The AI honeymoon is ending
Just two years ago, companies raced to adopt artificial intelligence, convinced it would quickly transform their businesses and deliver dramatic financial returns. Corporate executives approved massive budgets, employees competed to use the most AI tokens, and technology companies encouraged the belief that bigger AI spending meant greater innovation.
Today, that mindset is changing.
Businesses are not abandoning AI, but many are abandoning the idea that spending more automatically produces better results. Instead, executives are becoming far more disciplined about where AI is deployed, which models they use, and whether the technology is producing measurable returns. The shift reflects an industry that is maturing after an initial period of excitement, where expectations often exceeded practical results.
The ROI problem
One of the biggest obstacles facing corporate AI adoption is proving a return on investment.
Research cited by Forrester shows that many enterprises are delaying roughly one quarter of their planned AI spending as they struggle to connect AI investments with improvements to their bottom lines. Only about 15 percent of AI decision makers surveyed reported earnings increases tied directly to AI, and fewer than one-third could link AI to meaningful revenue growth.
Brian Hopkins, vice president of emerging technology at Forrester, says many chief information officers tell him they see employees becoming more efficient but struggle to convert those time savings into measurable business results.
“Because firms are having a hard time moving from individual task kind of efficiency into more process efficiency, they’re not seeing any earnings at the top line,” Hopkins said.
Saving fifteen minutes when writing an email may improve productivity, but unless that efficiency changes an entire business process, it often does not appear in quarterly earnings reports. That disconnect has caused many executives to question whether some AI projects deserve additional funding.
MIT finds a staggering failure rate
Perhaps the most sobering assessment comes from researchers associated with MIT’s Project NANDA.
Their report found that approximately 95 percent of enterprise generative AI pilot programs fail to produce rapid revenue acceleration. Rather than blaming the technology itself, researchers concluded that the biggest obstacle is what they describe as a “learning gap” inside organizations. Companies frequently purchase powerful AI tools but fail to integrate them into existing workflows where they can create measurable value.
The report found that executives often blame regulations or limitations in AI models, but researchers concluded the larger problem is poor implementation.
MIT also found companies frequently spend their AI budgets in the wrong places. More than half of enterprise AI spending goes toward sales and marketing tools, while the strongest returns often come from automating back office operations, reducing outsourcing costs, and streamlining internal business processes.
The research also found that organizations purchasing specialized AI solutions and partnering with experienced vendors succeed far more often than companies attempting to build their own AI systems from scratch.
Companies are becoming smarter buyers
The current trend is not an AI retreat so much as a shift toward smarter spending.
Instead of relying exclusively on premium AI systems, companies increasingly mix expensive frontier models with lower cost alternatives. Simpler tasks are assigned to inexpensive models while the most advanced systems are reserved for work requiring deeper reasoning.
Mike Saeks of Cursor compares using the most powerful AI models for routine work to “driving a Lamborghini to go to the grocery store to pick up milk.”
Cursor demonstrated just how significant the savings can be. Building a web browser entirely with OpenAI’s GPT-5.5 cost more than $10,000. Using a combination of specialized models reduced the cost to about $1,339 while still completing the project.
Businesses have also become far less loyal to individual AI vendors. Instead, they increasingly choose whichever model offers the best balance between performance and price for a specific task.
Some AI vendors are responding aggressively. Companies are offering free tokens, heavily subsidized usage, discounted enterprise contracts, and lower cost models to retain customers. Others have embraced open weight models that allow businesses greater flexibility and lower operating costs.
Success stories still exist
Despite the growing caution, there are many examples showing AI can generate substantial value when deployed thoughtfully.
Zoom has used Meta’s open Llama models for years and says fine tuning those models has significantly reduced costs while allowing the company to combine multiple AI systems for different tasks.
Harvey, the legal AI company, routes difficult legal questions to advanced models while allowing simpler work to be completed by less expensive systems. The approach maintains quality while lowering operating expenses.
Telnyx adopted a similar strategy after discovering that continuing to rely on premium AI subscriptions could have cost roughly $100,000 per day. By shifting much of its workload to open models while keeping premium systems for planning and review, the company dramatically reduced costs without abandoning AI altogether.
These examples suggest AI succeeds best when organizations carefully match the right tool to the right job rather than assuming one model should perform every task.
Not every benefit appears on a balance sheet
Some economists argue companies may actually be creating far more value than current financial statements reveal.
Researchers at the Peterson Institute for International Economics estimate AI could already be generating as much as $250 billion in economic activity that traditional productivity statistics simply fail to measure. Individual workers may be completing tasks faster and producing higher quality work even though those gains have not yet translated into the accounting metrics executives typically use to justify investment.
If that assessment proves correct, today’s disappointing ROI numbers may partly reflect the difficulty of measuring a technology whose benefits are still spreading through organizations.
An industry growing up
The current slowdown in AI spending should not be mistaken for a collapse in artificial intelligence.
Instead, corporations appear to be entering a more mature phase. The initial excitement encouraged companies to launch pilots almost everywhere, often without clear business objectives or measurable success criteria. That experimentation produced valuable lessons, but it also produced expensive disappointments.
Now executives are demanding proof before approving additional spending. They are favoring targeted applications over broad deployments, combining multiple AI models instead of relying on one provider, and focusing on projects with measurable business outcomes.
The technology itself continues to improve rapidly. The challenge is no longer whether AI works. It is whether companies have developed the experience, organizational knowledge, and business processes needed to convert AI’s remarkable technical capabilities into consistent financial returns.
Like many transformative technologies before it, artificial intelligence may ultimately deliver enormous economic value. The lesson emerging from corporate America is that the path from breakthrough technology to profitable business practice is rarely immediate. It requires experimentation, refinement, and patience. For many companies, that maturity is only now beginning to arrive.
