Let me be direct about something the official summary buries. The Stanford AI Index is genuinely excellent at tracking what AI can do. It’s less useful — and this is a deliberate structural choice on their part — at tracking what organizations actually get from deploying it. Those are two completely different questions, and mixing them is how you end up with breathless stats that make no strategic sense.

Here’s what’s unambiguously real: model performance has improved dramatically. Inference costs dropped 280-fold since 2022, landing at roughly $0.07 per million tokens by October 2024. Open-weight models closed the performance gap with closed models to just 1.7 percentage points on major benchmarks — down from an 8-point gap two years ago. If you’re a developer who hasn’t revisited open-weight models since 2023, you’re making decisions based on outdated assumptions. That gap matters.

The geopolitical picture is messier than the U.S. investment lead suggests. China closed benchmark gaps that were double-digit numbers not long ago, and leads the world in AI publications and patents. The U.S. still produces the lion’s share of notable models — about 90% in 2024 — but the perception that this is a settled race is not what the data shows. And here’s something I found genuinely interesting: countries with higher AI investment correlate with higher public skepticism about AI. The more you’re exposed to it, apparently, the more cautious you become. Make of that what you will.

The incident data is where I want to slow down. 233 AI-related incidents in 2024 — up 56.4% from 2023. These range from deepfake fraud to chatbot interactions linked to user harm. 85% of security professionals say they’ve taken action on AI-related security concerns, up from 79% the prior year. This is not a future risk. It’s a present operating environment.

This is the framework I promised in the introduction. It doesn’t exist in the Stanford report, the McKinsey survey, or the Gartner analysis individually. It only becomes visible when you read them together — and it explains, more clearly than anything I’ve found, why the ROI numbers are so ugly despite the adoption numbers looking so good.

The three datasets I’m triangulating: Stanford’s capability benchmarks, McKinsey’s organizational maturity self-assessments, and Gartner’s incident and abandonment projections. When you plot them against the same adoption timeline, a gap appears — and it’s at a very specific stage.

The gap concentrates between Experimenting and Expanding. That’s where 80% of organizations that fail to see EBIT impact get stuck. They’ve run successful pilots. They have internal champions. They even have budget. What they don’t have is a measurement framework that survived contact with production, and a governance structure that existed before the first incident rather than being built in response to one.

I want to name the specific mechanism, because it’s not obvious. At pilot stage, teams self-select both the use case and the success criteria. They pick workflows where AI helps, and they measure the things AI is good at. When they expand to other workflows — different teams, different data quality, less favorable use cases — the performance degrades. But the governance expectations were set by the pilot’s best-case results. The gap between those expectations and production reality is where abandonment happens.

Five cases. Four with sourced outcomes. One failure that I think is the most instructive thing in this entire article — and I’m going to give it more space than the successes, because that’s where the lessons are.

I’m going to be more selective here than most roundups. There are seven tool categories the AI Index data implies matter most. I’ll give you the category first, then representative options — because picking a specific tool without understanding the category is how you end up locked into something that doesn’t serve you six months later.

One thing I want to flag about this table: the Governance category is where I see organizations systematically underinvest relative to the actual risk profile they’re carrying. If you’re in a regulated industry — finance, healthcare, legal — and you’re not using at least one tool in that category, you are making a bet about incident frequency that the AI Index data does not support. The 56% incident increase happened to organizations that were largely doing what felt like responsible deployment.

Gartner’s 2025 Hype Cycle names agentic AI and AI-ready data infrastructure as the two fastest movers. McKinsey projects autonomous AI agents handling 15% of daily work decisions by 2028, up from essentially zero today. Deloitte forecasts that 50% of GenAI users will have launched agentic pilots by 2027.

I’m going to be direct about what I think those numbers mean. They mean the governance gap is about to get much worse before it gets better. If only 1% of organizations are currently mature in AI deployment, and 50% are about to launch agentic pilots, the collision between those two statistics is going to produce a lot of the incidents that end up in the AI Index 2027 report.

For SMBs specifically: the “GenAI Divide” Deloitte describes is real, but it’s not inevitable. The inference cost collapse — 280x since 2022 — means open-weight models that were enterprise-only two years ago are now accessible on free tiers. DeepSeek and Mistral run at a fraction of GPT costs with comparable performance on most commercial tasks. The gap isn’t access anymore. It’s measurement discipline and governance, which cost almost nothing to implement and almost everything to neglect.

For executives: Gartner estimates 40% of agent projects will be abandoned by 2027. That statistic is a prediction about organizations that deployed before they built the governance to sustain deployment. Whether your organization is in that 40% is a decision being made right now, not in 2027.