Two stories from the past few weeks capture important points about the current state of AI.The first concerns enterprise software giant Salesforce, which has aggressively adopted AI for customer service. CEO Marc Benioff proudly announced that the company was able to reduce its support staff from 9,000 to about 5,000 thanks to AI. Then reality intervened. According to reports in late 2025, the company is currently exiting AI due to widespread failures. The AI agent confidently gave incorrect answers, skipped instructions after eight steps, and lost focus when users asked unexpected questions. Customers complained that AI support was more time-consuming than traditional simple search functionality. Salesforce has now retreated to strict rules-based scripting, essentially admitting that, in their own words, they “had more confidence” in the technology than was guaranteed.The second story is a shift in the zeitgeist. Over the past few months, the conversation around AI and coding has completely changed. People who were skeptical six months ago, senior developers who actually write code for a living, are now saying that the days of humans writing code are numbered. Not in the distant future, but in the near future. All functionality is shipped by AI with minimal human intervention. Productivity gains are no longer gradual. They are structural.How can both be true? How does AI comprehensively fail at seemingly simple customer service while revolutionizing software development, which seems much more complex?The answer is because we’ve been thinking about AI wrong. We treat this as a single phenomenon that spreads across the economy at roughly the same pace. However, AI in business is not a single story. It’s a lot of parallel stories, moving at very different speeds. And the difference has little to do with the level of intelligence of the AI.I’ve written about this tension before. A year ago, I argued that “the fact that the revolution is real does not mean that every company that claims to be a part of it will succeed.” Just recently, I realized that there remains a huge gap between what AI can successfully demonstrate in a controlled environment and what it can actually deliver when faced with a chaotic real world. We now believe there is a way to understand this gap more precisely. It’s not random. It’s a structural thing.Let’s consider what makes coding fertile ground for AI. The code is formally structured and machine verifiable. That is, the code either runs and passes the test, or it doesn’t. The feedback loop is immediate. When the AI makes a mistake, the developer (or another AI agent) notices it, fixes it, and moves on. Errors are private and reversible. Next, think about customer service. Customers don’t talk about data schemas. Emotion, irony, and cultural context are very important. One wrong answer can lead to outrage on social media and regulatory complaints. Failures are public and often irreversible.The difference is not intelligence. That’s what I call error economics. AI thrives when mistakes are cheap, private, and fixable. We struggle when mistakes are costly, public, and persistent.Just a few days ago, we received a clear example of leadership disconnect. During Bajaj Finance’s Q3 conference call, CEO Rajeev Jain announced that AI listened to 2 billion calls and generated 100,000 new customer offers. “Next year we will be able to listen to 100 million calls,” he said proudly. The reaction on social media was predictably amusing. As the whole country knows, except apparently Mr. Jain, Bajaj Finance’s incessant spam calls are the butt of countless jokes. Here we had a CEO using sophisticated technology to optimize for things that customers actively dislike. Machine learning works perfectly. Lack of learning about customers.For investors, the implications are significant. When you hear “AI” being associated with a business function, ask yourself, “What if the AI is wrong?” If the answer involves customers, regulators, or reputation, progress will be slower than the vendor’s PPT claims. If the answer is “someone will notice and fix it” then that’s a whole different world.The story of AI in business is not one of universal acceleration. This is a selective escape velocity. Coding left the atmosphere and entered orbit. Customer service is still fighting gravity. Most other features fall somewhere in between and are mistakenly assumed to be closer to rockets than they actually are. The different worlds of AI are not converging. They are branched. And that divergence determines which investments succeed and which ones disappoint.(Dhirendra Kumar is Founder and CEO of Value Research)
