What If AI Is Just Bad at Most Things?

This is Part 2 of a three-part series on the future of AI adoption. Part 1: "What If Exposure Breeds Exhaustion?"

The demo always works. A product manager at a developer conference watches an AI agent book a flight, cross-reference her calendar, and draft a summary email in eleven seconds flat. She pulls out her phone to try it on her own calendar, and it books the wrong Wednesday.

This isn't a failure of imagination or effort. The technology on stage is real, the eleven-second workflow genuinely happened, and the gap between that moment and the wrong Wednesday is the entire story of AI agents in 2026. According to research replicated across Anaconda, Forrester, a16z, and MIT Sloan, 88% of AI agent pilots never reach production, regardless of company size. An S&P Global survey found that 42% of companies abandoned most of their AI initiatives in 2025, up from 17% the prior year. The MIT NANDA report put the number of enterprises achieving rapid revenue acceleration from AI pilots at roughly 5%.

Demos work because they're single-turn, controlled, and optimized for the exact task being shown. Real work is multi-step, messy, and full of edge cases nobody optimized for, which is why 88% of those pilots stall before they ever reach production.

The Compound Error Rate

The math behind agent failures is straightforward and unforgiving. If each step in a five-step automated workflow succeeds 95% of the time, the end-to-end success rate drops to 77%. At 90% per step it's 59%, and at 85% it falls to 44%, which is roughly a coin flip on whether the workflow completes at all. These aren't hypothetical numbers chosen to scare; they're the range where current models actually operate depending on task complexity.

Hallucination rates haven't been solved so much as narrowed. A 2026 benchmark across 37 models found rates between 15% and 52% depending on task difficulty, with the top reasoning models, including GPT-5, Claude Sonnet 4.5, and Gemini-3-Pro, all exceeding 10% on harder benchmarks. A 10% error rate in a single chatbot query means you verify the answer and move on, but in a twenty-step agent workflow where each step feeds the next, it means at least one step is probably wrong and you won't know which one until the output looks subtly off three steps later.

The perception gap may be more revealing than the error rates themselves. METR's developer productivity study found that experienced open-source developers using AI tools took 19% longer to complete tasks while believing they were 24% faster. A 2026 follow-up revised the slowdown to roughly 4%, well within statistical noise, but the perception gap persisted. Developers consistently overestimate how much the tools help, which means they underestimate how much time they spend going back to verify and correct the output. The tools aren't saving as much time as they feel like they're saving, and that misperception matters because it shapes purchasing decisions, enterprise rollout timelines, and the industry's own sense of how close agents are to being ready.

The Smartphone in 2008

The iPhone launched in June 2007, and Apple shipped 1.4 million units that first year. The device was revolutionary and also genuinely bad: no App Store (that came in 2008), no copy-paste (2009), no multitasking (2010), locked to a single carrier. The people who dismissed it as an expensive toy weren't making a category error; they were describing the product accurately as it existed on the day they held it.

Pew Research tracked smartphone ownership through the adoption curve: 35% of U.S. adults owned one by 2011, four years after launch, and the number hit 46% by early 2012 before finally crossing 50% in 2013. The "smartphone moment," where your uncle who hates technology has one and uses it without complaint, was somewhere around 2013 to 2015, after six to eight years of early adopters dealing with crashing apps, carrier lock-in, and battery life measured in hours.

AI agents in May 2026 might be somewhere around smartphones in 2008: past the proof of concept but before the product that works for normal people. The person saying "I tried it, it wasn't worth the hassle" isn't confused or behind the curve so much as accurately reporting their experience with a product that hasn't matured yet. The iPhone skeptics of 2007 were correct about the 2007 iPhone and wrong about the smartphone, and both of those things were true simultaneously. Holding that contradiction without collapsing it into either "it's the future" or "it's a fad" is the difficult part, and it's where most of the AI discourse falls apart.

The Honest Confession

Future Shock runs an agentic news pipeline with 22 active cron jobs, sub-agent orchestration, automated research, editorial review, and publishing workflows. We are the target audience for AI agents, and we built our entire publication on top of them.

