AI Disinvestment: When the Hype Cycle Starts Sending Invoices
It feels like the wheels are starting to come off the AI bus.
Do not get me wrong. AI is useful. It can help summarize documents, draft content, analyze data, support developers, speed up research, and reduce repetitive work. Used well, it can be a real force multiplier.
But that is very different from the story many companies were sold.
The pitch was not just, “AI can help your team work faster.” The pitch was closer to, “AI will transform everything, replace huge parts of your workforce, and create massive returns almost immediately.”
That version of the story is starting to look shaky.
Are We In An AI Bubble?
Maybe.
There is no doubt that AI will keep improving. The tools are already impressive, and the next few years will almost certainly bring better models, better integrations, and more practical business use cases.
But the current AI landscape often feels less like a refined business transformation and more like a brute-force effort.
More GPUs. Bigger data centers. Larger models. Higher token spend. More automation. More agents. More promises.
At some point, businesses have to ask a boring but important question:
Is this actually producing value?
That question is where the AI disinvestment conversation starts.
AI disinvestment does not mean abandoning AI. It means pulling back from wasteful, poorly planned, or low-value AI projects and redirecting money toward initiatives that actually support the business.
The Pilot Problem
A widely discussed MIT report found that 95% of enterprise generative AI pilots were failing to produce meaningful results. The key issue was not necessarily that the models were bad. The bigger problem was that companies struggled to integrate AI into real workflows in a way that changed productivity, revenue, or operations.
Source: Fortune: MIT report says 95% of generative AI pilots are failing
That should sound familiar.
A lot of AI projects start like this:
- An executive hears that competitors are “doing AI.”
- A team is told to find a use case.
- A pilot is launched.
- People use it for a few demos.
- The tool never becomes part of the actual business process.
- Everyone quietly moves on.
That is not transformation. That is experimentation without a strategy.
There is nothing wrong with experimenting. The problem is pretending every experiment is a business case.
The Productivity Paradox Is Back
Fortune also reported that thousands of CEOs said AI had no measurable impact on employment or productivity, raising comparisons to the old productivity paradox from the computer era.
Source: Fortune: CEOs say AI has had no impact on employment or productivity
That does not mean AI is useless.
It means productivity gains are not automatic.
Giving employees access to AI tools does not magically redesign business processes. It does not clean up messy data. It does not fix unclear ownership. It does not replace governance. It does not solve bad strategy.
AI can accelerate a good process.
It can also accelerate a bad one.
Replacing People With AI Is Not A Strategy
One of the most expensive mistakes companies can make is assuming AI is a direct headcount replacement.
Forbes reported that some companies that cut workers because of AI are now bringing people back.
Source: Forbes: Companies Fired Workers For AI. Now They Want Them Back
That should not be surprising.
People do more than complete tasks. They understand context. They notice when something feels wrong. They deal with exceptions. They understand customers, vendors, politics, process gaps, and institutional knowledge.
AI may reduce some repetitive work, but replacing experienced employees before the system is proven is a risky bet.
The better strategy is usually human amplification:
- Help existing teams move faster
- Reduce repetitive administrative work
- Improve first drafts and analysis
- Support decision-making
- Keep humans accountable for judgment, risk, and outcomes
The goal should not be “how many people can we replace?”
The better question is:
How do we make our best people more effective?
Operational Reality Still Matters
The difference between an AI demo and a working business system is massive.
Starbucks reportedly scrapped an AI inventory tool across North America after the system struggled with inaccurate counts and product identification issues.
Source: Reuters: Starbucks scraps AI inventory tool across North America
Pizza Hut is also facing a $100 million lawsuit from a major franchisee over an AI-powered delivery management system. The franchisee alleges that delivery times worsened, customer satisfaction dropped, and sales declined after the system changed how delivery coordination worked.
Source: Tom’s Hardware: Pizza Hut’s AI delivery system cooks up $100 million franchisee lawsuit
These are not abstract problems.
They are real-world examples of AI systems colliding with messy operations.
The AI may work in the lab. The workflow may look good in a slide deck. The vendor demo may be impressive.
But then the system hits real stores, real drivers, real customers, real inventory, real incentives, and real edge cases.
