From Hype to Reality: When AI Underperforms Humans
The latest wave of AI project failures is not about exotic use cases but basic operational tasks. Starbucks offers a vivid example. After nine months, the company scrapped an AI “Automatic Counting” tool, developed with NomadGo, that was supposed to automate inventory tracking for items like milk and syrups. In practice, the system repeatedly miscounted stock, mislabeled similar milk types, and even skipped items entirely. An internal video introducing the tool reportedly captured it missing a syrup bottle in plain view—an embarrassing symbol of the gap between promise and delivery. Starbucks has now told staff to return to manual inventory checks, conceding that human workers currently count better than the AI. The episode highlights a growing corporate AI pullback: when automation cannot reliably beat simple human routines, the efficiency narrative collapses, and the extra enterprise AI costs become difficult to justify.
Rising Enterprise AI Costs and the Uber, Microsoft, Klarna Reassessment
As AI use scales, enterprise AI costs are exploding faster than measurable benefits. Uber’s deployment of Anthropic’s Claude Code to around 5,000 engineers showed impressive usage: most engineers touched AI tools monthly, and a majority of code commits involved AI assistance. Yet COO Andrew Macdonald publicly admitted that linking these statistics to clear customer value or productivity improvements remains elusive. Microsoft, despite its AI-first narrative and investment in OpenAI, quietly began revoking internal Claude Code licenses, pushing engineers toward GitHub Copilot CLI instead—an implicit acknowledgement of AI ROI problems at scale. Klarna, once a loud proponent of replacing staff with an OpenAI-powered chatbot, saw customer satisfaction plunge, with responses becoming generic and failing on complex queries. The company has since rehired human agents, with its CEO conceding that chasing efficiency alone led to unsustainable quality trade-offs and undercut the business case.
Service Quality Backlash: When AI Hurts Customer Experience
Customer-facing AI implementations have become a flashpoint for frustration as companies discover that cheap automation can be very expensive reputationally. Klarna’s chatbot, at one point handling up to three-quarters of customer interactions, could not manage nuanced issues, driving down satisfaction by 22% and forcing a partial return to human support. A major bank saw a similar pattern after replacing dozens of call-centre staff with an AI voice bot, expecting call volumes to fall. Instead, calls and queue times surged, and managers had to get back on the phone. Within weeks, the bank reversed redundancies and apologized, acknowledging it misjudged how long elevated call volumes would persist. These cases show how AI project failures often stem from treating service quality as secondary to cost-cutting. When automated systems generate more friction, they erode brand trust and wipe out any theoretical savings on staffing or tooling.
The Tokenmaxxing Backlash and the AI ROI Reckoning
Inside Silicon Valley, a growing debate over “tokenmaxxing” captures the new skepticism around AI ROI. Tokenmaxxing refers to driving up AI token usage—often as a badge of innovation—without proving corresponding business value. Uber’s Macdonald voiced this tension directly, noting that even dramatic increases in AI usage do not yet translate into clearly higher output of useful features. Meanwhile, big firms are aggressively tracking AI adoption: some label employees as “AI builders,” organize AI-native teams, and even boast about staggering monthly token consumption. Critics argue a large share of this internal spend is wasted, with millions of tokens burned and little to show. Even Google’s CEO has warned that CIOs are increasingly alarmed by runaway AI budgets. The emerging consensus is that indiscriminate AI expansion is over; future investments will need to demonstrate tangible gains rather than simply bigger models or higher token counts.
Toward More Cautious, Value-Driven AI Adoption
The pullback does not mean companies are abandoning AI altogether, but expectations are being reset. Duolingo’s experience is emblematic: after loudly declaring itself “AI-first” and initially tying performance reviews to AI usage, leadership later dropped that metric when staff complained they were being pushed to use AI for its own sake. The message now is pragmatic—use AI where it genuinely improves your work, ignore it where it does not. Across industries, the pattern is similar. Early enthusiasm produced rushed deployments, ballooning enterprise AI costs, and over-optimistic assumptions about productivity. Now, firms are redesigning AI governance with stricter controls on budget, tighter measurement of outcomes, and a renewed willingness to admit when humans simply do the job better. The next phase of AI adoption will likely prioritize smaller, focused deployments with clear benchmarks rather than headline-grabbing, company-wide transformations.