Dennis Mortensen knows something about bad timing. For years, he and his team at x.ai did the unglamorous work of teaching machines to schedule meetings—hand-labeling somewhere in the neighborhood of 33 million data points, one tedious example at a time. They built neural networks from scratch. They pitched investors on a future where AI assistants would eliminate calendar Tetris. Between 2014 and 2021, they raised approximately $44.23 million to make it happen.
Then Bizzabo acquired the company's assets in June 2021. The price? Undisclosed, though anyone watching the deal understood it for what it probably was: an acqui-hire dressed up as an acquisition. By October 31 of that year, x.ai's scheduling product was dead.
Here's the thing, though—Mortensen wasn't wrong. He was just a decade ahead of schedule.
Today, the market for AI agents that x.ai helped pioneer is projected to reach an estimated $7.63 billion in 2025, per Grand View Research's March 2026 analysis. Companies building what Mortensen tried to build are raising rounds at valuations that would have sounded like science fiction in 2014. Sierra, which builds enterprise customer experience agents, hit $100 million in annual recurring revenue by November 2025. That's less than two years from launch to nine figures. Decagon closed a $250 million Series D in January 2026 at a $4.5 billion valuation. Cognition—the startup behind Devin, the coding agent that's become something of a developer obsession—raised over $1 billion at a $25 billion pre-money valuation this past May.
What changed between Mortensen's failure and their success? The infrastructure finally showed up. Models got good enough to actually work. And crucially, the market figured out what these things were supposed to do.
When There Were No Shortcuts
To understand the timing gap, you have to understand what building AI agents looked like before large language models became commoditized. There was no GPT API to call. No Anthropic. No pre-trained foundation models you could fine-tune over a weekend with a credit card and some coffee. Just you, your data scientists, and millions upon millions of examples of human behavior that someone—usually a small army of someones—had to manually encode into training data.
"We had to hand-label everything," Mortensen said in a June 2025 podcast, his tone still carrying a trace of fatigue from those years. The x.ai team labeled roughly 33 million elements: emails, calendar invites, the natural language patterns people use when they're trying to coordinate schedules across time zones and conflicting priorities. Every edge case demanded new training data. Every fresh context meant more examples.
The technical debt piled up, sure. But there was another problem, arguably more intractable: nobody knew what to make of an AI assistant that operated autonomously. The error rate—or more precisely, the perceived error rate—became a wall they kept running into. Mortensen noted in a January 2026 interview that even when the system worked correctly, users often assumed it had failed because they didn't understand its logic. Building the product was hard enough. Convincing people to trust it turned out to be harder.
By 2021, the path forward had narrowed to nothing. Bizzabo picked up the pieces in June. Four months later, the lights went off for good.
The Infrastructure Moment
Then came 2022 and 2023, and with them, ChatGPT—followed quickly by the rapid commoditization of frontier models. Almost overnight, the capabilities that had required millions in R&D and years of painstaking data work became accessible via API. Anthropic launched its Computer Use beta in October 2024. OpenAI followed with Operator, its computer-using agent, in January 2025, then added computer-use tools to its Responses API by March. Microsoft shipped Agent Framework 1.0 this past April, pulling together AutoGen and Semantic Kernel with Model Context Protocol interoperability baked in.
It wasn't just that the models got better. They got usable. Instruction-following actually worked. Developers stopped building custom NLP pipelines and started writing prompts instead. The barrier to entry didn't just lower—it collapsed.
Gartner's data captures the acceleration. This past May, the firm reported that only 17% of organizations had deployed AI agents so far—but more than 60% expected to deploy them within the next two years. That's the steepest adoption curve Gartner has tracked, possibly ever. By April, the firm was already warning about "AI agent sprawl," projecting that the average Fortune 500 company would be running upwards of 150,000 agents by 2028, up from fewer than 15 in 2025.
The ground-level numbers tell the same story. Gallup's Q1 2026 survey found that half of U.S. employees now use AI at work, with 28% using it daily or weekly. Morgan Stanley's earnings call transcript analysis showed that a quarter of S&P 500 companies cited quantifiable AI impact in Q1 2026—nearly double the 13% from Q1 2025. This isn't hype cycling through conference keynotes anymore. It's showing up in the financials.
Customer Service: The Beachhead
If you're looking for where agents landed first, customer operations is the obvious answer. Klarna disclosed in SEC filings that its AI assistant handled 69% of customer service chats in the twelve months ending June 30, 2025—work the company pegged as equivalent to over 700 full-time employees, with $39 million in cost savings for 2024 alone. Intercom's CEO claimed in March that the company's Fin agent was approaching $100 million ARR with roughly 8,000 customers. Around the same time, Intercom introduced Fin Apex, a vertical model it says handles more than two million conversations per week and outperforms general-purpose LLMs in support scenarios.
