The Great AI Shakeout

The software industry is entering a “very violent technology cycle.” That is not dramatic phrasing from a tech blogger with a flair for the apocalyptic. It is the cold assessment of John Zito [1], co-president of Apollo Asset Management, a firm that oversees roughly $700 billion in assets. When someone managing that kind of money uses the word “violent,” it is worth putting down your coffee and paying attention.

We have all seen the headlines. Artificial intelligence is the new electricity. The new internet. The new fire. Companies are racing to buy it, integrate it, and assure their shareholders they are “powered by it.” But behind the press releases and the breathless keynote presentations, a quieter, far messier reality is unfolding. The companies selling this glorious future are disappearing at an alarming rate.

You might sign a contract with a promising AI partner on Tuesday and by Friday, they have been "acqui-hired" [2] by a tech giant, their product is being sunsetted, and you are left holding a digital paperweight. This is not bad luck. It is a structural feature of the current market. We are in the middle of a gold rush where the shovels break, and the shovel-sellers vanish into thin air.

The question is no longer whether AI will transform business. It will. The question is whether you can buy it without getting burned.

The Killing Fields

Let’s start with the big number. According to Ken Smythe, founder of Next Round Capital Partners, 85% of AI startups will be out of business within three years [3]. Let that sink in. If you walk into a room with ten AI vendors today, eight or nine of them will not be there to renew your contract in 2029.

The “AI Graveyard” is already getting crowded. As of mid-2024, one tracker listed nearly 1,700 defunct AI tools, which represents close to 30% of all the tools that particular aggregator monitored. In 2024 alone, more than 600 new “bodies” were added to the list [4]. At that pace, we are not witnessing a correction. We are witnessing a mass extinction.

Much of this churn is driven by what industry insiders call the “wrapper” phenomenon. A staggering number of AI startups are nothing more than thin interfaces built on top of foundation models from OpenAI, Google, or Anthropic. They do not own the intelligence.

They rent it. Their entire value proposition is a user-friendly front end sitting atop someone else’s engine. The moment OpenAI and co. ships a new feature that replicates what the startup does, the startup becomes obsolete overnight. It is a fragile existence, and it makes these vendors dangerous partners for any enterprise buyer who needs to plan beyond the next quarter.

And it is not just the companies that are failing. It is the projects themselves. A 2024 report from the RAND Corporation found that more than 80% of AI projects fail to reach production - twice the failure rate of traditional IT initiatives [5].

So even if you pick a vendor that survives, the odds that your project actually delivers value are roughly those of a coin flip. Worse, actually.

The New Rules for Buying AI

So we know the risks. The vendor might go bust. They might be faking it. They might get hollowed out by a tech giant’s acquisition team. But here is the uncomfortable truth: you still have to buy this stuff. Standing on the sidelines is not a strategy. It is a slow-motion way of becoming irrelevant.

The problem is that traditional procurement methods are catastrophically ill-suited for AI purchasing. You cannot simply check a box for “financial stability” when the company is two years old and burning venture capital to keep the lights on. You cannot evaluate AI the way you evaluate a CRM platform or a fleet of laptops. The entire playbook needs to be rewritten, starting with four challenges that most procurement teams are not yet equipped to handle.

Challenge One: Find the Signal in the Noise

First, you need to understand where thevalue actually lives. A widely cited MIT study reveals a striking divide: 95% of organizations investing in AI pilots are getting zero return [6]. Zero. On the other side of that chasm, the remaining 5% are generating enormous value. Same technology. Wildly different outcomes.

The difference is not about picking the fanciest model or the startup with the slickest demo. The winners are not just buying tools, but they are reimagining their workflows. If you are purchasing a chatbot to help your team write slightly better emails, you are almost certainly in the 95%. That is a productivity toy, not a business transformer. But if you are deploying an agent to automate your accounts payable process, eliminating manual steps and compressing cycle times, you might be in the 5%.

The implication for procurement is profound: before you evaluate a single vendor, you need to evaluate the use case. If the business case does not involve a fundamental rethinking of how work gets done, the AI is window dressing.

Challenge Two: The Iceberg of Hidden Costs

When you buy an AI solution, the price on the box is the tip of a very large iceberg. The real costs often lurk beneath the surface, and they can sink a project faster than any technical failure.

Start with inference costs, the price of actually running these models in production. IBM found that computing costs are expected to jump 89% between 2023 and 2025, driven largely by AI workloads [7]. If your contract does not account for this trajectory, your bill will explode in ways your CFO will find decidedly unpleasant.

