People have been talking about it for a while, but now the industry is seeing part of that new rush to utilize what can be a game-changer for companies.
Of course, the rise of the AI agent is no small thing. Many people who had a front-row seat to the cloud and all of the disruption that it brought understand that it was a pebble in the ocean compared to what’s coming.
The prospect of simulating human decision-making, and handing knowledge work to large language models, is a big deal. It leads to replacing human workers with something much less costly and more durable – workers who never need lunch, or a bathroom break.
Yes, AI engines are far more efficient than humans in so many ways, and now we’re seeing that bear out in the enterprise markets.
Major Enterprise Use Cases for AI Agents
Reading through some reports on the utility of enterprise AI agents, I noticed that many of them refer to customer support, as well as marketing and process support or fulfillment, as popular implementations. A few other top use cases involve knowledge assistance, generative AI in existing workflows, and the daily usage of productivity tools by front-line workers. That last one speaks to the often-promoted idea of the “human in the loop” or HITL, and the desire that AI not replace humans, but augment their work instead. Practically, though, some of these AI agents leave us wondering: what is the HITL actually needed for?
Market Projections
Consultants and reporting companies are chiming in with rosy projections for the year ahead.
Market.us estimates $3.6 billion for the enterprise AI agent market in 2023 and $139 billion by 2033. Deloitte adds the following projection: 25% of companies expected to embrace AI agents by 2025, and 50% two years later. However, given that nearly all companies everywhere will want some of this functionality, the numbers, in both cases, are likely to be much higher.
And here’s this from a McKinsey report:
“McKinsey research sizes the long-term AI opportunity at $4.4 trillion in added productivity growth potential from corporate use cases.”
In a recent episode of the AI Daily Brief podcast, (one of my favorites), host Nathaniel Whittemore talks about a move in executive estimates of investment: from $9 million last quarter to $114, million in Q1 of 2025.
“I think that those of us, probably most of you, who are listening, who have used these tools, find them very quickly making their way into your daily habits,” Whittemore adds. “I would expect to see (the numbers do) nothing but increase in the coming quarters.”
He also talks about a KPMG study of companies launching enterprise pilots after experimentation with the technology, suggesting that doubled from 37% in Q4 to 65% in Q1, and that 99% of companies said they intend to deploy these agents at some point
“99% of organizations surveyed said that they plan to deploy agents, suggesting to me that 1% of organizations misread the question,” he adds. Whether or not 1% deliberately forswear the technology is probably beside the point – we have to anticipate that demand is going to be very high.
Remaining Barriers
Although models like OpenAI’s o3 are evolving quickly, and no-code tools are democratizing the process of creating applications, there are still some clear boundaries to what AI can do in the workplace.
A main one consists of accuracy challenges. The most common word applied to this for LLMs is “hallucinations.” Experts are finding, in general, that models with more inference are producing more hallucinations, and that’s a problem as these uses become more important to the companies that have already jumped on the bandwagon.
Case in point: a news story showing a customer support engine named Sam at Cursor, who apparently created a new policy erroneously, and started shutting down people‘s access to the platform. The shakeout showed why these kinds of mistakes make a difference.
Another concern is hacking, where bad actors could take advantage of the functionality to compromise systems. A third is regulation – what is the landscape around these agents going to be like?
All of these should be considered as top brass mull opportunities.
Some Advice
I also came across this handy chart and process description from Gartner, of which the firm’s magic quadrant report has been so helpful in the IT world.
Garner representatives suggest mapping the enterprise pain points, and then addressing them with the AI agents. Address them, to do what?
“Enhance customer experiences,” the authors write. “Streamline operations, and uncover new products, services, or revenue streams.”
Another approach that might deal with hallucinations is ensemble learning. Having one model check the work of another can prevent those hallucinations and mistakes from percolating into the places where AI agents help with production. Some are suggesting even access to web search can help mitigate a model’s hallucinations, which is another thing brought up on that AI Daily Brief episode.
Getting Ready
In so many of the events that I’ve been privileged to attend, and even host over the last year or so, I have heard the same refrains: that we have to get ready to welcome in AI agents into the fold.
What all of this tells us is that the idea of company AI agent adoption is not just a flash in the pan. It’s happening all around us, and we should be paying attention.