Check your research, MIT: 95%. AI projects They don’t disappoint – on the contrary.
According to recent data from G2nearly 60% of companies already have AI agents in production, and less than 2% actually fail after deployment. This is a very different picture than recent academic predictions suggesting widespread stagnation in AI projects.
As one of the world’s largest crowdsourced software review platforms, the G2 dataset reflects real-world deployment trends that show AI agents are proving significantly more hard-wearing and “better” than early generative AI pilots.
“Our report really shows that the agent is a completely different beast when it comes to AI in terms of failure or success,” Tim Sanders, head of research at G2, told VentureBeat.
Transfer of AI in customer service, BI, software development
Sanders points out that it is often invoked these days MIT studyreleased in July, only considered Gen. AI’s custom designs, Sanders argues, and many media outlets generalized this thesis to AI failures 95% of the time. He points out that university researchers analyzed public announcements, not circular data. If companies didn’t announce the impact on profits and losses, their projects were considered failures – even if they actually weren’t.
G2 AI Agent Analysis Report 2025in turn, conducted a survey of over 1,300 B2B decision makers and found that:
-
57% of companies employ agents in production, and 70% say that agents are the “basis of operations”;
-
83% are satisfied with the agent’s performance;
-
Businesses now invest on average more than $1 million a year, with 1 in 4 spending more than $5 million;
-
9 out of 10 plan to escalate this investment in the next 12 months;
-
Organizations saw 40% cost savings, 23% faster workflow, and 1 in 3 saw speed increases of over 50%, especially in marketing and sales;
-
Almost 90% of study participants observed higher employee satisfaction in departments where agents were delegated.
Leading utilize cases for AI agents? Customer service, business intelligence (BI) and software development.
Interestingly, G2 found a “surprising number” (about 1 in 3) of organizations that Sanders calls “let it rip.”
“They basically let the agent do the job and then either immediately rolled it back if it was a bad action, or they did a quality check so they could roll back the bad actions very, very quickly,” he explained.
At the same time, however, human-assisted agent programs were twice as likely to deliver cost savings—75% or more—than fully autonomous agent strategies.
This reflects what Sanders called the “dead embers” between “let it rip” organizations and “leave some human gates” organizations. “In a few years, a human will appear in the loop,” he said. “More than half of our respondents said there was more human oversight than we expected.”
However, almost half of IT buyers are not comfortable giving agents full autonomy in low-risk workflows such as data remediation or managing data pipelines. Meanwhile, think of BI and research as preparatory work, Sanders said; agents gather information in the background to prepare people to make final transitions and final decisions.
A classic example of this is mortgage lending, Sanders noted: Agents do everything at once until a person analyzes their arrangements and agrees to the loan.
If there are bugs, they are in the background. “He just doesn’t post on your behalf and put your name on it,” Sanders said. “As a result, you trust it more. You use it more often.”
When it comes to specific implementation methods, Salesforce’s Agent power “loses out” to off-the-shelf agents and proprietary solutions, gaining 38% of total market share, Sanders said. However, it appears that many organizations are moving to hybrid solutions with the goal of eventually introducing their own tools.
Then, because they want a trusted source of data, “they’ll start to cluster around Microsoft, ServiceNow, Salesforce and companies that have a true system of record,” he predicted.
AI agents are not driven by deadlines
Why are agents (at least in some cases) so much better than humans? Sanders pointed to the concept of the so-called Parkinson’s lawwhich states that “work expands so as to fill the time allotted for its completion.”
“Individual productivity does not lead to organizational productivity because people are only really driven by deadlines,” Sanders said. When organizations looked at gen AI projects, they did not change their goals; the dates have not changed.
“The only way to fix this is to move the goalpost up or deal with non-human beings, because non-human beings are not subject to Parkinson’s law,” he said, pointing out that they do not suffer from “human procrastination syndrome.”
Agents don’t take breaks. They don’t get distracted. “They just work, so you don’t have to reschedule,” Sanders said.
“If you focus on faster and faster QA cycles, which can even be automated, you will fix your agents faster than you fix your people.”
Start with business problems, understand that trust is built slowly
Still, Sanders thinks AI is following the cloud when it comes to trust: he remembers 2007 when everyone was quickly adopting cloud tools; then in 2009 or 2010 “there was a sort of trough of confidence.”
Combine this with safety concerns: 39% of all G2 survey respondents said they had experienced security incident since the implementation of artificial intelligence; In 25% of cases it was earnest. Sanders emphasized that companies need to think about measuring in milliseconds how quickly an agent can be retrained so that they never repeat bad behavior again.
He advised that IT operations should always be included in AI implementations. They know what went wrong with Gen AI and Robotic Process Automation (RPA) and can get to the bottom of explainability, which leads to much greater trust.
On the other hand, don’t blindly trust suppliers. In fact, only half of respondents said yes; Sanders noted that the No. 1 trust signal is an agent’s explainability. “In qualitative interviews we were constantly told yes [a vendor] “You can’t explain it, you can’t implement it and you can’t manage it.”
It’s also very crucial to start with the business problem and work backward, he advised: Don’t buy agents and then look for proof of concept. If leaders apply measures to the biggest problems, internal users will be more understanding when incidents occur and will be more willing to repeat, thereby developing their skills.
“People still don’t trust the cloud, they definitely don’t trust generational AI, they may not trust agents until they experience it, and then the game will change,” Sanders said. “Trust comes on a mule – you won’t just be forgiven.”