Every major AI company is selling agents right now. Salesforce has Agentforce. Microsoft has Copilot agents embedded across the 365 suite. Google has Agentspace. The pitch is consistent: deploy an AI that can take multi-step actions on your behalf, hand it a goal, walk away. What the demos don’t show is the failure rate once the environment gets even slightly unpredictable.
The core problem isn’t the models themselves - it’s that agents are being evaluated in controlled conditions and shipped into chaotic ones. A customer service agent that handles straightforward refund requests beautifully will stall, loop, or silently fail when a user presents a request that sits outside the training distribution. And unlike a chatbot that gives a wrong answer, an agent that takes wrong actions - submitting a form it shouldn’t, escalating a ticket incorrectly, misreading a file structure - creates downstream errors that are harder to audit and expensive to unwind.
This is what makes the current marketing push genuinely misleading rather than just optimistic. When Salesforce announced Agentforce in late 2024, the framing was squarely around autonomous execution and ROI. The fine print about human-in-the-loop requirements, the need for structured data environments, and the significant setup overhead required from enterprise IT teams was not the headline. It rarely is.
The reliability bar is being set by demos, not deployments
There’s a version of this that’s just normal tech marketing - show the best case, let early adopters discover the edge cases. But agents introduce a category of risk that earlier AI products didn’t. With a large language model answering questions, errors are visible and local. With an agent operating across tools, APIs, and databases, errors can propagate before anyone notices. The expected failure mode isn’t a bad answer. It’s a bad action taken confidently.
Researchers working on agent reliability - including published work from teams at DeepMind and academic institutions studying multi-step task completion - have consistently found that success rates drop sharply as task length and environmental complexity increase. This isn’t a secret. It’s just not the number on the sales slide.
The enterprise is about to find this out the hard way
Mid-sized companies without dedicated AI engineering teams are the target market for most of these agent products. They’re also the least equipped to build the guardrails, monitoring pipelines, and fallback logic that make agents safe to run in production.
The vendors know this. The question is whether they’ve priced the support contracts accordingly - or whether the reckoning is coming in the form of churn, eighteen months from now, when the pilots don’t become rollouts.