Every pitch deck I've seen this year includes the phrase "AI agent." Every startup is building one. Every enterprise is buying one. And almost none of them work.
Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027. Honestly, I think that's generous. The failure rate for AI initiatives hit 42% in 2025, up from 17% the year before — and agents are the hardest category to get right.
So why is everyone still building them?
The Solution-First Trap
Here's the pattern I keep seeing: A team discovers LLMs can do cool stuff in a demo. They build an agent around it. Then they go looking for a problem it solves.
This is backwards. Good products start with pain. Someone has a problem that costs them time, money, or sanity. You build the simplest thing that fixes it. If AI is the right tool, great. If a spreadsheet works, use the spreadsheet.
But "we built an AI agent" gets funding. "We built a really good spreadsheet" doesn't. So here we are.
What's Actually Broken
Non-determinism. When you call a function, you expect the same input to produce the same output. AI agents don't work that way. Run the same query twice, get two different results. Try explaining that to a compliance team.
Compounding hallucinations. Multi-agent workflows — where one agent's output feeds into another — multiply the hallucination risk at every handoff. Gartner's hype cycle report flagged this explicitly. One agent might be 95% accurate. Chain five together and you're at 77%. That last 23% is where the lawsuits live.
Data readiness. 42% of enterprises need access to 8+ data sources to make agents useful, and 79% expect data challenges to block their rollouts. Your agent is only as smart as the data it can access, and most enterprise data is a mess.
The "it works in the demo" problem. I've watched countless demos where an agent books a meeting, summarizes documents, and drafts emails flawlessly. Then you deploy it with real data, ambiguous requests, and edge cases. Suddenly it's booking meetings in the wrong timezone and summarizing the wrong document.
The Best AI Is the AI You Don't Notice
Gmail's smart reply. Spotify's Discover Weekly. Your phone's autocorrect (well, sometimes). These are AI features that just work. You don't think about them. You don't interact with an "agent." You get value.
The gap between these invisible AI features and the autonomous agent everyone's pitching is enormous. One filters spam quietly. The other tries to run your entire customer support operation and occasionally tells a customer your product is free.
What I'd Actually Invest In
If I were building an AI product right now, I'd skip the agent hype entirely and focus on:
- Single-task tools that do one thing reliably (summarize this doc, extract these fields, classify this ticket)
- Human-in-the-loop workflows where AI drafts and humans approve
- Data quality infrastructure because your AI is useless without clean data
None of this is sexy. None of it makes a good keynote demo. But it ships, it works, and it doesn't hallucinate your company into a PR crisis.
The Hype Will Pass
We've been through this cycle before. Blockchain was going to revolutionize everything. The metaverse was the future of work. Web3 was the new internet. Each had real technology underneath the hype — and each settled into a much smaller, more useful niche than the pitch decks promised.
AI agents will follow the same path. The technology is real. The applications are narrower than people think. The companies that survive will be the ones solving actual problems, not the ones with the flashiest demo.
Stop chasing the trend. Build something people need. If an agent is the right tool, you'll know — because the problem will demand it, not because your investor suggested it.
Top comments (0)