The Rise of the Agent Manager: How AI Is Redefining Work, Not Replacing It

Estimated Read Time: 6 minutes

The conversation around AI and employment has calcified into two opposing camps. One side predicts a white-collar apocalypse. The other promises frictionless human-machine collaboration. But here's what neither camp seems to discuss: what happens when you're actually inside a Fortune 500 company deploying AI agents right now, today, at scale?

Nancy Xu knows. As Salesforce's AI strategy leader—a role she stepped into after the company acquired her recruiting-focused AI startup Moonhub—Xu sits at an unusual vantage point. She's built vertical-specific agents herself. Now she helps some of the world's largest companies implement them across sales, service, finance, and recruiting.

Her perspective cuts through the noise: this isn't about replacement. It's not even really about augmentation. It's about something more fundamental—a shift in what we mean by "work" itself.

The question isn't whether AI will change jobs. It's whether we're ready to become managers of intelligence rather than executors of tasks.

Here's a problem most consumers never think about: when you're an enterprise deploying AI agents, which model should power which task?

It matters more than you'd think. Upgrade to the wrong model version, and you might actually degrade your agent's performance rather than improve it. Deploy a general-purpose model where you need specialized healthcare reasoning, and you've just created an expensive liability.

This is where Xu's team comes in. They run continuous internal benchmarks across OpenAI's latest releases, Anthropic's Claude models, and Google's Gemini, giving customers the confidence that when they upgrade, they're actually upgrading. "A lot of times our customers look to us as a guide for what models they should be using," she explains.

It's tedious, technical work—the kind that doesn't make headlines. But it reveals something crucial about where we are right now: we're not in a one-model-wins-all world. Even within a single platform like Salesforce's Agent Force, customers use a variety of different models. Sometimes the choice comes down to business relationships or cost. Increasingly, though, it's about specialized models trained for specific domains that outperform their general-purpose cousins in particular niches.

The AI landscape is fragmenting, not consolidating. And the companies that win will be those that can orchestrate across this complexity, not those betting everything on a single foundation model.

Ask Xu about her favorite use cases, and she doesn't point to flashy demos or proof-of-concepts. She points to ADCO, a staffing company with a deep partnership with Salesforce.

Here's what's happening there: if you're a candidate communicating with ADCO after business hours today, there's a 50% chance you're talking to an agent. Not a chatbot that routes you to FAQs. An agent that successfully resolves your inquiry—answering questions, updating your status, scheduling next steps.

For Xu, this is the future crystallizing: "I really believe in this vision of making everyone an agent manager. If you're on a customer success team, you can now manage agents. You previously didn't have the resources to delegate the work that you wanted to, and now you can make that possible."

But becoming an agent manager isn't just about offloading tedious tasks. It's about a fundamental cognitive shift—from thinking about the how of work to thinking about the why.

"It's no longer about necessarily all the tactical details of how you do something," Xu explains. "We'll elevate people to think about the why. What are the important strategic things we need to solve? Let me delegate the how to the agents to actually do it for me."

She offers a film production analogy that clarifies the distinction: "If you were on a movie shoot, you were the producer before and now you're the director. It's a different mindset."

The producer handles logistics—scheduling, budgets, resources. The director shapes vision and makes creative decisions. Same project. Different level of abstraction.

There's a concern lurking beneath the enthusiasm: are we just replacing one form of tedium with another? Instead of manually processing customer inquiries, will we spend our days tediously prompting and correcting agents?

Xu doesn't think so, and her reasoning is tied to where agents are headed. "What does agents look like in one year or two years? I think we're going to move to a world where agents become more long-horizon. They can work on longer objectives, longer tasks for you."

Her example is concrete: "You should be able to tell your agent, 'Hey, I want you to increase my CSAT score by X number of points within 6 months.' And your agent will help you figure that out in simulation with your evals and actually get you to that state."

In that future, the human's job isn't to figure out how to increase the customer satisfaction score—it's to decide that increasing it matters, and to make strategic decisions about the tradeoffs the agent should navigate along the way.

