Gartner warns semantics waste AI spend

AI agents are starting to look less like chatbots and more like junior operators inside companies. They pull data, make decisions, trigger workflows, and answer questions that sound simple. But Gartner’s warning this week is that many of these systems are being dropped into businesses that still haven’t agreed on what their own data means. That sounds abstract. It isn’t. It’s a fast way to get expensive wrong answers. ### What is Gartner actually warning about? The core point is simple: an AI agent can only reason with the context it has. If a company has messy or inconsistent definitions for business terms, the agent does not magically repair that mess. It amplifies it. Gartner said neglecting semantics makes agents inaccurate and inefficient, which means higher costs, more governance risk, and more chances that an automated workflow quietly does the wrong thing. (gartner.com) ### What does “semantics” mean here? Basically, semantics is the layer that tells a system what a thing actually is in business terms. Not just that a field is called “revenue,” but whether that means gross revenue or net revenue. Not just that a record says “customer,” but whether that means buyer, bill-to account, ship-to location, or parent company. The same goes for hierarchies, time periods, and metric grain. An agent can read the labels. (gartner.com) The problem is that labels are often not enough. ### Why do agents stumble on this faster than older software? Older analytics tools usually waited for a human to catch the mismatch. Agents are different because they chain steps together. One bad assumption early in a workflow can ripple through retrieval, reasoning, and action. If an agent pulls “customer revenue” from one system, combines it with “channel performance” from another, and both use different definitions, the output can still sound polished. (gartner.com) That is the dangerous part — fluent nonsense looks operationally ready. ### What kinds of mistakes is Gartner talking about? The examples are the boring-enterprise kind — which is exactly why they matter. Revenue may be defined differently across finance and sales. A weekly metric may get compared with a monthly one. Customer and channel hierarchies may not line up across systems. None of those issues are dramatic on their own. But an agent asked to optimize pricing, summarize performance, or route decisions across departments can get tripped by any one of them. (gartner.com) ### Why does this turn into wasted AI spend? Because companies end up paying for more than the model. They pay for tokens, orchestration, observability, integration work, and people cleaning up mistakes. If the business context underneath is shaky, the agent needs more correction loops and more guardrails, or it just produces low-trust output that employees stop using. Either way, the spend is real and the value leaks out. Gartner framed this as both an accuracy problem and an efficiency problem. (gartner.com) ### So what is the fix? Gartner’s advice is not “buy a smarter model.” It is to define business entities, metrics, and relationships before scaling agents. In plain English: decide what your company means by customer, product, revenue, region, and time period — then make those meanings explicit in the data layer agents use. That semantic foundation becomes the map the agent reasons over. Without it, you are asking software to navigate a city where every street has two names. (gartner.com) ### Why is this landing now? Because agent adoption is accelerating faster than governance. Gartner said in late April that the average Fortune 500 enterprise could go from fewer than 15 agents in 2025 to more than 150,000 by 2028, while only 13% of organizations think they have the right AI agent governance in place. That is the backdrop for this warning. The industry is racing to deploy agents. The plumbing underneath is not keeping up. (gartner.com) ### Bottom line? The news here is not that AI agents hallucinate. Everyone knows that. The sharper point is that even non-hallucinated answers can be wrong when a business has not defined its own terms. In enterprise AI, semantics is not polish — it is infrastructure. (gartner.com 1) (gartner.com 2)