A nice piece In Fierce Networks talks about the fact that while enterprise AI use is growing, more and more enterprises are finding AI cost overruns, not only versus expectations but versus any plausible return on the investment. I’ve dug through my enterprise comments over the last two years to get an idea of what the problems are, in the opinion of enterprises themselves. As you’ll see, many of their issues are related, so I’ll list the points in order of mentions, and note the relationships. The question, in the end, is whether AI is following the same path of hype-driven overuse and repatriation as the cloud has.
The number one problem I hear about is citizen AI. The form of AI whose “adoption is rising” is the cloud-hosted form, which is either fully usage priced or priced based on base-plus-usage, and this form can usually be adopted with only minimal management approval and rarely requires any IT oversight. My comments come largely (2:1 roughly) from IT professionals, but both line and IT types agree that there is really no supervision of AI usage by the approving authority, until costs become conspicuous. Workers are happy to have the company spend on things that make their jobs easier, and their management will often simply agree. Enterprises agree that this form of AI inevitably leads to misuse, meaning spending that doesn’t produce a valid benefit. How much of this goes on is hard to say because no enterprises indicated there was anyone in their company who even knew just who was spending on citizen AI, and how much.
The second problem, by a small margin, is related to this. AI is an expensed service. How many times did we hear that “repatriation” of cloud applications was necessary because of cloud cost overruns? That’s already happening with AI, as more and more workers exceed any base usage levels on their plans and end up needing tokens. However, with cloud computing, the stuff that was a cloud expense used to be a data-center capital project, so bringing it back was a consideration. With AI, companies find that it’s often not even clear that the AI expense is justified, and the path to running it in house to improve costs is rarely considered. Instead, the applications are simply abandoned, or scaled back by budget controls.
Problem number three is data access and sovereignty. Making a data center connection for an AI service will normally require some form of IT coordination, but in some cases workers may have access to raw data in a generalized form, like a spread sheet. Some AI models will even accept PDFs, which means that “printed” reports are available. Workers are encouraged by AI stories, and sometimes by their AI service providers, to expand the data sources, and this can run up costs and also end up giving an AI model proprietary data that governance policies say can’t even be used or hosted in the cloud. Where IT is involved in data access, enterprises find that there’s not enough control exercised over the data connection (RAG, MCP, or whatever) to manage cost and compliance.
Problem four is AI literacy. There are five AI providers typically recognized by enterprise AI users (Amazon, Google, Microsoft, OpenAI and Anthropic). All of them offer multiple models and tools, and all these are evolving over time. In addition, the AI plans available, including tools, costs, features, etc. are evolving. About a third of enterprises say that their line AI users are often switching between AI providers and services based on what they see online, or hear from the providers. As a result, it’s harder to develop a basis for confident AI planning, and often necessary for a user to retrain themselves and adapt their application of AI to new tools. This makes AI less effective.
Problem five is advanced AI features are more often expensive, and wrong. Enterprises say that “everyone who uses AI has found errors.” They also say that the more sophisticated the AI tool, the greater the chance it will make a mistake, that the worker(s) won’t catch it, and that it will generate a cost problem. Of course, everything said or written about AI encourages users to employ the most advanced tools, features, and models. One enterprise noted that generating a presentation was more than twice as likely to produce a problem with the data than generating a report, that generating a video was perhaps three times as likely to introduce errors as a presentation, and that the entire base usage quota for an AI plan would likely be consumed by creating a video less than a minute in length.
The final problem is AI encourages classic “theft of time” and also wastes company money on personal usage. Theft of time is a personal activity that’s undertaken on company time, and since AI tools can’t discriminate between a desire to create a product image and a personal one, almost all enterprises say that some (well, most) workers will use AI tools at work for personal reasons. Most enterprises say that they believe that some workers run up significant usage costs this way, too. Enterprise IT has recognized this from the first, and IT projects using AI often embed the AI within another application or in a workflow, rather than expose a chat/request type of interface.
You can see in the article I referenced that it’s based on chat-interfaced AI services, so the presumed AI of the future is that form of AI. This is most definitely not how IT organizations see it, but it’s surely how the AI giants are hoping things will go. However, these issues are all weighing on AI just as many weighted on cloud usage. Enterprises really do believe that the cloud providers hope is that AI will be a back door into a new “everything moves to the cloud” paradigm, that AI adoption will turn enterprises away from self-hosting.
Enterprises I chat with don’t see this happening, and neither do I, but what I think may be true is that it will take a long time for enterprises to come to terms with the “real AI”, and that means the AI hype wave likely has more runway available than cynical me might offer it.
