The Agents are Coming

Posted by

Earlier this week, Google introduced Gemini 2.0, a new edition of its flagship LLM that, according to CEO Sundar Pichai, “will enable us to build new AI agents that bring us closer to our vision of a universal assistant.”

That came just days after AWS made an agentic announcement of its own, roughly two weeks after Microsoft did the same. Anthropic beat both companies to the punch by about a month, but was in turn about a month behind Salesforce, which aims to have a billion agents working on behalf of clients by the end of next year.

“I’ve never been more excited about anything at Salesforce, maybe in my career,” said CEO Marc Benioff of the company’s Agentforce platform in October.

Venture capital firms are pretty excited too, apparently. According to CB Insights, they’ve pumped $2 billion into agentic startups in the last two years.

Agentic AI, in other words, is sprinting up the Gartner Hype Cycle toward the peak of inflated expectations at an Olympian’s pace right now. Are you feeling the burn as much as I am?

West McDonald, founder of AI consultancy GoWest.ai and someone you’ve met here before, certainly is. “Everyone is kind of pre-selling it,” he says. Leslie Joseph (pictured), a principal analyst at Forrester, agrees.

“There’s a crescendo of vendor noise out there that seems to suggest agents are already here and already mature and already awesome,” he says. “That’s not true.”

But make no mistake, Joseph continues, agentic AI is coming, perhaps sooner than any of us thinks, and IT management will be among the first markets to feel its impact.

“It’s ground zero for disruption,” Joseph says. “Human labor and billing rates and commercial models are all going to be in heavy flux in the next few years.”

Thanks for reading Channelholic! Subscribe for free to receive new posts and support my work.

Weighing the risks

Marketing hype has muddied the waters some, according to Joseph, but the basic outlines of what makes an agent an agent have been clear for a while. To qualify, a program must be capable of remembering what it did before, planning what it’ll do next, interfacing with applications to get things done, collaborating with other agents, and getting better at its assignments over time by learning from experience.

Above all else though, an AI agent must be smart enough to do all that entirely on its own. “The agent should be able to perform completely autonomously unless it is specifically required from a governance perspective or as part of its instructions to reach out to a human at a certain point in time,” Joseph says.

Per last week’s post, Zendesk, Intercom, and others already have agents autonomously resolving customer service issues, and in theory, makers of RMM and PSA software could be resolving help desk issues the same way. In reality, however, there’s a big difference between processing a disappointed consumer’s refund and keeping a hospital’s mission-critical infrastructure securely operational.

“What if it hallucinates?” asks McDonald (pictured) of the hypothetical agent responsible for that task. “What if it gets something wrong?”

Such questions have made Atera (per reporting a year ago and in August) one of the few management vendors comfortable deploying fully autonomous AI. Even McDonald, a self-described fan of AI who spends his days working with it, is skittish about the idea.

“I can’t wait to start using it on my workflows,” he says of agentic AI, “but very cautiously at first for things that aren’t necessarily going to have huge financial impacts, customer-facing impacts, or potential legal consequences.”

On the other hand, proceeding too slowly could ultimately be just as risky, Joseph warns. It’s not hard to imagine a future a few years from now, he notes, in which agentic AI has turned MSPs from providers of outsourced IT management labor that use a lot of technology into providers of outsourced IT management technology that use a little labor for truck rolls, some Level 3 work, and not much else.

“Humans are the learning layer” in that vision, Joseph says. “They’re taking exceptions, training the models, improving the data wherever necessary, or adding additional context.”

And because humans are expensive, he adds, the more an MSP continues to rely on them when others don’t, the less competitive it’ll be. Which is why Joseph says that if he were running a managed services business right now, especially a big one with a lot of private equity money riding on its success, “I would be investing heavily in experimentation around agentic AI with almost an urgency to ensure that if there’s a survivability issue from a firm perspective, I can understand what that is.”

Fortunately, it’ll be a while before autonomous agents are sophisticated enough to render managed services as we know it a thing of the past. Probably.

“I’m willing to look as far as the first half of 2025 and say I don’t think anything’s going to happen there,” McDonald says.

But beyond that, who knows? Agentic AI’s moving too rapidly to say. Indeed, the first major release of Agentforce arrived in September. The second one debuts next week.

Pia’s feeling confident

That said, the dangers of moving too fast on agents do probably outweigh the dangers of moving too slowly for the moment. Gerwai Todd (pictured), CEO of AI-for-MSPs vendor Pia, is acutely aware of it and treading carefully as a result.

“While we’ll play around with that stuff in the lab to be on the cutting edge in our knowledge and our experience, I don’t see that technology getting introduced into our application for some time,” he says.

