Companies cut AI spending as low-quality work piles up

The output problem

Microsoft executive vice president Ryan Roslansky has described AI-generated work documents as part of a broader flood of low-quality material. He recounted receiving a document that was clearly written by AI, asking AI to summarize it, and finding no new ideas in the result. He called it a closed loop.

That example points to a distinction companies may have underestimated: producing more text is not the same as producing more useful work. AI can make an employee look busy while shifting the cost of that appearance to the company.

SHRM chief executive Johnny Taylor Jr. told The Guardian that business leaders can feel as if they are driving a truck while building it at the same time. He said that, given the chance to start over, companies should take a more considered, gradual approach to AI adoption.

The concern is not limited to bad writing. A September review by Ernst and Young found that 36% of AI leaders at companies with annual revenue above $1 billion had already experienced a catastrophic AI-related incident, involving financial damage, operational disruption, data loss or reputational harm.

The harder case for adoption

Consumer AI agents may dominate headlines with alarming behavior, but practical use inside companies remains limited. Low-quality content may be the most common workplace consequence so far; the incidents reported by Ernst and Young show that the costs can be more serious.

I think the pullback is less about companies abandoning AI than about a gap between spending and measurable value. That gap is especially awkward for Microsoft: its own success depends in part on persuading other businesses to buy into AI’s promise. If the first visible return is more work that needs to be checked, summarized and discarded, enthusiasm alone will not sustain the budgets.

What remains unclear is how much of the problem comes from the models and how much from the way companies have introduced them. Until businesses can show that AI reduces meaningful work rather than multiplying low-value output, a slower rollout may look less like hesitation than basic cost control.