Zendesk made perhaps the boldest bet at its May Relate conference, launching what it calls an "Autonomous Service Workforce" with outcome-based pricing. Customers pay per verified resolution, not per seat or per interaction. It's a statement of confidence—or at least a statement that the company believes agents are reliable enough to tie revenue directly to results.
Software development is the other domain where agents have gained real traction. GitHub announced its coding agent in May 2025, an asynchronous system that opens pull requests without waiting around for human approval. A January 2026 study posted to arXiv estimated that between 15.85% and 22.60% of GitHub repositories showed signs of coding agent adoption in early 2025. By May of this year, Stack Overflow reported that 59% of surveyed developers were using agents in their workflows—nearly double earlier levels. And then there's Cognition, which closed that billion-dollar round at a $25 billion valuation, a number that would've been unthinkable just a couple of years ago.
Platform players are moving fast, too. ServiceNow's Now Assist surpassed $600 million in annual contract value and is targeting over $1 billion in 2026, according to the company's January earnings call. Salesforce is pushing Agentforce despite some profitability trade-offs flagged in December coverage. Google Cloud, Microsoft, and OpenAI are all shipping multi-agent orchestration frameworks and no-code builders. Even Apple overhauled Siri at WWDC 2026, adding what the developer docs describe as "Dynamic Profiles for multi-agent workflows." That's a lot of enterprise motion in a very short window.
The Governance Problem Nobody Wants to Talk About
The infrastructure arrived. The governance layer? Still scrambling to catch up. A Cloud Security Alliance survey from April found that 82% of organizations have unknown AI agents lurking in their environments, and 65% have already experienced agent-related incidents. NIST ran a large-scale red-team competition whose March takeaways highlighted vulnerabilities across tool-use, coding, and computer-use scenarios. Gartner's April guidance prescribes six steps for managing agent sprawl, emphasizing centralized registries and lifecycle controls—the kind of unglamorous plumbing work that doesn't generate headlines but becomes critical once you're running thousands of agents.
Regulatory timelines, meanwhile, are compressing faster than most companies anticipated. The EU AI Act's general application date lands on August 2, 2026, though specific obligations begin on different dates, with transparency and governance requirements phasing in over time. General-purpose AI model rules began last August, though existing models have until August 2027 for full compliance. In the U.S., the FTC has ramped up enforcement. It banned Air AI from marketing business opportunities in March over deceptive claims. In May, it levied a $930,000 settlement against Cox Media Group and others for false assertions about "AI listening" capabilities.
A widely cited report from MIT Sloan and BCG—published in late 2025 and referenced throughout early 2026—found that 35% of surveyed companies had deployed agentic AI by 2023, with another 44% planning to. But the same research flagged a governance tension: companies are deploying agents faster than they're building the frameworks to manage them. BCG's June commentary put it bluntly: AI is reshaping jobs faster than companies are reshaping work.
What Good Timing Looks Like Now
So what constitutes good timing in 2026? Mortensen's reflections, scattered across multiple recent interviews, offer a framework: validate the problem before you automate the solution. He's emphasized the value of concierge MVPs—manual services that prove demand and surface edge cases before you write a line of ML code. The technology exists now, yes. But the hardest questions remain behavioral and organizational, not technical.
Forrester's June report on agentic AI describes what it calls a "chase-catch gap" between ambitions and production value, urging companies to focus on orchestration and control rather than simply spinning up more agents. Gartner projects that by 2028, 45% of CIOs will lead AI agent systems outside traditional IT—a signal that these systems are shifting from experiments to operational infrastructure. The Stanford AI Index 2026 documents rapid enterprise diffusion while noting that "agentic systems" remain an emerging category with tooling that's still, to be generous, immature.
The market has crossed a threshold, though. Fifty percent employee adoption. Billions in committed revenue. Platform vendors shipping agent-first products as core offerings, not side projects. But perhaps the clearest signal is simply the money: Sierra valued at $10 billion, Decagon at $4.5 billion, Cognition at $25 billion pre-money. Investors are pricing in a future where autonomous agents are as ubiquitous as SaaS became in the 2010s.
Mortensen built too early. But he also validated the category, proved the use case, and absorbed the lessons that come from trying to sell a product the market isn't quite ready to understand. The companies scaling today are riding infrastructure he could only sketch in pitch decks—LLMs that follow instructions reliably, APIs that simulate computer use, frameworks that orchestrate multi-agent workflows without falling apart. The lesson isn't that he was wrong about where the market was headed.
It's that the distance between visionary and viable is measured in infrastructure, not just imagination. And in 2026, that infrastructure finally exists. Whether it's mature enough to support the weight being placed on it? That's a question we're all about to find out together.
