Agentic AI Productivity Falls at 30% of Firms, McKinsey Finds

The finding behind the headline

McKinsey's Technology Trends Outlook 2026 has put a number on a problem many engineering leaders already suspected. At about 30% of companies, productivity went down after teams started using agentic AI tools.222 The consultancy says the main cause is how companies roll the tools out: firms that add autonomous systems without changing how their development teams work run into operational trouble.2 The report adds that "almost everyone, it seems, is 'vibe coding'—but that doesn't always lead to value."8

The report's second statistic says more than the first. In one study it cites, AI tools raised coding activity by 180%, but shipped releases rose by only 30%.23 McKinsey's own conclusion is that more code does not reliably mean more products.8 For anyone who measures engineering output in commits or lines of code, that gap between activity and delivery is the core of the problem.

The report also describes a trust problem among the people expected to use these tools every day. About 46% of developers worldwide say they actively distrust the accuracy of AI tools. Only 33% say they trust them, and just 3% trust the outputs highly.22 McKinsey calls developer acceptance "a meaningful constraint" on getting value from agentic software development.8

Where the coverage agrees, and where it overreaches

The reporting was fast and highly consistent. Most outlets used nearly identical wire copy and repeated the same figures for the productivity decline, the coding-to-release gap and the trust gap.1710 That consistency makes the core numbers reliable. It also means the framing spread just as widely, and some of it needs correcting.

The first problem is scope. Headlines described the decline as hitting "companies using agentic AI," but the reporting makes clear that McKinsey was talking about agentic tools in software development, meaning autonomous coding systems.27 This is not a verdict on AI agents in customer service, finance or HR. Applying the 30% figure to every kind of agent deployment goes further than the evidence does.

The second problem is investment. Equity investment in companies building agentic software rose from negligible levels through 2023 to about $5 billion in 2025, then passed $61 billion in the first half of 2026.2 Some summaries presented this as a broad surge across the sector.4 But almost all of the 2026 figure comes from a single $60 billion acquisition of Cursor.22 Without that deal, the jump looks far less dramatic. The fair reading is that buyers are making very concentrated bets, not that money is flooding in evenly.

The third problem is wording. One outlet paraphrased the 46% figure as developers who "don't trust these tools much."5 That is softer than McKinsey's "actively distrust."8 The difference matters, because active distrust affects behaviour: developers who distrust a tool check its output twice, route around it, or quietly stop using it.

The pattern fits a broader correction

The McKinsey finding matches a wider pattern of companies reconsidering their AI rollouts. In September, McKinsey's State of AI 2026 survey found that 80% of respondents said AI had made them personally more productive. Yet only 37% reported a positive contribution to earnings, and only about 6% qualified as high performers.9 An Atlanta Federal Reserve working paper found that 90% of executives say AI has not yet produced productivity gains at their companies.13

Recent reporting from large companies shows the same thing. Meta reportedly scaled back plans to cut staff and lean more on AI agents after finding employees spent growing amounts of time fixing technical and security errors.13 Microsoft concluded that "adding agents to a broken process still leaves a broken process" and that "speeding up one step just creates a longer queue at the next."13 That is essentially McKinsey's coding-to-release gap seen from inside a company. Microsoft found that training, especially pairing early-career engineers with experienced colleagues, improved productivity.13

Analysts expect more pullbacks. Gartner projects that 40% of enterprises will demote or decommission autonomous agents by 2027 because of governance gaps.9 It also points to weak spots such as audit logs, approval thresholds and unclear ownership when an agent makes a mistake.17 Costs are another constraint: about one in five McKinsey survey respondents said AI operating costs, including tokens, had limited their use.9 Agents can call a model many times in a single task, so a bill that looked fine in a pilot can rise sharply once the agent runs across a whole department.17

Meeting assistants, copilots and the 'doom loop'

The McKinsey data covers coding, but office productivity tools face a similar risk. Microsoft executive Ryan Roslansky described receiving a document a colleague had clearly generated with AI. He summarized it with his own AI and came away with no new ideas. He warned of a "doom loop" in which machines do both the reading and the writing and leave behind an "ocean of sameness."13 It is the knowledge-work version of 180% more code but only 30% more releases: output rises while understanding does not.

Vendors are pushing ahead regardless. OpenAI has launched "dots," always-on agents that work through Slack and Teams to manage projects, prepare documents and build software. It says weekly users of Codex and ChatGPT Work now exceed 35 million.16 Microsoft is adding Autopilot to Copilot, an agent that works as a digital coworker with its own identity and permissions.16 Startups are giving agents names, avatars and seats in email threads.11 That raises a risk the coding data hints at: BCG researchers found managers caught 18% fewer errors when told work came from an "AI employee" rather than an AI tool.11 If human-like branding makes reviewers less careful, it could make the trust problem McKinsey describes worse instead of solving it.

Training has not kept up. PwC's survey of more than 49,000 workers found AI use at work rising to 64%, while the share who could access the learning resources they needed fell from 59% to 51%.16 An Indeed survey found more than half of employees say they are not getting the AI training they need.13 McKinsey itself says organizations are deploying AI at scale "without any proven road maps," with workforces that need upskilling and legacy systems that need updating.

What it means for jobs

The labor-market picture is mixed. AI was cited in more than 112,000 announced US job cuts in 2026, making it one of the top stated reasons for layoffs, according to Challenger, Gray & Christmas.13 Another count puts the figure above 120,000 through September.19 Yet a University of Pittsburgh study found that productivity did not rise after companies invested in AI and cut staff. Its authors also noted that share prices tend to fall after AI-related layoff announcements.13 Read together with McKinsey's report, the risk is clear: companies may be cutting people on the strength of productivity gains that have not arrived yet.

The effects on workers are mostly showing up in hiring and pay rather than mass firings. The Federal Reserve's October Beige Book said businesses in five of its 12 districts were using AI-related productivity gains to shrink through attrition, delay hiring or plan for fewer workers.19 Research reported by The New York Times found slower wage growth in AI-exposed occupations while employment held steady. That suggests companies are capturing the gains by holding down pay rather than cutting jobs.20 Stanford HAI data shows employment of US software developers aged 22 to 25 has fallen nearly 20% since 2024.9 Those are the same developers whose productivity McKinsey says is uneven.

Demand is also shifting toward people who can make agents work. McKinsey reports that job postings related to agentic AI rose 221% between 2024 and 2025.22 On Wall Street, postings that mention "agent orchestration" jumped 1,721% this year, and governance-related skills now appear almost twice as often as model-building skills.12 Workday found demand for basic AI skills such as prompting fell 25% after peaking in January. Demand for building AI tools and automating workflows rose 51%.15 Only 28% of leaders in Workday's survey expect AI to reduce headcount, while 40% expect to get more out of the staff they already have.15

The verdict

The McKinsey finding does not show that agentic AI fails. It shows that agents make whatever process they are added to more visible, along with its flaws. The 30% of companies that lost productivity most likely bought tools without redesigning how work moves from code to release, and without winning over the developers who have to trust what comes out.28 Spending is still rising, and the hiring market increasingly rewards orchestration and governance skills over prompting.1215 The companies most likely to see real returns are those that invest in workflow redesign and training, rather than those adding the most agents.