in practice.

That gap has a name. It’s called context.

What the 5% Do Differently

If 95% of organizations are seeing zero ROI, the interesting question isn’t “what’s wrong with AI?” It’s “what are the other 5% doing?”

McKinsey’s research on AI high performers — the roughly 6% of organizations achieving meaningful returns — found a consistent pattern. These organizations were 2x more likely to have redesigned end-to-end workflows before selecting models. They didn’t start with the technology. They started with the work: what information flows into each decision, what context is needed at each step, what the human needs to know and when they need to know it.

Then they engineered the pipeline to deliver that context.

The 5% aren’t using fundamentally different models. They’re using the same GPT-4, the same Claude, the same Gemini. What differs is the infrastructure around the model — the system that determines what the model sees when it processes a request. They invested in context delivery. The other 95% invested in model access and assumed the context would take care of itself.

It doesn’t.

The Missing Layer

Andrej Karpathy — co-founder of OpenAI, former Tesla AI Director — defined the discipline precisely: “the delicate art and science of filling the context window with just the right information for the next step.”

That definition contains the word most enterprises are ignoring: “just.” Not all the information. Not a data dump. Not “here’s every document in the knowledge base, good luck.” Just the right information. For the next step.

This is context engineering, and it’s the layer sitting between the model and the ROI that enterprises keep expecting but not receiving.

Consider what happens in a typical enterprise AI deployment. A company licenses GPT-4 or Claude. They connect it to their knowledge base through RAG — Retrieval-Augmented Generation. They build a chatbot or an internal assistant. And they expect the model to deliver value.

What actually happens: the model receives a question, retrieves a batch of vaguely related documents through vector similarity, and generates an answer based on whatever showed up in the retrieval results. Sometimes the answer is excellent. Sometimes it’s confidently wrong. The difference isn’t the model’s capability — it’s whether the right context happened to surface.

“Happened to surface.” That’s the phrase that explains the ROI gap. In the absence of context engineering, whether the model produces a useful answer is partially a matter of luck. The model is capable of generating the right answer. The system just doesn’t consistently deliver the information the model needs to do so.

The 5% achieving ROI didn’t leave this to chance. They built the pipeline.

The Cancellation Wave

The consequences of this gap are already materializing. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear value, or unacceptable risk.

Forty percent. Of the most hyped category in enterprise AI.

The pattern is predictable. Organizations launch agentic AI projects — autonomous systems that take actions, not just generate text. These agents need even more context than chatbots because they’re making decisions, not just answering questions. Without context engineering, the agents make decisions based on incomplete information. Incomplete decisions produce bad outcomes. Bad outcomes produce cancellations.

The projects that survive the cancellation wave won’t be the ones that picked the best model. They’ll be the ones that solved context delivery — that built the infrastructure to ensure their AI systems consistently receive the right information at the right time.

This is the same pattern we’ve seen in every enterprise technology wave. The value isn’t in the technology itself. It’s in the implementation layer that connects the technology to the actual work. CRM didn’t produce ROI until companies redesigned their sales processes around it. Cloud computing didn’t produce ROI until organizations re-architected for distributed systems. AI won’t produce ROI until enterprises engineer the context pipeline.

What This Means

The $2.52 trillion isn’t going away. AI spending will continue to grow. The question is whether organizations will keep spending on model access alone — or start investing in the context infrastructure that makes models productive.

This is the problem grāmatr was built to solve. Not by being another model. Not by being a bigger knowledge base. By engineering the context that makes every model more effective — classifying what each request needs, delivering precisely the right information, and improving that delivery with every interaction.

The models are good enough. The context pipeline is what’s missing.

If you want to see how context engineering works in practice, start with how grāmatr approaches it. If you want to see what it produces, look at the proof.