Deloitte's "State of AI in the Enterprise" 2026 report surveyed 3,235 business and IT leaders across 24 countries and six industries between August and September 2025. One number frames everything else: only 25 percent of organisations have moved 40 percent or more of their AI experiments into production. The other three-quarters are still operating in pilot territory — proofs of concept, contained deployments, demos that impress a steering committee but never touch the operating line.
This is no longer 2023, when the technology was new and failure was the expected outcome. This is the third year of serious enterprise generative-AI investment, inside organisations that already have budgets, executive sponsors, and strategic mandates. The open question is no longer whether AI works. It is whether an organisation can make it work repeatedly, in production, under its own governance — and most still cannot.
The production gap in detail
The 25 percent figure needs reading carefully, because the threshold is deliberately high. Forty percent or more of experiments reaching production is not a single model in a single workflow; it implies a repeatable deployment pipeline — the plumbing, the review gates, the ownership — that turns a successful pilot into the next one without starting from zero. A further 54 percent of respondents expect to cross that line within three to six months. That is the more revealing statistic, because near-term optimism of exactly this shape has recurred in Deloitte's surveys year after year. Many enterprises are perpetually "almost there".
The ambition is not the constraint. Investment is rising, confidence is rising, and access is rising. What lags is the conversion of all three into deployed, governed systems. The distance between "we expect to scale soon" and "we have scaled" is the most durable feature of enterprise AI, and it has nothing to do with model capability.
Access has outrun adoption. Workforce access to sanctioned AI tools grew by roughly half in a single year — from fewer than 40 percent of workers to around 60 percent. That is a real infrastructure achievement. But putting a tool on someone's desktop is not the same as changing how they work. Provisioning is a procurement decision; daily use is a behavioural and process change, and the second does not follow automatically from the first. The gap between the two is precisely the organisational work that a software rollout cannot do for you.
Gains are real, but unevenly captured
The data does not support the fashionable claim that enterprise AI has failed. Two-thirds of respondents — 66 percent — report measurable productivity and efficiency gains. That is substantial: most organisations are seeing a return, even if few have scaled that return across the business.
The transformation signal is sharper still. Twenty-five percent of leaders now say AI is having a transformative impact on their operations — more than double the share that said so a year earlier. Read alongside the production figure, the picture resolves into a familiar shape: the same quarter that has built repeatable deployment is the quarter now reporting transformation. Once AI clears production at meaningful scale, its effect compounds. Getting there is the entire problem.
What this produces is a widening split rather than a closing one. A minority is compounding advantage — better pipelines feeding better deployments feeding better data feeding the next deployment. The majority is iterating on pilots and re-running the same proof of concept with a new model. Incremental effort does not close that gap, because the leaders are not standing still while everyone else catches up.
The preparedness paradox
The most counterintuitive finding is that perceived preparedness fell. Despite more investment, wider access, and more experience, organisations rate themselves less ready than a year ago across the operational dimensions that matter. Technical-infrastructure preparedness sits at 43 percent, data-management preparedness at 40 percent, and talent preparedness at just 20 percent — the lowest of any dimension and the one leaders name most often as the binding constraint.
This is recalibration, not regression. As organisations move from a contained chatbot to agentic workflows, from a single-model pilot to a multi-model production estate, the definition of "ready" moves with them. The infrastructure that carried a customer-service assistant is not the infrastructure for an agent that acts across systems. The data governance that satisfied a recommendation engine does not satisfy a system making consequential decisions inside a regulated process. Preparedness is falling because the bar is rising faster than the capability — and the more an organisation learns about what scaled AI demands, the further off the target looks.
Talent is a structural barrier, not a hiring problem
At 20 percent preparedness, talent is the weakest link by a wide margin, and the response pattern explains why it stays that way. The most common talent adjustment is education — workforce-wide AI fluency, cited by 53 percent of respondents — ahead of upskilling and reskilling programmes at 48 percent and specialist hiring at 36 percent. Education first is understandable, and it is also insufficient on its own.
