Stanford’s AI Index Report 2026 contains two numbers that, placed next to each other, describe the state of enterprise AI more accurately than anything else published this year.
The first: 88% of organizations report using AI in at least one business function, up from 78% a year earlier (Ch4, Fig 4.3.1).
The second: depending on company size, between 3% and 10% have a fully scaled AI program (Ch4, Fig 4.3.6).
Adoption is close to universal. Production is close to rare. The distance between those two numbers is where most enterprise AI budget currently goes to die, and closing it is an engineering problem, not a strategy problem.
What does “scaled” actually mean here?
The AI Index draws on McKinsey’s 2025 State of AI survey, which asks organizations to place themselves on a five-stage ladder. The distribution is unflattering.
| Company revenue | Not using | Experimenting | Piloting | Scaling | Fully scaled |
|---|---|---|---|---|---|
| Under $100M | 9% | 39% | 22% | 25% | 5% |
| $100M–499M | 8% | 33% | 32% | 23% | 4% |
| $500M–999M | 5% | 31% | 32% | 29% | 3% |
| $1B–4.9B | 6% | 22% | 31% | 32% | 9% |
| $5B+ | n/a | 17% | 31% | 39% | 10% |
Source: Stanford AI Index 2026, Ch4, Fig 4.3.6. Figures are self-reported and directional.
Read the middle three columns. Across every revenue band, somewhere between half and two-thirds of organizations are experimenting or piloting, which is to say they have AI somewhere in the building but not in the critical path of anything that matters.
Agents are further back still. Across business functions, the share reporting no agent use at all runs from 66% in knowledge management to 91% in manufacturing, and scaled deployment is in single digits nearly everywhere (Ch4, Fig 4.3.8). The technology sector is the exception, with 24% reporting scaled agent use in software engineering. Everyone else is roughly where the marketing suggests they were two years ago.
There is a diagnostic question that cuts through self-assessment faster than any maturity model: if your model provider had an outage right now, who would notice, and what would they do? If the honest answer is “nobody,” you are experimenting. If it is “our customers, and we have no fallback,” you are scaling and under-engineered.
Is the money showing up?
Partially, and unevenly. The same survey asked organizations what AI has actually improved.
| Outcome | Improved | No effect |
|---|---|---|
| Innovation | 64% | 21% |
| Employee satisfaction | 45% | 31% |
| Customer satisfaction | 45% | 32% |
| Competitive differentiation | 45% | 33% |
| Cost | 38% | 31% |
| Profitability | 36% | 36% |
| Organic revenue growth | 33% | 39% |
| Market share | 25% | 49% |
Source: Stanford AI Index 2026, Ch4, Fig 4.3.5.
Innovation is a soft measure and it scores well. Profitability is a hard measure and it is a coin flip: as many organizations report no effect as report improvement. Revenue growth is worse than a coin flip.
The macro evidence agrees. A survey of 6,000 executives across the US, UK, Germany and Australia found widespread adoption alongside minimal realized productivity impact to date (Yotzov et al. 2026, reported in Ch4, Fig 4.4.28). Penn Wharton’s model puts AI’s current contribution to total factor productivity at 0.01 percentage points, effectively nothing.
Two readings are available. The optimistic one is Erik Brynjolfsson’s J-curve: organizations absorb the cost of adoption before the returns appear, and US productivity growth reaching 2.7% in 2025 against a 1.4% decade average may be the curve starting to turn. The pessimistic one is that a large share of what has been deployed was never capable of moving a financial metric.
Both readings point at the same operational conclusion. The organizations getting returns are the ones that got past the pilot.
Why do the pilots stall?
Not for the reason the pitch decks imply. The AI Index asked what blocks organizations from reaching fully scaled agentic AI (Ch3, Fig 3.3.10):
| Obstacle | Share |
|---|---|
| Security and risk concerns | 62% |
| Technical limitations | 38% |
| Regulatory uncertainty | 38% |
| Gaps in responsible-AI tooling and control | 36% |
| Resource or budget constraints | 34% |
| Unclear or insufficient business value | 32% |
| Immature vendor or ecosystem landscape | 28% |
| Organizational resistance | 23% |
| Lack of executive support | 9% |
Look at what is missing from that list. Not one respondent group named model capability. Nobody is stuck because the model isn’t smart enough.
