Two numbers belong on every CTO’s dashboard right now: 88 and 6.
Eighty-eight percent of organizations say they use AI regularly in at least one business function. Six percent can point to more than 5% of EBIT and call it AI’s doing.
Nearly everyone has a demo running. Almost nobody has turned it into something finance can verify.
Everyone Is “Using AI.” Almost Nobody Is Winning With It
Those figures come from McKinsey’s 2025 State of AI survey, which polled organizations across 105 countries. Adoption climbed to 88%, up from 78% the year before. Less discussed on earnings calls: out of 1,933 participants, only 109 reported that more than 5% of their EBIT and “significant value” were attributable to AI — roughly 6%, McKinsey’s definition of an “AI high performer.”
The gap isn’t a model problem. High performers aren’t running smarter foundation models than everyone else. They’re more than three times as likely to say their organization intends to use AI for transformative change, and they’re about three times more likely to redesign workflows around AI rather than bolt it onto an existing process. Buying access to a capable model is trivial now. Building the engineering discipline that turns that access into a working system is where the other 94% get stuck.
Why AI-Powered Engineering Is the Missing Layer
This is the layer AI-powered engineering fills — combining engineering expertise, AI development, and enterprise implementation so systems integrate with existing infrastructure, support live workflows, and produce value that outlasts the pilot phase.
Closing the 88/6 gap isn’t about picking a better foundation model. It’s the unglamorous work: integration, evaluation infrastructure, and workflow redesign — the difference between a working pilot and a system the business actually depends on.
What Separates the 6% From Everyone Else
McKinsey’s breakdown of high-performer behavior reads like an organizational story more than a technology one. High performers commit real budget to back their ambition: more than a third allocate over 20% of their digital budget to AI, a share most competitors don’t come close to matching.
That commitment shows up structurally, not just in spend. Half of AI high performers are actively redesigning workflows around AI rather than layering it onto processes built for a pre-AI world — and layering is exactly what produces an impressive demo and a disappointing rollout.
The high performers aren’t asking where they can automate a task. They’re asking what the workflow should look like if AI were the starting assumption. That’s a harder engineering exercise, and it’s the one most enterprises skip.
The Production Gap Deloitte Keeps Finding
Deloitte’s 2026 State of AI in the Enterprise report, based on a survey of 3,235 business and IT leaders across 24 countries, tells a strikingly consistent story from a different vantage point.
| Metric | Finding |
|---|---|
| Experiments in production | Only 25% of organizations report that 40% or more of their AI experiments have reached production |
| Expected to reach that bar soon | 54% expect to hit it within the next three to six months |
| Mature agent governance | Just 21% of companies report having a mature model for governing autonomous AI agents |
That last number matters most. Nearly three-quarters of companies plan to deploy agentic AI within two years, yet governance maturity isn’t scaling at the same rate — even as data privacy, regulatory compliance, and model quality sit near the top of leaders’ risk list.
Getting a model to do something impressive in a sandbox is the easy 20%. Getting that capability to run reliably, securely, and auditably inside a business with real compliance obligations and real legacy infrastructure is the hard 80% — and it’s where most enterprise AI investment currently stalls.
Stuck in Pilot Limbo Isn’t a Technology Problem
This is the stage where most initiatives quietly stall. The pilot worked. Leadership liked what they saw. Then it sits there — technically alive, never killed, never scaled — while everyone waits for someone to make the call.
That in-between state has nothing to do with the model losing its shine. Nobody built the plan for what comes after the applause: who owns it, what it needs to plug into, who signs off before real users depend on it every day.
Every enterprise AI initiative runs into this same wall eventually. The transition from pilot to production is where most initiatives quietly die — not at the model-selection stage, but at the stage nobody budgeted real engineering time for.
What This Means for Engineering Leaders
Enterprises aren’t short on AI capability. Foundation models are strong, adoption is nearly universal, and pilots are easy to spin up. What’s scarce is the engineering discipline to take something that works in a demo and make it work inside a real business — with its real data, real compliance requirements, and real legacy systems attached.
That’s not a strategy problem or a model problem. It’s an engineering problem, and most enterprises are underinvesting in it relative to what they’re spending on the models themselves.
The 6% aren’t winning because they found a smarter model. They’re winning because getting to production was the actual project — not the afterthought that happens once the demo gets applause.
For engineering leaders weighing where to invest next, that’s a useful filter to run any AI initiative through: is the plan mostly about which model to use, or does it account for the integration, governance, and workflow redesign work that separates a pilot from a system the business can rely on?
That honest answer, more than anything happening at the model layer, predicts which side of the 88/6 split an initiative lands on.
Related: 2026 Is the Year AI Grows Up: From Hype to Real-World Power
