Check your research, MIT: 95 % AI projects Not failing – it’s far from it.
According to new data G2About 60 % of the companies already have AI agents in production, and once it is deployed, less than 2 % fails. It paints a very different image from recent educational predictions that suggest large -scale AI project stagnation.
As one of the world’s largest crowded software review platforms, G2’s datastate reflects the trends of adoption of real world-which shows that AI agents are far more durable and “sticky and” sticky “than the early generative AI pilots.
“Our report is really pointing out that the agent is a different animal when it comes to AI in connection with the failure or success.”
Customer Service, BI, to hand over AI to Software Development
Sanders said it was now cited The study of MITSanders says, released in July, only General AI customs projects were considered, and many media outlets generally made it common for 95 % of the failure. Researchers at the university analyzed public announcements instead of closed loop data, he said. If the companies did not announce the impact of P&L, their plans were considered a failure – even if they were not really.
G2 Reports of insights of 2025 AI agentsOn the contrary, surveyed more than 1,300 B2B decision makers, found that:
57 % of companies are agents in production and 70 % say agents are “the core of operations”.
83 % are satisfied with the agent performance.
Enterprises are now investing more than 1 million annually, an average of more than $ 5 million.
9 of 10 plans to increase this investment in the next 12 months;
Organizations have seen 40 % cost saving, 23 % faster workflow, and 50 % plus speed benefits, especially marketing and salvation.
About 90 % of the study reported the high satisfaction of the employees in the department where agents were deployed.
Well -known cases for AI agents? Customer Service, Business Intelligence (BI) and Software Development.
Interestingly, G2 got a “amazing number” (about 1 in 3) in which Sanders called ‘chicked it’ to the organizations.
He explained, “They basically allowed the agent to do something and then they would either return it immediately if it was a bad action, or qa so that they could withdraw the bad movements very quickly.”
At the same time, although, the cost savings in agent programs with humans in the loop – 75 % or more – was likely to double the fully independent agent strategies.
This shows that Sanders is called ‘dead hat’ between which this is ‘chair’, leave the organizations and ‘some human gates’ organizations. He said, “Loop is going to be human in years. “More than half of our respondents told us that we have more human surveillance than we expected.”
However, almost half of the IT buyers are comfortable to give full independence in low -risk workflows such as low -risk workflower or data pipeline management. Meanwhile, Sanders said, think of BI and research as a pre -work. Agents collect information in the background to prepare humans to make the final pass and final decisions.
Sanders noted that an excellent example of this is a mortgage debt: Agents do everything unless humans analyze their results and yes or loans.
If there are mistakes, they are in the background. Sanders said, “It doesn’t just publish you and make a name for it.” “So as a result, you have more confidence in it. You use it more.”
When it comes to specific deployment methods, sales force Agent force According to Sanders, ready -made agents and abroad construction “wins”, which has 38 % of all market shares. However, many organizations seem to be hybrid, with the purpose of ultimately aimed at the interior tools.
After that, because they want a reliable source of data, “they are going to curb Microsoft, Servinu, Sales Force, Real Records Companies.”
AI agents are not based on deadline
Why are the agents (at least in some cases) why better than humans? Sanders pointed to a concept Parkinson’s lawWhich states that ‘work is spread so that the time available to fulfill it can be met.’
Sanders said, “Individual production capacity does not lead to organizational productivity because humans are just driven by deadline.” When organizations saw General AI projects, they did not move round posts. The deadline has not changed.
He said, “The only way you fix this is to either move the round post or deal with non -humans, because the inhuman Parkinson is not subject to the law,” he said, indicating that he was not suffering from “human prejudice syndrome”.
Agents do not take a break. They do not engage. Sanders said, “They are just grinding so you don’t need to change the deadline.”
“If you focus on fastest and fastest QA bicycles that can be automatic, you fix your agents much faster than fixing your humans.”
Start with business issues, understand that trust is a slow pace
Nevertheless, when Sanders speak of confidence, they see the cloud following: He remembers in 2007 when everyone was quick to deploy cloud tools. Then until 2009 or 2010, “there was a type of trust.”
Mix in this with security concerns: 39 % of all the respondents of the G2 survey said they would experience Security incident Since the deployment of AI; 25 % of the time, it was severe. Sanders emphasized that companies must think about measurement in millions of seconds how quickly an agent can be re -trained to repeat the bad action.
He advised that II should always include IT operations in the deployment of II. They know what has been wrong with General AI and robotic process automation (RPA) and they can go down the explanation, which is a lot of confidence.
Flip side, though: Do not trust the shopkeepers with your eyes. In fact, only half the respondents said they did. Sanders noted that the number 1 trust signal agent is explained. “In a quality interview, we were repeatedly told, if you (a vendor) can’t explain it, you cannot deploy it and manage it.”
He suggested that it is important to start with a business problem and work backwards: Do not buy agents, then find evidence of concept. If leaders apply agents to the biggest pain places, internal users are more pardoned when events occur, and more pleased to repeat it, so increase their skills.
Sanders said, “People still do not trust the cloud, they certainly do not rely on General Ai, may not trust the agents unless they experience it, and then the game changes.” “Trust reaches a mule – you just don’t get apology.”