Your competitor just announced they're 'transforming with AI.' Their stock ticked up. Their CEO is on podcasts. And you're sitting there wondering if your team's chatbot experiment—the one that's been 'almost ready' for eight months—will ever actually do anything useful.
You're not alone. I've watched this pattern repeat across dozens of businesses. The pilot works great in the demo. It works great in the test environment. Then it hits real data, real edge cases, real workflows—and it stalls. Or worse, it starts hallucinating answers that sound confident but are completely wrong.
A new research paper is making the rounds this week, and it's saying something most AI vendors really don't want you to hear: the math on AI agents might not add up. Hold that thought—I'll show you why this is actually good news for businesses that approach AI differently.
What Does the Research Actually Say?
A paper titled 'Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models' claims to mathematically prove that LLMs—the technology behind ChatGPT, Claude, and every AI assistant you've heard of—are incapable of carrying out computational and agentic tasks beyond a certain complexity, according to Wired's coverage.
This isn't some fringe claim. OpenAI's own scientists published research in September 2025 admitting that 'accuracy will never reach 100 percent' for AI models. They proved their point accidentally—when they asked three models to provide the title of the lead author's dissertation, all three made up fake titles.
Let that sink in. The people building these systems are telling you the systems will always make things up sometimes.
Why Are 60% of Companies Seeing Almost Nothing?
Here's where it gets interesting for business owners. According to BCG's latest research, only 5% of companies worldwide are 'AI future-built'—meaning they've figured out how to actually generate significant value from their AI investments. Another 35% are scaling AI and starting to see returns.
The remaining 60%? They're investing substantial money and seeing almost no material value.
That's not a technology problem. That's a strategy problem.
The numbers get worse when you look at enterprise deployments. Gartner's 2025 AI Deployment Survey found that almost 80% of enterprise AI projects don't make it past the pilot stage. Even among those that do, less than 40% work well in real-world use.
What Separates the 5% Who Are Winning?
This is the part that matters. Because while most companies are failing at AI, the companies that get it right are seeing results that change the competitive math entirely.
BCG's research shows that AI future-built companies achieve five times the revenue increases and three times the cost reductions compared to everyone else. Five times. That's not incremental improvement—that's a different business.
So what are they doing differently?
First, they're not treating AI as magic. Himanshu Tyagi, cofounder of AI company Sentient, told Wired: 'The value has not been delivered.' He points out that dealing with hallucinations can disrupt an entire workflow, negating much of the value of an agent. The companies succeeding understand this limitation and design around it.
Second, they're building systems around LLMs rather than relying on pure LLMs. As one researcher explained: 'A pure LLM has this inherent limitation—but at the same time it is true that you can build components around LLMs that overcome those limitations.' Companies like Harmonic are using formal mathematical reasoning to verify LLM output by encoding outputs in the Lean programming language—essentially having the AI check its own work.
Third, they're measuring AI by business outcomes, not technical metrics. Organizations that measure AI agents by metrics like lines of code or chat completions risk undervaluing—or overvaluing—their investments. The winners measure what actually matters: revenue generated, costs reduced, hours saved.
The Uncomfortable Truth Most Vendors Won't Tell You
Here's where it gets philosophically interesting—and practically important.
Hallucinations might not be a bug. They might be a feature.
One AI researcher quoted in Wired made this provocative claim: 'I think hallucinations are intrinsic to LLMs and also necessary for going beyond human intelligence. The way that systems learn is by hallucinating something. It's often wrong, but sometimes it's something that no human has ever thought before.'
That's cold comfort when your AI agent just sent a customer the wrong product specifications. But it explains why the 'just make it stop hallucinating' approach isn't working. You can't eliminate hallucinations without eliminating what makes these systems valuable in the first place.
The winning companies understand this. They use AI for tasks where creativity and pattern-matching matter, and they build verification layers for tasks where accuracy is critical. They don't ask AI to be what it can't be.
What This Means if You're in the 60%
If your AI projects have stalled, failed to deliver, or just feel stuck—you're not doing anything wrong. The approach most companies are using is designed to fail.
The problems aren't usually algorithmic limitations. According to research on AI agent failures, they come from weak databases, poor planning for growth, and not having enough monitoring after the system goes live. In short: the problems are about how systems are built and managed, not how smart the AI is.
BCG recommends what they call the 10/20/70 rule for AI success: 10% technology, 20% data, 70% people and processes. Most companies have those percentages inverted.
What Should You Do About It?
- **Audit your current AI projects.** If something's been in pilot for more than 6 months without clear business metrics, either kill it or redesign it around measurable outcomes.
- **Shift your measurement.** Stop tracking technical metrics like completion rates or response times. Start tracking revenue impact, cost reduction, and hours saved.
- **Design for verification.** Any AI workflow handling high-stakes decisions needs a human checkpoint or automated verification layer. The hallucination problem isn't going away.
- **Invest in the 70%.** Before buying another AI tool, ask whether your processes and people are ready to use it. Most aren't.
- **Watch the leaders.** The companies in the 5% are publishing case studies. Study their approaches—not their tool choices—to understand what actually drives value.
If you're developing your AI strategy, the evidence is clear: the technology works, but only when the surrounding systems are designed to support it. And if you're evaluating AI tools for your business, don't start with capabilities—start with how you'll measure success.
What This Research Tells Us About the Next 12 Months
- **80% of enterprise AI projects never leave the pilot stage**—this isn't a technology problem, it's a strategy problem rooted in poor data foundations and inadequate process design.
- **The 5% of 'AI future-built' companies are seeing 5x the revenue gains**—proving the technology works when implemented correctly.
- **Hallucinations may be inherent to how LLMs work**—design for verification rather than hoping accuracy improves.
- **The 10/20/70 rule matters**: spend 70% of your AI budget on people and processes, not technology.
- **Measure AI by business outcomes** (revenue, costs, hours saved)—not by technical metrics like completion rates or chat volumes.