The agent becomes less like a tool and more like a thought partner. Not something you micromanage, but something you direct.

When asked about labor market disruption, Xu doesn't dismiss the concern. "We do have to acknowledge that work is transforming. A lot of people say the industrial revolution and the farming revolution—those are great analogies, but the truth is AI is moving a lot faster than some of these other previous revolutions."

The speed matters. The industrial revolution gave society generations to adapt. AI is giving us years, maybe less.

But Xu's optimism is grounded not in denial of disruption but in the belief that the transformation can be managed: "Our job is to enable our customers to uplift everyone in their team so that they can become managers of agents."

The companies that thrive won't be those that eliminate headcount most aggressively. They'll be those that successfully transform their workforce from task-executors to agent-orchestrators—people who understand what needs to happen and can effectively direct artificial intelligence to make it happen.

As agents take on more consequential work, a thorny question emerges: when something goes wrong, who's responsible?

Xu's answer is refreshingly specific. It depends on the failure mode.

"If it's a copyright issue because your foundational model was trained on a particular type of data, that really sits at the model layer," she explains. "If it's an issue related to, 'Hey, you're not properly anonymizing your end users' information and you're sending it to these third party models without their permission,' that's on the platforms."

This is why Salesforce invested heavily in what they call a "trust layer"—systems that handle data anonymization, access controls, and governance before information ever reaches a third-party model. The most successful agent implementations don't bolt on trust as an afterthought. They design for it from day one.

And they require buy-in at the business level, not just from IT. "The most successful agent implementations we see come from business line leaders who have buy-in," Xu notes. "They're working with the developers, but there's buy-in from the business leaders on what the end goals for the business are."

Translation: if your VP of sales doesn't understand why agents matter and what they're trying to achieve, your expensive AI implementation is probably going to flounder.

How do agents actually spread through organizations? Top-down mandate or grassroots groundswell?

The answer is both—but in different contexts.

Complex agents for service centers or sales operations require centralized, top-down efforts. "It takes a large team effort to build an agent like that," Xu explains. "Usually there's a centralized effort that's building the agent, and then it's about empowering individuals to get on board and use that agent."

But simpler, task-oriented agents—document summarizers, case classifiers—can spread more organically. Xu points to Slack's AI features as an example: "So many people love Slackbot. They use Slack and it's not someone saying, 'Hey, everyone in the organization has to use it.'"

The pattern mirrors previous enterprise software waves: complex systems need executive sponsorship and change management. Lightweight tools can spread virally if they solve real pain points.

As agents become more sophisticated, definitional boundaries start to dissolve. Salesforce's Agent Force supports something called A2A—agent-to-agent communication. When one agent calls another agent to complete a subtask, is that one massive agent with sub-agents? Or is that two discrete agents talking to each other?

"I think ultimately the lines become: who are the core platforms that you're using to build your agents and how do they engage with each other," Xu says.

For Salesforce, the strategic bet is clear: build agents that are open, that can communicate with other platforms' agents, and position Salesforce as the orchestration layer. It's a vision of the future that's less about Salesforce-versus-Microsoft-versus-Google and more about creating an interoperable ecosystem where agents from different providers work together seamlessly.

The value shifts from model performance alone to integration, orchestration, and trust. Raw intelligence becomes a commodity. How you deploy it matters more.

The future of work isn't humans versus machines. It's humans learning to operate at a higher level of abstraction—defining objectives rather than executing steps, making strategic decisions rather than tactical ones, directing intelligence rather than being the intelligence.

This transition will be faster and more disruptive than previous technological revolutions. But companies like Salesforce are betting the path forward runs through elevation, not elimination. Not replacing workers but transforming them into agent managers.

The question for businesses and individuals alike: are you ready to shift from doing the work to directing it?

Because the agents are here. They're getting more capable by the month. And the role that remains isn't about having all the answers.

It's about knowing which questions to ask.

This article is based on insights from Nancy Xu, AI Strategy Leader at Salesforce and founder of Moonhub, discussing the practical realities of deploying AI agents at enterprise scale.