Nearly Half of Companies Cut Back AI Agents After Costs Blew Past the Benefits, KPMG Survey Finds

Nearly half of company leaders have pulled back AI agent deployments because the bills got bigger than the benefits. That's according to KPMG's Global AI Pulse survey for the second quarter of 2026, covering 2,145 senior leaders across 20 countries at firms with over $50 million in annual revenue, as reported by Forbes contributor Sandy Carter.

Forty-nine percent of leaders in that survey said they scaled back or delayed AI agent deployments once operating costs outweighed the benefits. This follows a string of reports this week documenting how enterprises deployed AI agents without the security, cost controls, or governance to manage them.

This isn't a retreat from AI. Seventy-nine percent of leaders still call AI a top investment priority, up from 74% the prior quarter, and average AI spending held steady at $188 million per organization, according to the same KPMG data. The share of firms saying AI is now part of everyday work jumped to 22% from 13% the quarter before, the largest single-quarter jump KPMG has tracked. In Asia-Pacific, 81% of companies say AI is already delivering meaningful value, up from 69% three months earlier.

Companies aren't quitting. They're realizing they built something they can't meter.

Why the bills exploded

Most AI vendors have shifted from flat subscriptions to usage-based, token pricing, according to Forbes. Every action an agent takes, every retry, every failed attempt, every unnecessary tool call racks up metered cost. Complex agent tasks chew through tokens fast, and most companies never built the visibility to see it coming.

VentureBeat's Pulse Research, surveying 107 enterprises on agent orchestration, found one in five enterprises has no real-time way to stop a runaway agent before the bill arrives. That's not a hypothetical. That's the mechanism driving KPMG's 49% pullback.

The same VentureBeat survey found the typical enterprise runs 3.1 orchestration platforms at once, not because they love complexity but because they're deliberately avoiding lock-in to any single model provider. Microsoft AI Foundry/Copilot Studio shows up in 70% of stacks, OpenAI's Agents SDK in 68%, Anthropic's Claude Platform in 47%. Flexibility across models drives purchasing decisions (29%) far more than picking the best model (10%).

Most of what companies call "agents" aren't actually agentic. VentureBeat found 47% of respondents say only 26-50% of their so-called agents are genuinely orchestrated autonomous systems. The rest are chatbots wearing an agent costume. Companies are paying agent-level token costs for chatbot-level functionality.

The governance gap is real, and it's expensive

A WitnessAI survey covered by Corporate Compliance Insights found 68% of companies exceeded their AI project budgets in the past year. Only 4% said their AI projects always stay on budget. Just 9% said 75% or more of their AI projects delivered measurable returns.

The incidents aren't cheap either. Twenty-one percent of leaders in the WitnessAI survey reported a single AI security incident costing $1 million or more in the past year. Forty-three percent said total costs from AI-related incidents exceeded $2 million.

Gartner analyst Max Goss told a London conference audience in April that the average Fortune 500 company will run more than 150,000 AI agents by 2028, according to SAP's coverage of the Gartner estimate. Yet only 13% of organizations believe they have the right governance in place to manage that. A separate SAP LeanIX survey found 98% of companies have deployed or plan to deploy AI agents, but fewer than half have any real inventory of the agents they've already built.

Goss's warning is blunt: block agents outright and employees just go around the controls with unauthorized "shadow AI," which he says creates bigger risks than the sanctioned tools ever did. That's a fair concern for IT leaders worried that clamping down too hard backfires. But the alternative KPMG's data describes, letting cost and permissions sprawl unchecked until finance finds an unexplainable bill, hasn't worked either.

Not a bubble

Carter's Forbes piece is explicit that this is not a bubble bursting. Prediction market Polymarket has priced roughly a 15% chance of an AI bubble bursting by year end, and that number circulated alongside the KPMG statistic on social media. But the KPMG data shows spending holding steady and AI's priority ranking rising, not collapsing. What's happening is a correction: companies overbuilt agent deployments in 2025 and early 2026 without cost metering or governance, and now finance departments are forcing a redesign.

The open question is whether "redesign" actually means fixing the metering and permissioning gaps VentureBeat, SAP, and WitnessAI all documented, or whether it just means quietly shrinking agent scope until the invoices look normal again. KPMG's own data doesn't say which. The next quarterly Global AI Pulse survey, expected in the fall, should show whether the 49% pullback figure keeps climbing or whether the governance fixes analysts are calling for actually take hold.