This is the time of year for predictions and trend analyses, and as data science and artificial intelligence become increasingly important to the global economy, it’s vital that leaders watch emerging AI trends.

Nobody seems to use AI to make these predictions, and we won’t either, as we share our list of AI trends that will matter in 2025. But we will incorporate the latest research whenever possible. Randy has just completed his annual survey of data, analytics, and AI executives, the 2025 AI & Data Leadership Executive Benchmark Survey, conducted by his educational firm, Data & AI Leadership Exchange; and Tom has worked on several surveys on generative AI and data, technology leadership structures, and, most recently, agentic AI.

Here are [two of the five] 2025 AI trends on our radar screens that leaders should understand and monitor.

1. Leaders will grapple with both the promise and hype around agentic AI.

Let’s get agentic AI — the kind of AI that does tasks independently — out of the way first: It’s a sure bet for 2025’s “most trending AI trend.” Agentic AI seems to be on an inevitable rise: Everybody in the tech vendor and analyst worlds is excited about the prospect of having AI programs collaborate to do real work instead of just generating content, even though nobody is entirely sure how it will all work. Some IT leaders think they already have it (37%, in a forthcoming UiPath-sponsored survey of 252 U.S. IT leaders); most expect it soon and are ready to spend money on it (68% within six months or less); and a few skeptics (primarily encountered by us in interviews) think it’s mostly vendor hype.

Most technology executives believe that these autonomous and collaborative AI programs will be primarily based on focused generative AI bots that will perform specific tasks. Most people believe that there will be a network of these agents, and many are hoping that the agent ecosystems will need less human intervention than AI has required in the past. Some believe that the technology will all be orchestrated by robotic process automation tools; some propose that agents will be fetched by enterprise transaction systems; and some posit the emergence of an “uber agent” that will control everything.

Here’s what we think: There will be (and in some cases, already are) generative AI bots that will do people’s bidding on specific content creation tasks. It will require more than one of these agentic AI tools to do something significant, such as make a travel reservation or conduct a banking transaction. But these systems still work by predicting the next word, and sometimes that will lead to errors or inaccuracies. So there will still be a need for humans to check in on them every now and then.

The earliest agents will be those for small, structured internal tasks with little money involved — for instance, helping change your password on the IT side, or reserving time off for vacations in HR systems. We don’t see much likelihood of companies turning these agents loose on real customers spending real money anytime soon, unless there’s the opportunity for human review or the reversal of a transaction. As a result, we don’t foresee a major impact on the human workforce from this technology in 2025, except for new jobs writing blog posts about agentic AI. (Wait, can agents do that?)

2. The time has come to measure results from generative AI experiments.

One of the reasons why everybody is excited about agents is that as of 2024, it has still proved difficult to demonstrate economic value from generative AI. We argued in last year’s AI trends article that the value of GenAI still needed to be demonstrated. Data and AI leaders in Randy’s 2025 AI & Data Leadership Executive Benchmark Survey said they are confident that GenAI value is being generated: Fifty-eight percent said that their organization has achieved exponential productivity or efficiency gains from AI, presumably mostly from generative AI. Another 16% said that they have “liberated knowledge workers from mundane tasks” through the use of GenAI tools. Let’s hope that these highly positive beliefs are correct.

But companies shouldn’t take such confidence on faith. Very few companies are actually measuring productivity gains carefully or figuring out what the liberated knowledge workers are doing with their freed-up time. Only a few academic studies have measured GenAI productivity gains, and when they have, they’ve generally found some improvements, but not exponential ones. Goldman Sachs is one of the rare companies that has measured productivity gains in the area of programming. Developers there reported that their productivity increased by about 20%. Most similar studies have found contingent factors in productivity, where either inexperienced workers gain more (as in customer service and consulting) or experienced workers do better (as in code generation).

In many cases, the best way to measure productivity gains will be to establish controlled experiments. For example, a company could have one group of marketers use generative AI to create content without human review, one use it with human review, and a control group not use it at all. Again, few companies are doing this, and this will need to change. Given that GenAI is primarily about content generation for many companies right now, if we want to really understand the benefits, we’ll also have to start measuring content quality. That’s notoriously difficult to do with knowledge work output. However, if GenAI helps write blog posts much faster but the posts are boring and inaccurate, that’s important to measure: There will be little benefit in that particular use case.

The sad fact is that if many organizations are actually to achieve exponential productivity gains, those improvements may be measured in large-scale layoffs. But there is no sign of mass layoffs in the employment statistics. Additionally, a Nobel Prize winner in economics this year, MIT’s Daron Acemoglu, has commented that we haven’t seen real productivity gains from AI thus far, and he doesn’t expect to see anything dramatic over the next several years — perhaps a 0.5% increase over the next decade. In any case, if companies are really going to see and profit from GenAI, they’re going to need to measure and experiment to see the benefits.