The Enterprise AI Budget That Swam Away

Every enterprise now seems to have a budget shaped like a rubber band: stretch it toward artificial intelligence, and it snaps the other way. Last week McKinsey reported that many companies are blowing past their AI budgets because agents and agentic software development usage exploded over the last year. The consultancy gathered clues from search queries, news articles, patents, research publications, equity investments, and talent demand. If AI is a casino, the report counted the chips by peering through the slot machines.
The twist is that spending more is not yet delivering more everywhere. Agentic coding is becoming standard in firms, but the glow of novelty is fading faster than a laptop battery in a conference room. Only one quarter of companies said the tools had meaningfully sped up product development, while thirty percent saw productivity fall after teams began using agentic AI. It is the digital equivalent of buying a rocket sled to walk to the mailbox, then wondering why the mailbox is on fire.
I once watched a small fintech team celebrate a new coding agent like a puppy they had adopted. They called it Coderius and gave it a dashboard that looked suspiciously like a spaceship console. By Thursday, Coderius had produced a thousand lines of elegant but unlovable code, and the humans spent Friday rewriting comments, tests, and error messages as if performing an autopsy on a very shiny duck. The lesson stuck with me because I have done the same thing with my own smart fridge, which cheerfully suggested I buy salmon for a dinner that was, in fact, pasta.
McKinsey senior partner Martin Harrysson argues that value will not arrive by sprinkling agents over old routines. Companies need smaller, highly leveraged teams that supervise agents through execution. In other words, humans are no longer just building every brick; they are hiring very obedient ghosts to stack bricks, then checking whether the ghosts have invented castles. Prakhar Dixit, a McKinsey partner, adds that two week sprints can become a continuous loop of drafting, testing, debugging, and documentation between agents and human reviewers. The tricky part is keeping speed while protecting maintainability, quality, and control.
Still, the money keeps flowing like a river after a dam has a philosophical crisis. Agentic software development investment is forecast to grow more than twelve times from 2025 to 2026, while Gartner puts global AI spend at $2.7 trillion this year, up nearly half. SpaceXAI bought Anysphere, maker of Cursor, for sixty billion dollars, and Anthropic and OpenAI are pushing coding tools into the enterprise mix. Adoption also depends on developer trust, and nearly half of developers worldwide say they distrust AI accuracy. The fix may be boring: clean docs, structured knowledge graphs, indexed repositories, and enterprise context systems. Companies with tidy codebases are giving their agents a map. Companies without one are asking them to navigate by vibes. So until the manuals improve, the humans should still keep pencils sharp, budgets guarded, and their sense of humor well charged.