Agentic AI Productivity Fell at 30% of Companies: McKinsey

Your AI coding assistant might be making your team slower. That is the uncomfortable takeaway from McKinsey's Technology Trends Outlook 2026, which found that productivity fell in nearly a third of companies after teams started using agentic AI tools, according to coverage of the report by Mint. It is the sharpest public data yet on the agentic AI productivity trade-off, and it complicates the assumption that adoption equals progress. The report put it bluntly: "Almost everyone, it seems, is 'vibe coding'—but that doesn't always lead to value."

The gap between activity and output is the detail that should worry anyone measured in commits. One study cited in the report found that AI tools increased coding activity by one hundred eighty percent, but shipped releases rose by only thirty percent. More code does not always lead to more products, the report added. Developers are typing more and shipping roughly the same, which is the kind of stat that makes a performance review conversation awkward.

Part of the drag is trust, and the numbers are stark. McKinsey found that about forty-six percent of developers worldwide actively distrust the accuracy of AI tools, according to BIIA's write-up of the report, compared with thirty-three percent who trust them, while only three percent said they highly trust AI-generated output. "Developer acceptance of agentic software tools remains a meaningful constraint," the report noted, adding that for companies hoping to benefit from agentic development, the trust gap matters. Tools nobody fully trusts still get used every day, which is how you end up reviewing machine-written code at midnight.

Then there is the money, moving in the opposite direction from the evidence. Equity investment in agentic software development companies grew from nearly nothing through twenty twenty-three to about five billion dollars in twenty twenty-five, then crossed sixty-one billion dollars in the first half of twenty twenty-six, driven largely by a sixty billion dollar acquisition of Cursor, according to the report. Relevant job postings rose two hundred twenty-one percent between twenty twenty-four and twenty twenty-five. The industry is pouring money in at a pace the evidence has not caught up with. For anyone managing a roadmap, the implication is plain: a bigger commit graph is not the same thing as a faster team.

What this means for the entry-level pipeline

For students and new grads, the timing is genuinely strange. Companies are hiring for roles built around tools that their own data says can slow teams down, and pay premiums for AI-exposed roles have surged even as the productivity case wobbles. Knowing how to prompt an agent is becoming table stakes. Knowing when to ignore one and do the work yourself might be the skill that actually gets noticed. The trust numbers point the same way: reviewing, testing, and catching what the model got wrong is the job now, and it pays to be the person who does it well.

The playbook is still unwritten

McKinsey's warning was aimed squarely at the how, not the what. "Without a systematic approach, agentic software development can lead to unintended outcomes," the report stated, adding that organizations are racing to deploy AI at scale without any proven road maps, with workforces that need upskilling and legacy systems that need updating. This site recently covered the other side of the same coin: AI is everywhere at work, but the strategy behind it has not caught up. The fix is unglamorous. Decide what the tool is for, measure whether it actually shipped anything, and keep humans in the review loop. Companies that skip those steps are, by the report's own numbers, the ones most likely to end up slower than when they started.

There is a version of this problem aimed at identity rather than output. As agents take on more of the workflow, the race to verify which agent did what is becoming its own industry. The agentic AI productivity numbers explain why that matters. If three in ten teams get slower with these tools, someone will eventually ask what the agent actually did all quarter, and "vibe coded" is not an answer that survives an audit.