Then there is maintenance. AI is not “set and forget.” Models drift. They degrade in accuracy as the world changes around them, like a map that slowly becomes wrong as new roads are built and old ones close. Maintaining them costs 15 to 30 percent of the initial build cost every single year. That is not a line item most procurement teams think to budget for.

And then there is the cost often neglected: change management. For every Swiss franc you spend on the technology, plan to spend three Swiss francs training people to use it. AI that nobody adopts is just an expensive server humming in a data center somewhere.

Challenge Three: The Data Problem Nobody Wants to Discuss

Here is an inconvenient truth: you cannot buy AI if you do not have the data to feed it. The RAND report flagged “lack of necessary data” as a top reason for project failure. You might purchase the most sophisticated tool on the market. But if your data is messy, siloed, or riddled with bias, the AI will fail. Garbage in, garbage out is not just a cliché in this context. It is a death sentence.

This transforms AI procurement into something more than a software evaluation exercise. You need to vet vendors not just on the elegance of their algorithms, but on their data handling capabilities. Can they help you clean your data? Do they offer pre-trained models that perform reliably with imperfect inputs? If a vendor tells you to “just hand over your data and we’ll figure it out,” consider that a bright red flag. Real AI implementation is a data engineering project first and a modeling project second. Any vendor who does not understand that order of operations is not ready for enterprise work.

Challenge Four: The Security Minefield

Security is perhaps the most insidious risk, because the threat often comes from inside the building. A Sophos report [8] highlights how “Shadow AI” is creating massive, largely invisible vulnerabilities. Employees across your organization are already using free AI tools to do their jobs, thereby pasting sensitive corporate data into public models, uploading proprietary documents to consumer-grade chatbots, running customer lists through unvetted applications. They are not being malicious. They are being resourceful. But the effect is the same.

When you source an AI solution, you are not just buying a tool. You are buying a way to bring that shadow usage into the light by giving people a sanctioned, secure alternative to the Wild West of free-tier AI.

If your vendor does not offer enterprise-grade security features like single sign-on, data residency controls, comprehensive audit logs, you are effectively inviting a data breach. In a regulated industry, that breach could cost more than the AI will ever save you.

The second threat is structural, and it persists even when you pick a reputable vendor and do everything by the book. Many AI providers, particularly those offering cloud-based models, reserve the right to use your data to retrain and improve their systems. Many bury it in terms of service that nobody reads and unlike a conventional data breach, which is a discrete event you can investigate and contain, data baked into a model’s training weights is essentially irrecoverable.

Buying Software When the Seller Might Vanish

So how do you actually do this? How do you buy software from companies that might not exist next year? You cannot just cross your fingers and hope for the best. You need a strategy, and the first and most important element of that strategy is deceptively simple: do not go it alone.

The AI market is too chaotic, too fast-moving, and too full of convincing imposters for any single procurement team to navigate by itself. You need a sourcing partner who genuinely understands the landscape and not a generalist consultant who applies the same framework they use for buying office chairs and cloud storage to a fundamentally different category of technology.

Generalists check the wrong boxes. They look for “financial stability” in companies that are two years old and burning venture capital. They see a big-name investor on the cap table and read it as a safety net. A specialist sees that same investor and reads it as an exit risk.

A good partner knows how to spot the fakes. They understand that a junior developer can build a visually impressive tool on top of GPT in a single weekend. They will ask the rude questions that polite procurement teams skip.

Most critically, a specialist helps you negotiate for the right things. Most procurement teams negotiate for price. In AI, price is secondary. What you need to negotiate for is exit. You need contractual terms that protect you, not the vendor. Basically a “prenup” that ensures you keep your data, your integrations, and your operational continuity when the relationship ends. Because in this market, it very well might.

The "Boring" Future We Need

The best possible outcome for AI is that it becomes boring. Not boring in the sense that it stops being powerful, but boring in the way that electricity is boring, so deeply woven into the fabric of business that nobody thinks to mention it. We stop talking about “AI projects” and start talking about “business projects” that happen to use AI. It becomes infrastructure, not spectacle.

But to get there, we have to survive the current chaos. We have to navigate the violent cycle, sidestep the falling startups, and build a foundation that can withstand the shakeout. The organizations that do this well will not merely survive. They will emerge with a massive competitive advantage, because they will have working, stable, value-generating AI while their competitors are still rebooting failed pilots, renegotiating with their third vendor in two years, and explaining to the board why the “transformative” investment has not transformed anything.

The AI revolution is real. The hype is also real. The trick is learning to tell the difference and buying accordingly.