That is where many AI projects fail.
The Cost Problem Is Getting Harder To Ignore
AI is not free.
Advanced models can be expensive to run, especially when companies start using agentic workflows that make multiple calls, process large context windows, or run continuously in the background.
The cost problem is not limited to software subscriptions. It also shows up in infrastructure.
xAI’s Colossus project describes itself as a massive AI training supercomputer.
Source: xAI: Colossus
At the same time, Wccftech reported that xAI was using only about 11% of its GPU capacity, citing reporting from The Information.
Source: Wccftech: xAI using just 11 percent of GPUs
That may be a temporary utilization issue, and AI infrastructure is complicated. But it highlights an important business lesson:
Buying a massive amount of capacity is not the same thing as turning it into productive value.
The same is true for small and mid-sized businesses.
Buying AI tools is easy.
Getting measurable value from them is harder.
AI Disinvestment Is Not Anti-AI
This is the part that gets lost in the hype cycle.
Being skeptical of AI spending does not mean being against AI.
It means being against waste.
It means asking normal business questions:
- What problem are we solving?
- Who owns the outcome?
- What process is changing?
- What data does the system need?
- What happens when the AI is wrong?
- What does this cost at full usage?
- What security and privacy risks are being introduced?
- How will we measure success?
- When do we shut it down if it does not work?
That last question is important.
Every AI initiative should have an exit ramp.
If the project does not reduce cost, improve quality, reduce risk, increase revenue, or meaningfully improve speed, then it should be reviewed. If it still cannot justify itself, it should be retired.
That is not failure. That is responsible governance.
The Security And Risk Side Cannot Be An Afterthought
AI projects often create new risks faster than companies can manage them.
Common issues include:
- Sensitive data being pasted into public tools
- Employees connecting AI tools to local files without oversight
- Shadow AI subscriptions outside IT visibility
- Unclear data retention and training policies
- Weak access controls around AI agents
- Unreviewed plugins, connectors, and MCP servers
- AI-generated code entering production without proper review
- Vendors making broad AI claims without clear security documentation
This is where business strategy, cybersecurity, legal, compliance, and operations need to work together.
AI should not be treated as just another software purchase. In many cases, it changes how data moves, how decisions are made, and who or what has access to sensitive business information.
That requires governance.
Not bureaucracy. Governance.
There is a difference.
A Better AI Strategy
A practical AI strategy should start small and stay tied to business outcomes.
A good approach looks more like this:
Pick a real business problem
Do not start with “we need AI.” Start with a pain point.Use AI to support people, not blindly replace them
Look for ways to make existing teams faster and more consistent.Measure the outcome
Track time saved, error reduction, customer impact, revenue impact, or risk reduction.Control the data
Know what information users can upload, where it goes, and how it is stored.Define acceptable use
Employees need clear guidance on what they can and cannot do with AI tools.Review vendors carefully
Understand their security posture, data handling, retention, access controls, and contractual terms.Watch the spend
Token usage, premium subscriptions, and agentic workflows can get expensive quickly.Shut down what does not work
A failed pilot should not become a permanent line item.
The Real Opportunity
AI is not going away.
But the easy money phase may be ending. The companies that win with AI will probably not be the ones that bought the most tools, launched the most pilots, or replaced people the fastest.
The winners will be the companies that slow down long enough to ask better questions.
Where can AI help our people?
Where can it reduce friction?
Where can it improve quality?
Where does it create unacceptable risk?
Where is the ROI real, and where are we just chasing the hype cycle?
AI can absolutely be valuable.
But value does not come from plugging in a chatbot and hoping magic falls out.
It comes from strategy, governance, process design, security, and clear business ownership.
That may not be as exciting as the sales pitch.
But it is a lot more likely to work.
Need Help Reviewing Your AI Strategy?
If your business is using AI, considering AI tools, or trying to figure out whether the investment is actually worth it, now is the time to step back and review the strategy.
MN Risk can help evaluate AI use cases, security risks, vendor claims, governance gaps, and whether your AI projects are aligned with real business outcomes.
AI should make your business stronger.
It should not become another expensive tool nobody can explain.