The enterprise AI deployment wave that began in earnest in 2024 has now been running long enough to produce something rarer than benchmark results or analyst projections: real operational data. And the picture that data paints is more useful, and more honest, than anything the marketing decks predicted.

The headline finding from multiple enterprise surveys in early 2026 is this: agentic AI deployments that succeeded did so primarily in one category of work -- high-volume, rule-bounded, exception-light processes where the cost of a mistake is low and the throughput gain is high. Invoice processing. Tier-1 support triage. Document classification. Compliance checklist review. In these domains, the productivity numbers are real. Cycle times cut by 60 to 80 percent. Error rates lower than human baselines once the systems stabilize. Cost per transaction meaningfully lower.

The deployments that failed -- and a substantial number did -- concentrated in a different category: judgment-intensive work where the exception is the rule, where context accumulates across interactions, and where a wrong answer does not surface until downstream. Legal contract review with novel clause structures. Customer escalations involving multiple prior touchpoints. Financial underwriting decisions where the inputs are structured but the risk assessment requires industry-specific pattern recognition that was never in the training data.

What makes this analytically interesting is not that agents fail at hard things. That was predictable. What is interesting is the mechanism of failure. The most common failure mode in 2025 and early 2026 was not catastrophic error. It was confident drift: agents producing plausible-looking outputs that were subtly wrong in ways that took weeks or months to detect, because the error rate was low enough that it looked like noise. By the time the pattern was visible, the outputs had been embedded in downstream processes.

The enterprises that caught this early share a structural characteristic: they built the measurement layer before they built the deployment layer. They defined what "wrong" looked like before they launched. They had human review processes that were not just for exceptions but for random sampling of non-exception cases. They treated the first six months of production as a calibration exercise, not a success story.

The enterprises that caught it late, or have not caught it yet, treated deployment as the finish line. They measured inputs and outputs but not the quality of the reasoning in between.

The practitioner take is blunt: an agentic system without an audit layer is not a production system. It is a pilot wearing a suit.

Source: McKinsey State of AI Report 2025

OpenAI's commercial strategy through early 2026 has shifted noticeably toward the enterprise segment, with dedicated account management, custom fine-tuning pipelines, and SLA commitments that were not part of the original API product. The shift reflects a competitive reality: the commodity model layer is compressing margins, and the defensible revenue is in deep workflow integration where switching costs are structural, not contractual. For enterprise leaders evaluating AI vendors, the implication is clear: the pricing you negotiate today is not the pricing floor. The companies building moats around workflow integration will have pricing power that pure API providers will not. Source: The Information

The true cost of enterprise AI is arriving on balance sheets in a form nobody's pilot budget anticipated: not compute per query, but the fully-loaded cost of the data infrastructure required to make the models actually useful. Retrieval pipelines. Vector databases. Embedding refresh cycles. Guardrail layers. Observability tooling. A Gartner analysis from late 2025 found that for every dollar enterprises spent on model API costs, they were spending three to five dollars on the surrounding infrastructure. The model cost is the visible number. The infrastructure cost is the real one. Enterprises that planned AI budgets based on API pricing are now mid-project and out of budget.

The EU AI Act's high-risk system requirements are now active for a category of enterprise deployments broad enough to affect most multinationals operating in Europe: AI systems used in employment decisions, credit and insurance underwriting, and critical infrastructure management. The compliance posture of most enterprises, per conversations with practitioners, ranges from "we have a legal memo" to "we have a task force." Very few have what the Act actually requires: documented risk assessments, registered systems in the EU database, ongoing human oversight mechanisms with documented protocols, and post-market monitoring plans. The enforcement window is open. The question is who gets the first headline. Source: European Commission AI Act Implementation

Google's strategy of embedding AI capabilities directly into Workspace -- Docs, Sheets, Gmail, Meet -- at no additional marginal cost for existing enterprise customers has shifted the competitive dynamic for standalone AI productivity tools in a way that is only now becoming visible in enterprise purchasing data. The question enterprise leaders need to answer is not whether Google's embedded AI is better than the best-in-class standalone tools. It usually is not, yet. The question is whether the friction reduction of having it already authenticated, already integrated, and already on the contract outweighs the capability gap. For the majority of knowledge workers, the answer appears to be yes. For the 20 percent of power users who have actually changed how they work, it is not. Source: Google Workspace Blog

Anthropic's commercial trajectory through 2025 and into 2026 has been notable not just for its growth rate but for its customer mix. The company has attracted a disproportionate share of regulated enterprise customers -- healthcare systems, financial institutions, defense contractors -- who cite its documented safety research and constitutional AI approach as a factor in vendor selection. This is the first time "safety credibility" has functioned as a sales differentiator at meaningful scale. Whether that premium holds as competitor models close the capability gap, or whether safety becomes table stakes rather than a differentiator, is one of the more interesting strategic questions in the enterprise AI market right now. Source: Anthropic

That is McKinsey's estimate of the annual value that could be unlocked by generative AI across global industries -- a number that has been cited in approximately ten thousand slide decks. The number almost never cited alongside it: the same report estimates that capturing that value requires retraining or redeploying roughly 375 million workers globally by 2030. Those are not separable facts. The gap between the first number and the second is where most enterprise AI strategies currently live.

They built the capability. They proved it works in controlled conditions. They launched. And then they stopped measuring.

Not because they are lazy or careless. Because the implicit theory of the project was that deployment was the hard part, and once you got there, the system would either obviously work or obviously fail. The idea that it would work, mostly, but in subtly wrong ways that would take months to surface -- that was not in the project plan.

The companies that are getting real value from agentic AI right now are the ones who treat deployment as the beginning of the measurement problem, not the end of the engineering problem. They have someone, usually multiple people, whose job it is to find out where the system is quietly wrong. They treat that as a permanent function, not a launch checklist item.

That function does not have a great name yet in most organizations. It is not QA in the traditional sense. It is not an audit in the compliance sense. It is something closer to what epidemiologists call surveillance: the ongoing, systematic monitoring of a system's behavior in the wild, specifically looking for the low-frequency errors that would never surface in a dashboard.

The enterprises that build this capability will be the ones writing the case studies in 2028. The others will be writing the postmortems.

Onward.

The Agentic Enterprise by Spearhead