Here is an uncomfortable idea for anyone who has watched enterprise AI bills climb this year: the prices are still too low. In a Harvard Business Review analysis this month, RedThread Research analyst Stacia Garr argues that major software vendors are quietly absorbing the real cost of the GPUs, inference computing and tokens behind their AI features to accelerate adoption. Premium AI agents get marketed as "unmetered," "complimentary" or "included." The effect, Garr writes, is that "organizations are effectively being paid to adopt AI," building workflows on pricing that does not reflect what the underlying compute actually costs.
That sets up a trap that springs when the subsidies end. Companies that wired AI agents into daily operations at today's flattering rates are exposed to usage-based billing that scales in ways a software licence never did, especially for agents that run long, multi-step tasks and burn tokens far faster than a simple chatbot query. Garr's sharper point is that this is not a software problem at all. As firms swap fixed labour costs for variable AI consumption, she argues, they need to stop treating AI as a purchase and start treating it as an organizational design challenge, with new approaches to budgeting, workforce planning and risk.
The awkward part is that most buyers cannot yet say whether the spending is working. IBM figures cited by InfoWorld suggest only 29 percent of executives can confidently measure AI's return, and just 25 percent of AI initiatives deliver the ROI expected of them. Kyndryl found 61 percent of senior leaders feel more pressure to prove that return than they did a year ago. The result is a strange combination: budgets ballooning, subsidies masking the true bill, and leaders unable to point to what any of it bought.
A small industry is forming to answer that question. Tempo, a maker of Atlassian tools, has launched Workforce Intelligence, which ties AI cost directly to individual Jira work items and sets it beside the human labour cost on the same record. "The amount of money people are spending on AI is enormous, and a very large percentage of it is wasted," chief executive Vic Chynoweth said, describing customers stuck between "I've got to move faster and deploy AI" and "I need to moderate my AI spend." His complaint about existing metrics is pointed: knowing that 100 percent of staff have "touched an LLM," or counting how much code they generated, tells you almost nothing, because AI has commoditized code that "may or may not have been useful."
That reframing, from consumption to outcomes, was the common note across the week. At AI TechWorld, Kognitos chief executive Binny Gill put it bluntly: "If we cannot show where the value comes from, we do not have an AI strategy. We have an AI bill." Salesforce's Marc Benioff, insisting enterprise adoption is still just beginning, argued the eventual test will not be chips or data centres but "what value are you delivering to the customer." And CompTIA research found 59 percent of organizations now prioritize integrating AI into existing systems rather than simply handing staff more tools, a sign the experimentation phase is giving way to a harder accounting.
The through-line is that 2026 is turning into a reckoning that the vendors' own generosity has delayed. Cheap tokens made it easy to deploy everywhere and ask questions later. As the true cost surfaces, the companies that thrive will not be the ones using the most AI, but the ones that can prove, line by line, that a given agent earned more than it consumed. The bill is coming due. The unsettling news from Harvard Business Review is that it will be larger than the invoices so far suggest.