We've come across some interesting news about how the AI industry is moving forward. If you haven't started yet, it definitely feels like this space slows down. (Not an economist, I'm not calling it a bubble, but there are a lot of opinions out there.) The GPT-5 came out last month and disappointed everyone. Meta has created a very sudden pivot, reorganising the entire AI function and halting all employment right after seducing unlimited funds into space adoption and pleading. Microsoft appears to be slowing down its investment in AI hardware (PayWall).
This does not mean, of course, that one of the major players will stop investing in AI. The technology isn't approaching anything like the grand results and even remotely, like AGI, which many analysts and writers (including me) predicted would not, but there is an incentive to continue moving forward as there is a level of utilization that is still going on among businesses and individuals.
5% success rate
In this trend, I read a new MIT report on AI in business this week. I would recommend it to anyone looking for real information about how AI adoption is moving from regular workers and C-Suite. The report has several headline takeaways, including claims that only 5% of AI initiatives in business settings create meaningful value. (And also, AI doesn't actually take on people's jobs in most industries, and in some industries AI hasn't affected at all.) It appears that many businesses have adopted AI without a strategic plan for what it is supposed to be, and that adoption would actually help them achieve their goals.
In fact, this sees a lot. Executives who are heavily separated from the daily work of organizations that are held by FOMO about AI, decide that AI must become part of the business, but do not consider how this fits into the business that already has.
Driver or Magic Wand?
Regular readers will know that, of course, they do not insist that AI cannot or should not be used if it serves their purposes. And it's far away! I build AI-based solutions to business problems in my organization every day. But I firmly believe that AI is a tool, not a magic. It provides a way to perform tasks that are unfeasible for human workers, and can accelerate the speed of tasks that otherwise have to be done manually. It helps you to make information clearer and better understand long documents and texts.
But what it doesn't do is make the business a success on its own. To be part of 5% rather than 95%, the application of AI must be established based on strategic thinking and planning, and most importantly, there must be clear expectations of what AI can and can do. Small projects that improve certain processes, despite being less attractive and headline producers than hype, can bring about great returns without betting on a massive turbulence or “revolution” of your business. The MIT Report explains how a huge number of projects begin as pilots and experiments, but it never actually happens in production. I argue that much of this is because there was no plan or under-eye expectations.
The author spends considerable time noting that many AI tools are considered inflexible and/or incompatible with existing processes and fail to adopt between ranks and files. When you build or buy AI solutions that don't work in the business that exists today, you're throwing away your money. Solutions should be designed with business in mind and do not imply a failure in strategic planning, or are not flexible or compatible in the way necessary.
Security transactions for versatility
I had more thoughts on the subject of flexibility when I was reading. The authors of MIT emphasize that the internal tools companies provide to their teams often “do not work” in some way, but in reality, many of the stiffness and limitations imposed on internal LLM tools are for safety and risk prevention. Developers don't build tools that don't work intentionally, but there are limitations and requirements to follow. In short, there is a trade-off here. If LLM is very open and there are little or no guardrails, either allow users to do more or answer more questions. But it can be quite costly, potentially liable, providing false or inappropriate information or even worse, doing it.
Of course, normal users may not be thinking about this angle when lifting up the ChatGpt app on their mobile phone with their personal accounts during work days. The Infosec community is rightly wary of this kind, with some circles calling it “Shadow Ai” instead of Shadow It. The risks from this behavior are catastrophic – it is possible that your own company data can be freely handed over to AI solutions without surveillance. This problem is really, really difficult to solve. Employee education is a clear step at all levels of the organization, but this shadowy degree of AI is likely to last, and security teams struggle with this as we speak.
Conclusion
I think this leaves us in an interesting moment. I think the winner of the AI rat race will be someone who is thoughtful and careful, who applies AI solutions sparingly and does not try to raise the success model up to date in order to pursue new shiny ones. A slow, stable approach can help hedge against risks, such as customer backlash against AI and many others.
Before closing, I would like to remind everyone that when the condos have concrete results, these attempts will try to build a palace equivalent. Elon Musk knows that by running illegal gas generator-powered data centers, it is polluting Memphis suburbs with immunity. Data centers make up a double-digit percentage of all electricity generated in some US states. Water supply is exhausted or contaminated by these same data centers that provide AI applications to users. Remember that the choices we make are not abstract and are conscientious about when and why we use AI. 95% of failed AI projects were more than just expensive in terms of the time and money spent by the company.
Read more about my work at www.stephaniekirmer.com.
Read more
https://garymarcus.substack.com/p/gpt-5-overdue-overhyped and-under-under-under-durwher
https://fortune.com/2025/08/18/sam-altman-openai-chatgpt5-launch-data-centers-investments
https://www.theinformation.com/articles/microsoft-scales-back-ambitions-ai-chips-overcome-delays
https://builtin.com/artificial-intelligence/meta-superintelligence-reorg
https://mlq.ai/media/quarterly_decks/v0.1_state_of_ai_in_business_2025_report.pdf
https://www.ibm.com/think/topics/shadow-ai
https://futurism.com/elon-musk-memphis-ilegal-generators
https://www.visualcapitalist.com/mapped-data-center-electicity-consumption-by-state
https://chicago.suntimes.com/environment/2025/08/20/data-centers-ai-artificial-intelligence-chicago-illinois-great-lakes-michigan-drink-jb-pritzker
https://www.esi.org/articles/view/data-centers-and-water-sumion
