The MIT Initiative on the Digital Economy and McKinsey released separate survey results this month looking at how enterprise AI investment is actually converting into deployed production systems. The headline numbers are similar. Roughly 85 percent of enterprises report active generative AI pilots. Roughly 25 percent report at least one pilot that has moved into full production. The gap between those two numbers has barely moved since the first wave of surveys after ChatGPT's release in late 2022, and the reasons keep coming back to the same set of issues.
The dominant reason pilots stall is not model capability. The models are good enough. The problem shows up in the connective tissue around the model. Data quality in most enterprise source systems is worse than anyone wants to admit. Permissioning and access control have not been designed for situations where an AI agent might pull from fifteen systems on behalf of a user with varying entitlements in each one. The governance frameworks to decide what an AI can and cannot do with which data have not caught up with the speed of the pilot work. When a pilot graduates to production, it hits each of these issues simultaneously, and the program slows down while the infrastructure teams catch up.
A second reason is that the value case for a given pilot is often easier to demonstrate than to measure in production. A coding assistant pilot shows immediate qualitative benefits to the engineers using it. Measuring the dollar value of those benefits requires baselining pre-AI productivity against post-AI productivity across enough teams and enough time to strip out noise. Most enterprises have not done that rigorously. When the pilot moves to a procurement conversation about scaling to ten thousand seats, the CFO asks for the hard number, and the team that ran the pilot often does not have it.
Vendor sprawl is also quietly eating into ROI. A typical large enterprise is now paying for Microsoft Copilot or Google Duet, plus a couple of category-specific AI tools for their contact center, plus a couple for marketing content, plus a developer assistant, plus whatever internal projects the data science team has built on foundation model APIs. Each tool is priced on a per-seat or per-transaction basis. The combined spend across the portfolio is often running two to three times what any single budget line item shows. CIOs are starting to rationalize.
The MIT survey found that companies reporting the highest production deployment rates shared a few specific characteristics. They invested heavily in data infrastructure before they bought AI tools. They built a central AI governance function with real authority and real staff, not a committee. They tied pilot approval to a specific business owner who was accountable for the production outcome. And they set a hard limit on pilot duration, typically 90 days, after which the pilot had to either move to production or shut down. Companies without those structures tended to accumulate pilots indefinitely without much production output.
There is an emerging pattern in the companies that are getting real production wins. The wins tend to be in narrowly defined workflows with clear inputs and outputs, high transaction volume, and measurable time or cost outcomes. Insurance claims intake, customer service triage, software development code review, contract redlining, and finance month-end close processes are where the early at-scale deployments are showing up. The more ambitious cross-functional agent scenarios, where an AI is supposed to plan and execute multi-step work across systems, are still mostly in pilot.
For buyers thinking about AI spend in 2026, the survey data points to a few practical conclusions. Invest in data infrastructure in parallel with AI tool purchases, not after. Force pilots to a 90-day clock with clear production criteria. Baseline productivity before you start so you have a credible ROI story when the CFO asks. Pick two or three workflow targets, not fifteen. And be realistic about vendor pricing models. The per-seat pricing that looks affordable at a 500-user pilot gets expensive fast at 50,000 seats, and the real leverage in negotiating with the big platform vendors comes from having a credible alternative.
Model providers know the pilot-to-production gap is a problem for their own growth. The major AI labs have all shipped enterprise features in the last six months aimed specifically at closing it. Anthropic's Agent SDK, OpenAI's enterprise deployment tooling, and Google's Vertex AI updates are all oriented around making it easier to deploy, monitor, govern, and audit AI workloads at scale. Early adopter feedback suggests these tools help, though none of them solve the upstream data quality problem.
The broader takeaway from the MIT and McKinsey data is that the enterprise AI story in 2026 looks less like a revolution and more like a long, uneven grind. The companies making progress are doing the boring work on data, governance, and workflow redesign that does not get covered in press releases. The companies stuck at pilot are the ones that bought the tools first and assumed the infrastructure would catch up. A year from now the pilot-to-production ratio will still be the number that matters, and the distribution will still be uneven across industries and across companies within the same industry.