To the contrary, much as Tesla’s full self-driving mode is “intended for use with a fully attentive driver, who has their hands on the wheel and is prepared to take over at any moment,” Pia’s self-driving AI acts only under the watchful supervision of an experienced technician. Whenever it receives a service request, Todd explains, the system calculates a “confidence score” measuring its certainty that it both understands the user’s need and knows how to fix it.

“Without a 100% confidence score on a particular use case, you have to have human oversight,” he says.

Until very recently, only the simplest requests, like password resets, approached that bar. The platform update Pia announced in October closes the gap for more of the IT issues MSPs encounter regularly.

“I like to see the [average] confidence of what we’re capable of doing upwards of 90% or more,” Todd says. “We were slightly below that, and we’ve now achieved that goal.”

It wasn’t easy, though. That slim margin of progress followed a months-long refinement of Pia’s private, managed services-specific language model. According to Todd, however, the impact has been significant.

“All of this is dependent upon the inputs, but we’re seeing a lot of improvement in the results our customers are getting,” he says. Onboarding a new user, for example, takes as much as 45 minutes when done manually. “What we’re finding is we’re able to do it in a matter of five to ten minutes,” Todd notes.

The financial implications can be profound, he continues. Go back 30 years and the few companies offering remote technical services were lucky to support 100 users per technician. When the first generation of RMM applications arrived the next decade that figure began climbing to perhaps 200 users. Today, according to Todd, Pia has partners managing 350 users per tech.

“What that’s doing, in my opinion, is reinventing the entire MSP model to where they can run at better margins, to where they can reinvest in their business, and not be running at 10% on average but running closer to 20% net income on average,” he says.

Future updates, according to Todd, will expand both the range of problems Pia’s software can address and the precision with which it addresses them.

“VPN issues come in a lot of different flavors,” he says by way of example. “The more you can parse that down to something specific, the better off you are.”

As with everything about autonomous AI, however, figuring that out will take time. Which is fine, Todd notes, because it will also take time to overcome the even bigger, distinctly human challenge Pia and companies like it face at this early moment in hyperautomation history: acclimating MSPs to the uncomfortable sensation of taking their hands off the steering wheel to let software do the driving.

“With technology professionals, the longer and more time we get to work with a given technology, the better we’re going to be at understanding it,” Todd says. “It’s a learning process.”

Share

Helpt prefers humans

Of course, when it comes to AI, some people never learn. Those are Helpt’s kind of people.

“Anytime they know that they’re talking to an AI, the first thing they think about is how to circumvent that to get a real human being,” says co-founder David Sohn (pictured right). As long as users like that continue to exist, he adds, “a human’s always going to play a part in delivering warm and fuzzies to a person having an issue.”

At a time of increasing interest in chatbots and offshore labor, Helpt’s decidedly contrarian mission is to provide those warm and fuzzies via flesh-and-blood technicians based entirely in the U.S. to MSPs who believe technical know-how isn’t what produces loyal customers.

“It’s the experience that they get. It’s the support that they get. It’s the feeling that they get when they have a problem and need some assistance,” says fellow founder Matthew Pincus (pictured left).

Not that Helpt’s team lacks technical know-how, Sohn adds. Part of what the young company’s roughly 60 partners appreciate about Helpt is its knack for diagnosing issues that only humans—for now, at least—can handle.

“If you pull your car into a mechanic and say, ‘hey, it’s making a funny noise,’ the mechanic is going to have to deploy a ton of experience to understand what’s happening,” Sohn argues. “If you go to an AI chatbot right now and say, ‘hey, my car is making a funny noise,’ what kind of response do you think you’d get?”

Helpt (which raised $850,000 of seed capital over the summer) has three tiers of service for MSPs who aim to respond usefully. The first, called HelptNow, offers basic triage and ticket entry support from trained technicians. The second, named HelptFlow, adds immediate resolution of simple end user issues. The third, dubbed HelptFul, offers complete Tier-1 outsourced help desk support.

AI plays a role in all three services, but only on the back end as an automated source of documentation and knowledge base content. “We want our agents to focus on the caller, the issue, the ticket, and provide that great customer experience,” Pincus says. “The only way for them to be able to focus on what really matters to us is to feed them the information they need.”

Which, unlike a lot of what help desks do, is something AI handles well, according to Sohn. “This article’s going to come out and it’ll be like, ‘two dummies don’t think AI’s going to take over the world,’” he jokes. “That’s not what we’re saying. We think it will.” It’s just that real success in managed services will always require technically and emotionally intelligent work that only humans, ultimately, can do.

“We want to be the humans to do it,” Sohn says.

And as long as we’re talking about AI and agents…

Both topics come up in the latest episode of the podcast I co-host during an interview with Nadir Merchant, Kaseya’s general manager of IT operations products and reigning AI expert. Give it a listen here.

Also worth noting