AI Is Everywhere Except the Balance Sheet

A developer argues that most companies are investing in AI without seeing financial returns, citing Stanford's AI Index showing only 39% of organizations report profit or cost savings from AI. The post emphasizes the need to tie AI projects to clear business outcomes and warns that Gartner expects over 40% of agentic AI projects to be shelved due to unclear ROI.

Originally published on lavkesh.com Most companies today will tell you they are doing something with artificial intelligence, that it is woven into their operations, and that it is changing everything. They talk about agents, about models, about the way data flows. This is the common wisdom, the thing you hear at every conference. But fewer than four in ten of these same organizations can point to any actual profit or cost savings from their AI efforts. The Stanford AI Index for 2026 put the number at 39 percent, which means a large majority are spending money on AI without seeing it reflected on the balance sheet. This disconnection feels familiar, like watching a team celebrate the deployment of a new service to production, calling that the win. The real win, of course, is what that service does for the business, how it solves a customer problem, or how it reduces a specific operational cost. The same pattern played out with cloud migrations, where companies moved everything to a new data center provider and declared victory, only to find their costs increased and their reliability stayed flat. The enthusiasm for new technology often overshadows the hard work of defining clear, measurable outcomes. When I was in energy management, we had a system that could predict equipment failures. The initial excitement was around the prediction accuracy, how many false positives versus true positives. But the real value came when we could show that acting on those predictions reduced unplanned downtime by a specific percentage, saving millions in lost production and repair costs. That required integrating the AI output into the maintenance schedule, training technicians, and tracking the financial impact of each avoided failure. This is where many AI projects falter. The engineers build a clever model, the product team finds a place for it, and the leadership announces the adoption. Everyone feels good. But then the project budget balloons, the model drifts, and the operational costs of maintaining it start to eat into any theoretical gains. Gartner expects more than 40% of agentic AI projects will be shelved by next year because the return on investment is unclear and the costs are too high. It is not enough to say the AI is running; you must also say what it is earning or saving. The problem starts when the conversation about AI is mostly technical, or aspirational. We talk about what AI can do, rather than what this specific AI is actually doing for this specific line item . The focus shifts to adoption metrics - how many users, how many models, how many inferences per second - instead of business metrics like reduced customer churn, faster transaction processing, or increased sales conversion rates that can be attributed directly to the AI's influence. Think about the data pipelines required to feed these models, the constant retraining, the monitoring for bias or drift. Each of these steps carries a cost, both in terms of infrastructure and human effort. If your AI is automating a task that costs a hundred dollars a month in human time, but the AI itself costs two hundred dollars a month to run and maintain, you are not winning. This seems obvious, but I have seen teams so focused on the 'coolness' of the tech that they ignore the basic arithmetic. Establishing a clear line of sight from an AI feature to a profit and loss statement requires collaboration between engineering, product, and finance teams, a language spoken by all of them. It means defining success not just in terms of model accuracy, but in dollars and cents. You need to know what problem the AI is solving, what it costs to solve it with AI, and what the alternative cost would be. Without this, you are just throwing money at a promising idea. Perhaps the biggest challenge is that measuring true bottom-line impact is hard, harder than counting how many models are deployed or how many data points are processed. It means asking tough questions, sometimes years after the initial investment, and being willing to admit when something is not working. It means treating AI not as a magic bullet, but as another tool in the engineering toolbox, one that must justify its existence like any other piece of software.