The failure modes are instructive. Sub-agents that report tasks as complete when they aren't, where the code says "done" but the output says otherwise, and the only way to catch it is to verify every result by hand. Model fallbacks that silently degrade output quality when the primary model hits a rate limit, changing not just the speed but the behavior of the entire orchestration layer. Memory systems that remember fifty conversations perfectly and forget the critical one, creating a kind of false confidence that's actually worse than having no memory at all, because at least with no memory you know to double-check everything.

Simon Willison, one of the most capable AI practitioners working today, described hitting a wall after about two hours of sustained agent orchestration at a conference in March 2026. The cognitive load isn't in generating or prompting; it's in supervising, watching the output closely enough to catch mistakes that look plausible until you check the details. That kind of sustained vigilance is genuinely exhausting in a way that just doing the work yourself often isn't.

The honest version of "AI agents work" is that they work the way a car worked in 1910: you can get where you're going, but you need to know how to fix an engine. The effort-to-value ratio is manageable for people who already understand the failure modes and prohibitive for everyone else, which is why most people will keep using chatbots for single-turn queries the way they used to use Google, copying and pasting answers into documents. That's not a failure of adoption. It's a rational response to the current state of the technology.

The Counterarguments

Two objections deserve genuine weight.

"It's already good enough for hundreds of millions of people." ChatGPT has over 900 million weekly active users as of February 2026, more than doubling in a year. GitHub Copilot is embedded in millions of developer workflows. People are getting real value from AI right now, not from agents but from chatbots and copilots. Summarization, translation, first-draft writing, code completion, search augmentation: these work, today, for a lot of people. The "AI is bad at most things" framing arguably ignores the things it's demonstrably good at.

This objection is largely correct, and it sharpens the argument rather than undermining it. The things AI does well are single-turn, low-stakes, human-verified tasks. The moment you ask it to chain decisions autonomously, to be an agent rather than an assistant, a reliability cliff appears that most of the usage numbers don't capture. The gap between "useful chatbot" and "reliable agent" may turn out to be wider than the gap between "no AI" and "useful chatbot."

"Maybe it's worse than early. Maybe agents are fundamentally the wrong abstraction." The smartphone parallel assumes AI agents will follow the same maturation curve: clunky early product, engineering improvements, eventual mass-market reliability. But autonomous multi-step workflows might be more like self-driving cars, where the last 10% of reliability requires more engineering effort than the first 90% and the timeline stretches from years to decades. The compound error problem isn't just an engineering challenge that improves with scale. Every new capability introduces new failure modes, and the reliability floor might not rise as fast as the capability ceiling.

This is the harder objection to dismiss. The METR follow-up showed the productivity gap narrowing with newer tools, which suggests real improvement, but the pace is slower than marketing claims by a wide margin. The honest answer is that nobody knows yet whether agents follow the smartphone curve or the self-driving curve. If the architectural critique is right, the timeline extends from years to potentially decades. Either way, the practical implication for most people in May 2026 is the same.

The Patience Argument

The product manager from the opening is not wrong about the conference demo or about the wrong Wednesday. Both things happened, and the distance between them is the whole story of AI agents right now.

This is a timing problem, not a technology problem, or at least that's the bet this series is making. The smartphone in 2008 was genuinely bad by the standards of what it would become by 2015, and describing it accurately as bad in 2008 wasn't pessimism. It was observation. The people who waited until 2013 to buy a smartphone didn't miss anything except the frustration of using an early product that wasn't ready for them.

There's a position in the AI discourse that almost nobody occupies: it's okay to wait. Not because the technology is fake, not because it won't eventually become essential, but because the current version genuinely isn't good enough for most people's actual needs and pretending otherwise doesn't serve anyone. The people who need AI agents to work will keep building, the tools will keep improving, and at some point the gap between the demo and the deployment will close enough that the product manager's phone books the right Wednesday without her thinking about it.

That day isn't today, and saying so out loud is more useful than another demo.