The Future of AI Isn't Autonomous

There's a word circulating in every boardroom, every product roadmap, every vendor pitch right now: autonomous. As in, autonomous AI. Agents that run without human input. Systems that make decisions on their own.

It's a compelling word. It's also a distraction.

At HumanX 2026 in San Francisco, three of the most consequential operators in enterprise AI — Eoghan McCabe, CEO of Intercom; Shishir Mehrotra, CEO of Superhuman; and Jaime Teevan, Chief Scientist at Microsoft — sat down with Ian Martin of Forbes to talk about what agentic AI actually looks like when you build it, deploy it, and run a business on top of it. The consensus was clear: autonomy isn't the goal. Agency is.

And those are very different things.

What "Agentic" Actually Means (And Why It Matters)

Autonomous AI implies a system that operates independently, without meaningful human involvement. Agentic AI is something more nuanced, and, for enterprise leaders, far more useful.

Teevan put it simply: "When we think about an agentic future, it is the orchestration of work." Not replacement, not pure automation, but orchestration: humans and agents working together, with the agents handling scale, legibility, and process continuity in ways humans can't.

The distinction matters because it reframes how organizations should think about deployment. The question isn't "what can we automate?" It's "how do we redesign work so that agents and humans each do what they're actually good at?"

The Intercom Playbook: From SaaS to Frontier AI

Eoghan McCabe gave the room one of the most candid and specific accounts of an AI-led business transformation on record.

Intercom was founded in 2011 as a platform helping SMBs and mid-market companies communicate with customers — marketing, sales, service, success. Effective, profitable, human-powered. Then ChatGPT launched, and McCabe saw the writing on the wall: "It became patently clear that nebulous chatting to customers was a low-intelligence effort. It's kind of the lowest-level white-collar work out there."

The conclusion was uncomfortable but obvious: AI wouldn't just disrupt the workflow tools — it would disrupt the work itself.

So Intercom built Fin, a customer service AI agent. And then they made the harder call: they started cannibalizing their own SaaS revenue to grow it.

"Your precious revenue that you cling on to is certainly not yours to keep," McCabe said. The logic was simple, if nerve-wracking: there are trillions of dollars spent on salaries doing service, marketing, and sales work, and only tens of billions on the software to support it. Disrupting the work itself opened a far larger market.

Three years in, Fin has reached $100 million in revenue, representing roughly a quarter of Intercom's total. The new-business segment of Fin is growing nearly 5x. The SaaS side is growing in double digits; the AI side in triple digits.

And the workforce impact? More nuanced than most people expect. Customer service request volume at Intercom has increased 3x since Fin launched. The service team has neither grown nor shrunk. "AI is augmenting and adding to human work," McCabe said. "It's just going to serve our needs far better and create just higher standards and a lot more growth than we expected."

Why Vertical AI Wins

One of the most practically useful arguments of the session was McCabe's case for vertical, purpose-built agents over general-purpose models.

"We've learned that verticalized models beat the generalized models," he said. Intercom trained an open-weights model on service interactions specifically — and found it outperformed general models from major foundation labs on customer service tasks, while running faster and cheaper.

The broader insight: a sophisticated agent isn't a raw model in front of a customer. It's a collection of models, wrapped in logic, trained for a specific use case, with a clear distinction between deterministic and non-deterministic components.

"With the most sophisticated of these systems, you can mix deterministic and non-deterministic logic," McCabe explained. The non-deterministic elements — the language model — handle nuance and judgment. The deterministic elements handle your internal policies: refund rules, legal escalation triggers, compliance guardrails. Together, they produce an agent that is fast, consistent, and won't surprise you.

Most enterprise deployments, he noted, aren't there yet: "A lot of the agents in the market are not that. They are GPT wrappers."

Superhuman Go: The AI Superhighway

Shishir Mehrotra came to the conversation from a different angle, one focused on the infrastructure through which agents operate, not just the agents themselves.

Mehrotra runs Superhuman, which rebranded from Grammarly last year to reflect the breadth of the company's product portfolio: Grammarly, Coda (docs), a popular email client, and the newest product, Superhuman Go. The rebrand signals an ambition: to turn what was the world's most-used writing assistant into a platform for all agents.

The core idea is what Mehrotra calls the "AI superhighway" — the ability of Grammarly to work across a million unique surfaces per day, across web apps, desktop apps, and mobile apps. "We can understand what you're doing. We can annotate it in a way that's unobtrusive to you and the application, and we can make changes on your behalf."

Today, that superhighway carries one car: a grammar teacher. With Superhuman Go, Mehrotra wants to open the lanes to anyone building agents. The vision is a future where, as you write a customer email, you have a grammar agent on one shoulder, a sales agent on another flagging product mismatches, a support agent surfacing the customer's recent outage history, and a digital chief of staff flagging a scheduling conflict — all operating where you already work.

That's the agentic model: not an AI taking the wheel, but a fleet of specialists riding along, each doing exactly what they're built for.

What Microsoft Knows About Scale

Jamie Teevan brought the perspective of an organization already running agentic systems at a scale most companies can barely imagine. And her core observation reframes the competitive opportunity for enterprises thinking about deployment.

"When you look at how work really gets done, it's people working together," she said. The agentic future isn't about individual agents replacing individual workers — it's about designing systems where groups of agents and groups of humans collaborate simultaneously.

One of Teevan's more provocative points: in many conversations about AI, people focus obsessively on what makes humans unique. She'd rather flip it. "It's really interesting to look at what's unique to AI models, because that actually is exactly how you go and change the systems."

Her examples were specific. Models are legible in ways humans aren't — you can externalize their prompts, their grounding context, their evaluation measures, their outcomes. Models operate at a scale that humans physically cannot. A model could synthesize the perspectives of every person in a 1,000-person room; a human panelist cannot. These aren't threats to human value. They're design parameters.

The implication for enterprise AI leaders: stop asking what AI might take from your workforce, and start designing systems that exploit what AI can actually do that humans can't.

The Accountability Problem No One Is Solving

The panel didn't pretend everything is figured out. The most honest moment came near the end, when the conversation turned to governance.

"Ultimately, humans can hold accountability in a way that AI can't," one panelist observed. "Having the structures where people are responsible for the outcomes is foundationally important, no matter what guardrails we build."

This is the unsolved equation in enterprise agentic AI. The systems are increasingly capable of autonomous execution. The legal and organizational structures to manage that execution lag behind. Building sophisticated agents is now tractable. Building the accountability architecture around them is still very much a work in progress.

On the job question — more or fewer skilled jobs in five to ten years — all three panelists answered the same way: more. Teevan added one qualifier that stayed in the room: "Metacognitive level of the work." The jobs that grow won't be the ones agents do well. They'll be the ones that require knowing how to work with agents — designing them, overseeing them, directing them.

Watch the full session on-demand. 

Register now for HumanX 2026 Amsterdam (September 22–24, 2026) or HumanX 2027 Las Vegas (March 7–10, 2027).