Check your research, MIT: 95% of AI projects aren’t failing — far from it.
According to new data from G2, nearly 60% of companies already have AI agents in production, and fewer than 2% actually fail once deployed. That paints a very different picture from recent academic forecasts suggesting widespread AI project stagnation.
As one of the world’s largest crowdsourced software review platforms, G2’s dataset reflects real-world adoption trends — which show that AI agents are proving far more durable and “sticky” than early generative AI pilots.
“Our report’s really pointing out that agentic is a different beast when it comes to AI with respect to failure or success,” Tim Sanders, G2’s head of research, told VentureBeat.
Handing off to AI in customer service, BI, software development
Sanders points out that the now oft-referenced MIT study, released in July, only considered gen AI custom projects, Sanders argues, and many media outlets generalized that to AI failing 95% of the time. He points out that university researchers analyzed public announcements, rather than closed-loop data. If companies didn’t announce a P&L impact, their projects were considered a failure — even if they really weren’t.
G2’s 2025 AI Agents Insights Report, by contrast, surveyed more than 1,300 B2B decision-makers, finding that:
57% of companies have agents in production and 70% say agents are “core to operations”;
83% of are satisfied with agent performance;
Enterprises are now investing an average of $1 million-plus annually, with 1 in 4 spending $5 million-plus;
9 out of 10 plan to increase that investment over the next 12 months;
Organizations have seen 40% cost savings, 23% faster workflows, and 1 in 3 report 50%-plus speed gains, particularly in marketing and saless;
Nearly 90% of study participants reported higher employee satisfaction in departments where agents were deployed.
The leading use cases for AI agents? Customer service, business intelligence (BI) and software development.
Interestingly, G2 found a “surprising number” (about 1 in 3) of what Sanders calls ‘let it rip’ organizations.
“They basically allowed the agent to do a task and then they would either roll it back immediately if it was a bad action, or do QA so that they could retract the bad actions very, very quickly,” he explained.
At the same time, though, agent programs with a human in the loop were twice as likely to deliver cost savings — 75% or more — than fully autonomous agent strategies.
This reflects what Sanders called a “dead heat” between ‘let it rip’ organizations and ‘leave some human gates’ organizations. “There’s going to be a human in the loop years from now,” he said. “Over half of our respondents told us there’s more human oversight than we expected.”
However, nearly half of IT buyers are comfortable with granting agents full autonomy in low-risk workflows such as data remediation or data pipeline management. Meanwhile, think of BI and research as prep work, Sanders said; agents gather information in the background to prepare humans to make last passes and final decisions.
A classic example of this is a mortgage loan, Sanders noted: Agents do everything right up until the human analyzes their findings and yay or nays the loan.
If there are mistakes, they’re in the background. “It just doesn’t publish on your behalf and put your name on it,” said Sanders. “So as a result, you trust it more. You use it more.”
When it comes to specific deployment methods, Salesforce’s Agentforce “is winning” over ready-made agents and in-house builds, taking up 38% of all market share, Sanders reported. However, many organizations seem to be going hybrid with a goal to eventually stand up in-house tools.
Then, because they want a trusted source of data, “they’re going to crystallize around Microsoft, ServiceNow, Salesforce, companies with a real system of record,” he predicted.
AI agents aren’t deadline-driven
Why are agents (in some instances at least) so much better than humans? Sanders pointed to a concept called Parkinson’s Law, which states that ‘work expands so as to fill the time available for its completion.’
“Individual productivity doesn’t lead to organizational productivity because humans are only really driven by deadlines,” said Sanders. When organizations looked at gen AI projects, they didn’t move the goal posts; the deadlines didn’t change.
“The only way that you fix that is to either move the goal post up or deal with non-humans, because non-humans aren’t subject to Parkinson’s Law,” he said, pointing out that they’re not afflicted with “the human procrastination syndrome.”
Agents don’t take breaks. They don’t get distracted. “They just grind so you don’t have to change the deadlines,” said Sanders.
“If you focus on faster and faster QA cycles that may even be automated, you fix your agents faster than you fix your humans.”
Start with business problems, understand that trust is a slow build
Still, Sanders sees AI following the cloud when it comes to trust: He remembers in 2007 when everyone was quick to deploy cloud tools; then by 2009 or 2010, “there was kind of a trough of trust.”
Mix this in with security concerns: 39% of all respondents to G2’s survey said they’d experienced a security incident since deploying AI; 25% of the time, it was severe. Sanders emphasized that companies must think about measuring in milliseconds how quickly an agent can be retrained to never repeat a bad action again.
Always include IT operations in AI deployments, he advised. They know what went wrong with gen AI and robotic process automation (RPA) and can get to the bottom of explainability, which leads to a lot more trust.
On the flip side, though: Don’t blindly trust vendors. In fact, only half of respondents said they did; Sanders noted that the No. 1 trust signal is agent explainability. “In qualitative interviews, we were told over and over again, if you [a vendor] can’t explain it, you can’t deploy it and manage it.”
It’s also critical to begin with the business problem and work backwards, he advised: Don’t buy agents, then look for a proof of concept. If leaders apply agents to the biggest pain points, internal users will be more forgiving when incidents occur, and more willing to iterate, therefore building up their skillsets.
“People still don’t trust the cloud, they definitely don’t trust gen AI, they might not trust agents until they experience it, and then the game changes,” said Sanders. “Trust arrives on a mule — you don’t just get forgiveness.”
