Check your research, MIT: 95% AI projects Not failing – far from it.
According to new data from g2About 60% of companies already have AI agents in production, and less than 2% actually fail after deployment. This paints a very different picture from recent academic forecasts that suggest widespread AI project stagnation.
As one of the world’s largest crowdsourced software review platforms, G2’s dataset reflects real-world adoption trends – showing that AI agents are proving far more durable and “sticky” than early generative AI pilots.
“Our report really indicates that agentic is a different animal when it comes to AI failure or success,” Tim Sanders, head of research at G2, told VentureBeat.
Handing over AI to customer service, BI, software development
Sanders points out that is now frequently referenced mit studySanders argues that, released in July, only General AI is considered custom projects, and many media outlets generalized that AI fails 95% of the time. He explains that university researchers analyzed public announcements rather than closed-loop data. If companies did not declare the P&L impact, their projects were considered failures – even if they did not actually fail.
G2’s 2025 AI Agents Insights ReportIn contrast, more than 1,300 B2B decision makers were surveyed and found:
-
57% of companies have agents in production and 70% say agents are the “core of operations”;
-
83% are satisfied with the agent’s performance;
-
Enterprises are now investing an average of more than $1 million annually, with 1 in 4 spending more than $5 million;
-
9 in 10 plan to increase that investment over the next 12 months;
-
Organizations have seen 40% cost savings, 23% faster workflow, and 1 in 3 report more than 50% speed gains, especially in marketing and sales;
-
Nearly 90% of study participants reported high employee satisfaction in departments where agents were deployed.
Major 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 if it was a bad task they would either roll it back immediately, or do QA so they could roll back bad tasks very quickly,” he explained.
At the same time, however, agent programs with a human in the loop were twice as likely to deliver cost savings – 75% or more – compared to fully autonomous agent strategies.
This reflects what Sanders called a “dead heat” between ‘Let It Rip’ organizations and ‘Leave Some Humans Out the Door’ organizations. “There’s going to be a human in the loop several years from now,” he said. “More than half of our respondents told us there is more human oversight than we expected.”
However, about half of IT buyers are comfortable giving agents full autonomy in low-risk workflows like data remediation or data pipeline management. In the meantime, think of BI and research as prep work, Sanders said; Agents gather information in the background to prepare humans to make the final pass and final decision.
A classic example of this is a mortgage loan, Sanders said: Agents do everything right until a human analyzes their findings and says yes or no to the loan.
If there are mistakes, they are in the background. “It’s not just published on your behalf and doesn’t put your name on it,” Sanders said. “So as a result, you rely on it more. You use it more.”
When it comes to specific deployment methods, Salesforce’s agentforce Sanders reported that readymade is “winning” over agents and in-house builds, taking 38% of total market share. However, it appears that many organizations are going hybrid with the goal of eventually standing up tools in-house.
Then, because they want a reliable source of data, “they’re going to crystallize around Microsoft, ServiceNow, Salesforce, companies with real systems of record,” he predicted.
AI agents are not driven by deadlines
Why are agents (at least in some cases) so superior to humans? Sanders pointed to a concept called parkinson’s lawWhich states that ‘the work expands so as to cover the time available for its completion.’
Sanders said, “Individual productivity does not lead to organizational productivity because humans are really only motivated by deadlines.” When organizations looked at General AI projects, they did not move goal positions; The deadline did not change.
“The only way you fix this is to either move the goal post higher or deal with non-humans, because non-humans are not subject to Parkinson’s Law,” he said, pointing out that they do not suffer from “human delay syndrome.”
Agents don’t take breaks. They don’t get distracted. “They just grind it out so you don’t have to make changes to the deadlines,” Sanders said.
“If you focus on faster and faster QA cycles that can also be automated, you get your agents fixing things faster than your humans.”
Start with business problems, understand that trust is built at a slow pace
Still, Sanders sees AI following the cloud when it comes to trust: He remembers 2007 when everyone was quick to deploy cloud tools; Then by 2009 or 2010, “there was a bit of a trust crisis.”
Combine this with security concerns: 39% of all respondents to G2’s survey said they have experienced security incident Since the deployment of AI; 25% of the time, it was serious. Sanders emphasized that companies should think about measuring in milliseconds how quickly an agent can be trained not to repeat a bad action again.
He advised to always include IT operations in AI deployment. They know what went wrong with general AI and robotic process automation (RPA) and can get to the bottom of the explanation, which leads to a lot of trust.
However, the flip side is this: don’t trust sellers blindly. In fact, only half of the respondents said they did so; Sanders said the No. 1 trust signal is agent interpretability. “In qualitative interviews, we were told repeatedly, if you (a vendor) can’t explain it, you can’t deploy and manage it.”
It’s also important to start with a business problem and work backwards, he advised: Don’t buy agents, then look for proof of concept. If leaders deploy agents at the biggest problem points, internal users will be more forgiving when incidents occur, and more willing to iterate, thereby building their skills.
“People still don’t trust the cloud, they certainly don’t trust Gen AI, they can’t trust agents until they experience it and then the game changes,” Sanders said. “Trust comes riding on a mule – you don’t just get forgiveness.”
