Only about 20% of all AI deployments succeed — and only a fraction of those typically involve autonomous Agents rather than AI Assistants. In my view, what actually drives these successes is the sound judgment of decision-makers who chose the safer path — one more appropriate for today's corporate reality.
The Success of AI Agents Isn't as Large-Scale as It Seems
In the AI industry, seemingly contradictory data circulates: 80% of AI projects fail, yet 88% of companies deploying AI agents achieve a positive return on investment (ROI). I haven't seen these figures presented side by side with a unified explanation, which is why they may appear paradoxically contradictory. In reality, they simply represent different stages of the same transformation story. These data points reveal the actual, far smaller than commonly believed, scale of "AI Agent success."
What Do These Numbers Really Tell Us?
The 80% failure rate covers all AI projects — from unfinished proof-of-concepts to pilots without a clear objective, naturally including failed AI Agent deployments as well. The 88% ROI figure, on the other hand, pertains to a small, elite group of companies that survived the gauntlet of failures, built the infrastructure, gathered the data, and brought their deployments to completion. In other words, "88% ROI" represents only a small slice (we don't know what percentage of successful deployments are actually Agent-based projects) of the 20% of projects that succeeded at all.
Assistants First, Then Agents
In theory, an "AI Agent" is a system that operates autonomously — it can independently analyze, decide, and execute tasks. From my observations — confirmed by data from IDC and EY reports — most companies are not yet ready for full AI Agents. They lack the data, competencies, and organizational readiness. That is why, when deploying so-called Agents, companies are in fact deploying Assistants — systems that support humans but don't operate independently. This is a "human + machine" model where AI executes tasks while humans retain control. Hence my conclusion that 2025 is not the "year of AI Agents," but rather the year of learning how to prepare for Agents — step by step — partly through building AI Assistants.
What Actually Works?
Research from IDC and Google Cloud shows that positive ROI emerges where AI supports specific processes rather than trying to replace entire roles. Companies that deploy AI with real tasks in mind achieve as much as $3.7 in return for every dollar invested within a year. This isn't technological magic — it's the result of a mature strategy: a well-designed workflow, meaningful data, and human oversight. So if you're thinking about AI Assistants (not Agents) for your company, you're probably right.
My Takeaway for Boards
Don't ask: "How do we replace people with AI?" Ask instead: "How do we relieve people of repetitive tasks so they can focus on decisions and relationships?" Because the truth is, 2025 and likely 2026 as well won't be "the years of AI Agents," but years of prudent leadership in the AI era. Before we build an AI Agent, we first need to build an organization capable of working alongside one.
Dear Reader, if you believe this topic is relevant to your company and would like to discuss with me and your Board how to effectively match AI solutions to your company's realities, I invite you to get in touch — Leszek Giza.