Handing out car keys before anyone builds the road. That’s roughly what happened to enterprise software in the first half of this year.
Gartner puts it plainly: by the end of 2026, forty percent of enterprise applications will embed a task-specific AI agent, up from under five percent twelve months ago. Eighty percent of the applications shipped or updated in the first quarter already have one baked in. If you work anywhere near a corporate IT estate, you are already, in some quiet way, managing an agent you didn’t explicitly ask for.
Here’s the part that doesn’t make the keynote slide: only thirty-one percent of enterprises have that agent actually running in production, according to S&P Global Market Intelligence and McKinsey. Not “embedded.” Running. Doing the job, in front of customers, with consequences. Telecoms, retail, and banking are out front at 47–48 percent — not because they’re braver, but because they did the unglamorous work first: identity, data foundation, the governance layer nobody puts on a roadmap slide. Healthcare sits at 21 percent. Public sector at 18. They’re not behind because they’re slow. They’re behind because getting it wrong there costs more than a bad quarter.
So: eighty percent shipped, thirty-one percent driving. That gap between “embedded” and “in production” isn’t a rounding error. It’s the industry’s actual state of readiness, and it’s smaller than most vendors would like you to notice.
It gets worse further down the funnel. Only twenty-five percent of AI initiatives currently deliver the ROI they promised. Only sixteen percent reach enterprise-wide scale. McKinsey finds that just ten percent of organisations are meaningfully scaling agents within any single function — not across the business, within one. And Gartner, not an organisation prone to theatrics, predicts forty percent of agentic AI projects will be cancelled by 2027, for the oldest reasons in the book: cost, unclear return, and governance that never showed up.
I’ve spent the better part of two decades building the unglamorous layer underneath these numbers — federated analytics frameworks, AI governance models, the plumbing that has to exist before anyone lets a system act on its own. So none of this surprises me. What’s worth saying out loud is why it keeps happening anyway.
Boston Consulting Group has a framework for this that I think about often: roughly ten percent of AI value comes from the algorithm, twenty percent from the technology and infrastructure around it, and seventy percent from people, process, and how the organisation actually redesigns itself around the new capability. Deloitte’s 2026 enterprise survey found that eighty-four percent of companies have not redesigned a single job or workflow around AI. They bought the agent. They did not touch the seventy percent.
That’s the governance gap, precisely. It isn’t a compliance department’s problem, and it isn’t really a technology problem either — the models work fine. It’s the gap between installing a capability and rebuilding the organisation around what that capability is now accountable for. You can ship an agent into a workflow in a sprint. You cannot redesign who owns a decision, who gets blamed when the agent is wrong, or what “good” looks like when the person doing the work is no longer entirely a person — in a sprint. That takes the kind of unglamorous, unfunded, slow work that never makes it into a Q1 board deck. Every number above is what happens when an organisation skips it and ships anyway.
The industry keeps asking how fast agents can drive. Nobody’s asking who’s allowed to fail the test.
Sources: Gartner (agent embedding and cancellation forecasts); S&P Global Market Intelligence and McKinsey, Q1–Q2 2026 enterprise AI surveys (production adoption by industry); McKinsey, “The State of AI” (Nov 2025); Boston Consulting Group, 10-20-70 Framework (Sept 2025); Deloitte, “State of AI in the Enterprise” (2026).
Leave a Reply