What MIT got wrong about AI agents: New G2 data shows they’re already driving enterprise ROI

Check your research, MIT: 95% of AI projects don’t fail – far from it.

According to new data from G2Nearly 60% of companies already have AI agents in production, and less than 2% actually fail once deployed. That paints a very different picture than recent academic predictions that point to a widespread stagnation of AI projects.

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 to be much more durable and ‘sticky’ than early generative AI pilots.

“Our report really points out that agentic is a different beast when it comes to AI when it comes to failure or success,” Tim Sanders, G2’s head of research, told VentureBeat.

Transfer to AI in customer service, BI, software development

Sanders points out that it is now often referred to MIT studyreleased in July, only considered custom gen AI 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 did not announce a P&L impact, their projects were considered a failure – even if in reality they were not.

G2s 2025 AI Agents Insights Reporton the other hand, surveyed more than 1,300 B2B decision makers and concluded that:

• 57% of companies have agents in production and 70% say agents are “the core of their business”;

• 83% of them are satisfied with the performance of agents;

• Companies now invest an average of more than $1 million per year, with 1 in 4 spending more than $5 million;

• 9 in 10 plan to increase this investment over the next twelve months;

• Organizations have seen 40% cost savings, 23% faster workflows, and 1 in 3 report more than 50% speed gains, especially in marketing and sales;

• Nearly 90% of survey participants reported higher employee satisfaction in departments where officers were deployed.

The top 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 had the agent do a task and then they would immediately roll it back if it was a bad action, or they would do QA so they could retract the bad actions very, very quickly,” he explained.

At the same time, agent programs that involved a human 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. “Years from now there will be a human in the loop,” he said. “More than half of our respondents told us there is more human oversight than we expected.”

However, nearly half of IT buyers are comfortable giving agents full autonomy on low-risk workflows such as data recovery 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 people to make final passes and final decisions.

A classic example of this is a mortgage loan, Sanders noted: Agents do everything up until the moment the human analyzes their findings and accepts or rejects the loan.

If there are any errors, they are in the background. “It just doesn’t publish on your behalf and 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 offers Agent power “is gaining” from off-the-shelf agents and in-house builds, which take 38% of the total market share, Sanders reported. However, many organizations seem to be going hybrid with the goal of eventually developing internal tools.

Then, because they want a reliable data source, “they will crystallize around Microsoft, ServiceNow, Salesforce, companies that have a real system of record,” he predicted.

AI agents are not bound by deadlines

Why are agents (at least in some cases) so much better than humans? Sanders pointed to a concept called Parkinson’s lawstating that ‘the work is expanding to fill the time available for its completion.’

“Individual productivity does not lead to organizational productivity, because people are really only driven by deadlines,” Sanders said. When organizations looked at gen AI projects, they didn’t move the goalposts; the deadlines did not change.

“The only way to solve that is to move the goal post up or to interact with non-humans, because non-humans are not subject to Parkinson’s law,” he said, noting that they do not suffer from “human procrastination syndrome.”

Cops don’t take breaks. They don’t get distracted. “They just grind so you don’t have to change the deadlines,” Sanders said.

“If you focus on faster and faster QA cycles, which may even be automated, you will fix your agents faster than your people.”

Start with business problems and understand that trust builds slowly

Still, Sanders sees AI following the cloud when it comes to trust: He remembers in 2007 when everyone was quick to adopt cloud tools; In 2009 or 2010 there was a kind of decline in confidence.

Combine this with safety concerns: 39% of all respondents in G2’s survey said they had a security incident since the use 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 retrained to never repeat a bad action again.

Always include IT activities in AI implementations, he advised. They know what went wrong with gen AI and robotic process automation (RPA) and can get to the bottom of the explainability, leading to much more trust.

But on the other hand: don’t blindly trust suppliers. In fact, only half of respondents said they did; Sanders noted that the most important trust signal is the explainability of agents. “In qualitative interviews we were asked again and again whether you were [a vendor] You cannot explain it, you cannot deploy and manage it.”

It’s also critical to start with the business problem and work backwards, he advised: don’t buy brokers, then look for a proof of concept. When leaders deploy agents to the biggest pain points, internal users will be more forgiving when incidents occur and more willing to iterate, expanding their capabilities.

“People still don’t trust the cloud, they certainly don’t trust the generation of AI, they may not trust agents until they experience it, and then the game changes,” Sanders says. “Trust comes with a mule – you don’t just get forgiveness.”

Source link