How Vercel runs on AI agents

Tom Occhino presented at SaaStr AI Deploy, and the most interesting part wasn't a product announcement. It was a set of operating numbers from a company that built hundreds of AI agents for itself before selling the tools to anyone else.

The concept that frames everything Occhino says

Every minute a company spends configuring infrastructure, building bespoke frameworks, and wrangling DevOps is energy that isn't going into product, pipeline, or customers. Occhino calls this "undifferentiated heat loss."

That frame matters now because most of what teams are about to build for agents is also generic. Logging, tracing, orchestration, deployment pipelines, rollback, access controls. Every team that purpose-builds all of that is spending energy on something that doesn't differentiate them. Vercel's pitch is that it absorbs that tax so product teams spend their energy on differentiated product work, not plumbing.

How the model of software itself changed

Agents invert the traditional software model. The UI shrinks to a review surface. The real work happens in headless, autonomous software running on its own.

The more important difference: agents act autonomously on your behalf. They wake up on a trigger, an inbound lead, a fraud signal, a monitoring alert, take action against the rules and goals you gave them, and escalate to a human only when they have to.

Occhino's framing is that every company needs at minimum two agents, and these are a starting point, not a roadmap item. One for employees, sitting inside the collaboration tools they already use, handling internal knowledge and workflows so people stop navigating legacy systems. One for customers, handling support, transactions, and self-service across any channel.

The production numbers

This is what running their own agents looks like at Vercel in production:

Marketing: 96% of marketing content is produced by agents. The team that exists isn't gone; it's doing the work agents can't do well yet. The agents produce the volume; the humans set the bar, approve the edge cases, and handle judgment calls that require context an agent doesn't have.

Support: 93% of support interactions are handled without a human in the loop. At Vercel's scale, that's a structural change in what a support team looks like. The 7% that reaches humans tends to be genuinely complex, which means the humans are running harder problems rather than routing tickets.

Sales development: The SDR function was reabsorbed into a broader go-to-market team. Agents handling inbound qualification and routing absorbed enough of the SDR workload that the team structure changed. Occhino's framing isn't that people were fired; it's that the function merged because the volume problem that SDRs existed to solve was no longer a volume problem.

Vercel sells to technical buyers, but the underlying capability is available to anyone.

"Agents as a service" don't work

Vercel's conclusion after two years and hundreds of agents is that you can't buy a one-size-fits-all agent off the shelf any more than you can buy a one-size-fits-all website. Every team builds its own.

That doesn't kill the build-versus-buy question. It moves it. You don't buy the agent. You buy the infrastructure and the tools that make building the agent cheap, fast, and safe. Vercel uses Vercel to build Vercel, and that dogfooding is the entire argument for the platform.

For more on how AI is reshaping go-to-market execution at the infrastructure layer, our breakdown of how AI agent stacks are reshaping GTM at SaaStr, Owner, and Klaviyo covers adjacent moves from operators building in the same direction. And if the trust dynamics around AI-assisted content are a live question for your team, the AI strategy trust problem is worth reading alongside this.

What operators building inbound should take from this

The direct application for founders and operators running LinkedIn-led inbound isn't "build 96% of your content with agents." It's a more specific reframe.

The bottleneck isn't the quality of the insight. It's the finding-and-executing work that eats the hours. Scrolling to find the right post to comment on, identifying which conversations your buyer's network is actually in, figuring out which comments got traction and why. That's the same undifferentiated work Occhino describes. It taxes without differentiating. The operators who win at LinkedIn inbound consistently are the ones who've compressed that tax so the time they spend on the platform is the high-judgment part, not the logistics.

Vercel's story is a data point on where that trajectory leads at production scale.

The headcount reframe

The SDR team reabsorption is the number that tends to generate the most friction in rooms with sales leaders. Occhino's framing is worth borrowing: the people didn't disappear, the function merged.

SDRs exist to solve a volume problem. When agents solve that, what remains is judgment work: nuanced enterprise qualification, relationship-building, conversations that require genuine human read on what the prospect cares about. That work still requires people. It just doesn't require a team organized around volume.

For operators thinking about their own growth function, the question isn't "should I replace my outreach effort with AI." It's "which parts of my current process are generic, and what would I do with the time if those parts weren't my problem."

Vercel's two years of production data suggest the answer is: a majority of the task volume, across marketing, support, and sales development.