Teaching a tool is not the same as redesigning the work. Showing a claims adjuster how to use an AI assistant is education. Redesigning the claims process so AI handles routine adjudication while the adjuster concentrates on complex, contested cases is organisational change. The first yields incremental productivity. The second is what moves the production metric. The capabilities that scaled AI actually requires — process redesign, cross-functional integration, data governance, exception handling, measurement — are not primarily skills in operating a model. They are operating-model capabilities, and they are learned by deploying into production and running the result, not by completing a training module.
Agentic AI raises the bar again
Asked where agentic systems — those that plan, reason, and act with limited oversight — will have the highest impact, respondents cluster around five domains: customer support first, then supply-chain management, research and development, knowledge management, and cybersecurity. Tellingly, these are not where most enterprises began. Early adoption concentrated on content generation and analysis, where AI assists a human at the keyboard. The agentic domains are structurally harder: long decision chains, multiple data sources, orchestration across systems — exactly the integration, governance, and process discipline the preparedness data says most organisations lack.
The result is a compounding gap. Close to three-quarters of companies plan to use agentic AI at least moderately within two years, yet only around one in five report a mature governance model for autonomous agents. The organisations best placed to capture agentic value are the ones that already solved the unglamorous problems of scaled deployment. Everyone else faces both at once: fix the production-scaling problem, and prepare for systems that demand even more maturity — under a regulatory regime that is tightening in parallel. Obligations for providers of general-purpose AI models under the EU AI Act have applied since August 2025, and the broader high-risk regime is moving through its implementation phases. Governance is no longer optional plumbing; it is a precondition for putting an autonomous agent anywhere near a regulated process.
What this means for DACH Mittelstand
The Deloitte picture maps cleanly onto what we see across DACH mid-market companies. Ambition is high — executive sponsorship, allocated budget, a strategic mandate. Activation is low — pilots and contained deployments that never reach the production threshold. And the binding constraints are organisational, not technological.
The talent constraint bites harder here. Competition for AI expertise is intense, and the classic Mittelstand strengths — deep domain knowledge, long tenure, a disciplined operational culture — have not yet been converted into AI operating capability. That conversion is the whole job, and it does not happen on a hyperscaler's budget. The companies making progress treat scaling as an organisational-transformation programme with technology components, run inside their own constraints — a handful of well-chosen workflows, a real owner, a review cadence — not as a technology project with organisational side-effects.
The data constraint is equally structural. Decades of ERP-centric architecture have produced data estates that are comprehensive but siloed — excellent for reporting, poorly suited to feeding AI systems that need integrated, current, governed data. Deloitte's global 40 percent figure for data-management preparedness is, in our experience, generous next to the typical DACH enterprise that has not yet invested in integration for AI-specific workloads.
The production-scaling problem is solvable, but not by running more pilots. It requires an honest assessment of the organisational barriers — infrastructure readiness, data maturity, talent, governance — and a structured route from experiment to production. That is precisely what the AI Operating System diagnostic measures, and it is the same set of dimensions the leaders in the 25 percent had to solve first.
Closing the ambition–activation gap
The 54 percent expecting to cross the production threshold within six months may be right. Or they may be repeating the optimistic near-term forecast that Deloitte's data has documented for three years running. What separates the two is whether an organisation treats scaling as a deployment problem or an operating-model problem.
The 25 percent that have already scaled share a profile. They redesigned workflows rather than bolting AI onto existing ones. They built data infrastructure for AI workloads rather than repurposing reporting pipelines. They designed the operating model — roles, review cadences, exception handling, measurement — before deploying at scale. And they treated talent as capability-building, not tool training. None of that is exotic, and none of it is reserved for enterprises with hyperscaler budgets. It is disciplined operating work, and it is what the next two years will reward as agentic systems and tightening regulation raise the bar again.
A Fit Call pinpoints where your AI work is actually stalling — infrastructure, data, talent, or governance — and what it takes to move your first workflows from pilot into production, before another planning cycle closes with the gap still open.
References: Deloitte, "State of AI in the Enterprise," 2026 edition, global report and press release (3,235 leaders, 24 countries, six industries, surveyed August–September 2025); European Commission, AI Act regulatory framework.