Every item on that list is a delivery problem. Security scoping, integration, controls, observability, cost modelling, proof of value. This is the ordinary work of building software that a large organization can actually run. It is unglamorous, it is well understood, and it is precisely what does not happen when an AI initiative is owned by an innovation team with a twelve-week mandate.
Note also that executive support is the smallest blocker at 9%. The mandate exists. The engineering behind it doesn’t.
What separates the 10% from everyone else
Across the report’s data, and across the engagements where we have watched this go well and badly, the same eight things distinguish an AI system in production from an AI demo that impressed a steering committee.
1. A baseline measured before anything was built. You cannot demonstrate improvement against a number you never captured. This is the most common irrecoverable mistake, because by the time anyone asks, the old process is already gone.
2. Evaluation suites, running in CI. Not vibes, not a spreadsheet of twenty hand-checked outputs. When the provider ships a new model version (and they will, on their schedule, not yours), an eval suite is the only thing standing between you and a silent regression in production.
3. Grounding you control. Retrieval against a corpus you own, with citations surfaced to the user. The report’s hallucination data is a design constraint, not a footnote: across 26 top models, error rates on a knowledge-versus-belief benchmark span 22% to 94%, and several models collapse when a false premise is presented as something the user believes (Ch3).
4. A designed failure path. Agents on the OSWorld benchmark improved from roughly 12% to 66.3% task success in a single year, which is remarkable, and still means failing about one attempt in three (Ch2). Systems that degrade gracefully and hand off to a human are shippable at that rate. Systems that assume success are not.
5. Traceability. Model version, prompt version, and retrieval corpus version, logged together per inference, so that in six months somebody can reconstruct why the system produced a specific output. Every governance conversation eventually lands here.
6. Permissions scoped tighter than the user’s. The 62% who named security as their blocker are mostly right to. An AI feature that inherits broad service credentials is a data exfiltration path wearing a friendly interface.
7. A run cost that survives volume. Per-transaction inference cost, modelled at 10x current usage, with model-tier routing and caching designed in. A pilot’s economics almost never survive contact with production traffic unexamined.
8. Portability. The frontier has converged: six labs now sit within 25 Elo points of each other, and the AI Index itself concludes that competition is shifting “toward cost, reliability, and domain-specific performance” (Ch2). Meanwhile vendor transparency fell, with the Foundation Model Transparency Index dropping from 58 to 40 in 2025 (Ch3). An abstraction layer that lets you swap providers is cheap to design in and expensive to retrofit. We wrote about this in detail in AI portability in enterprise applications.
Where to start if you are in the 55%
Start with the work that is not about the model.
Instrument the baseline. Map the data (access, quality, lineage), because “the model isn’t accurate enough” is, in our experience, usually “the data behind it was never fit for this.” Write the eval suite before the feature. Scope the permissions. Decide where the human sits and how their override is recorded. Then build.
One more number worth sitting with. Across 12,000 European firms, AI adoption raised labour productivity by 4%, with a 5.9 percentage point gain for every 1% of spend that went to training (Aldasoro et al. 2026, in Ch4, Fig 4.4.28). Enablement has the best-evidenced multiplier in the entire report, and it is the line item most often cut first.
The AI Index’s own framing of where the returns are is worth quoting exactly, because it is a scoping rule disguised as a finding: gains are strongest “when work can be divided into well-defined, repeatable tasks with clear quality monitoring.” If your candidate use case fails any of those three tests, that is not a reason to abandon it, but it is a reason to expect the harder version of the project.
The gap between 88% and 10% will close. It will close for the organizations that treated AI as software with unusual failure modes, rather than as a category of procurement. If you are trying to work out which side of that you are on, our AI implementation discovery checklist is the assessment we run before scoping any AI implementation engagement, and pilot to production in regulated enterprises covers the same crossing under compliance constraints.
All figures cited from the Stanford HAI Artificial Intelligence Index Report 2026 (9th edition). Organizational adoption and outcome data originate in McKinsey’s 2025 State of AI survey, is self-reported, and should be read as directional. Model performance figures are a March 2026 snapshot.
