Two and a half years ago, any startup coming out of Y Combinator had a predictable shopping list. Salesforce or HubSpot for customer relationships. A call recorder. An email sequencing tool. Data enrichment software. Maybe a forecasting platform. By the time a founder finished assembling the stack, they were juggling fifteen or twenty subscriptions and probably hiring someone just to keep it all running.
The current generation of YC founders is skipping most of that entirely. In my view, the behavior of AI-native YC founders is a prelude to how the rest of the world will be operating in 6 to 18 months from now. So it’s worth paying attention.
From what I’m hearing, nearly zero percent of the accelerator’s most recent cohorts use tools like Salesforce or HubSpot — down from close to 100 percent just three or four years ago. But the shift isn’t just about which vendor they’re choosing. It’s about whether they need the software at all.
These founders aren’t replacing one CRM with another. They’re bypassing the CRM, the person who maintains it, and the workflows that fed into it — with AI agents that do the work themselves.
The End of Software as a Category?
For two decades, the enterprise software playbook was simple: identify a business process, build a tool to manage it, charge per seat. The result was an explosion of point solutions. Each one required configuration, maintenance, and someone to operate it.
AI agents collapse this model entirely. They don’t need interfaces designed for humans to navigate. They don’t require manual data entry. And they don’t charge you for seats when the “user” is another piece of software.
“Before AI, software fragmented into point solutions because building was hard — you had one company per feature,” says Keith Peiris, founder of Lightfield, an AI-native CRM. “Now building is fast enough that you can bring everything back together.”
Investors see this as a once-in-a-generation opening. “There are rare technology shifts where core systems of record, like the CRM, are ‘up for grabs’ because of a new technology architecture,” says Seth Rosenberg, partner at Greylock. “This happened with on-prem to cloud, which created Salesforce.”
The difference now is that AI-native systems aren’t just better interfaces on the same database. “All data — structured and unstructured, internal and external — needs to be in one place,” Rosenberg says, “and reasoning models can then automate work with full context.”
Peiris saw this shift early. He previously built Tome, an AI presentation tool that scaled to 25 million users and raised $81 million from investors including Reid Hoffman. When he recognized that improving AI models might commoditize his own product, he shut Tome down with $30 million still in the bank.
“We kept running into this limitation where the AI just didn’t know enough to be useful,” he explains. “It didn’t understand our users, their company, or the context of what they were trying to communicate.”
“It takes a lot of courage to pivot from something that is ‘working,’ but with significant questions on durability, to build something that has the shot to be one of the iconic companies of the AI era,” Rosenberg says.
That context problem — the reason AI assistants often produce generic outputs — is exactly what the new generation of agentic tools solves. They don’t sit on top of your data. They are your data layer, capturing information automatically and acting on it without waiting for human instruction.
Not Just the Software — The Work Around It
The most expensive part of legacy software was never the subscription. It was everything else.
“The real cost of using traditional SaaS is not with the software,” Peiris says. “It’s with the people you need to keep on your payroll to set it up for your business and keep it running.”
A startup stitching together fifteen different tools needs someone to configure them, maintain them, and make them communicate. Add that salary to the subscriptions, and the true cost of a software stack can double. “I’ve talked to founders who started at what they thought was $50,000 a year and somehow ended up at $1 million,” Peiris says.
But AI agents don’t just eliminate the software. They eliminate the work that flowed in and out of it.
Tyler Postle, co-founder of Voker.ai, spent months manually logging customer conversations into HubSpot. “Using HubSpot, I was a data hygienist,” he says. “Using Lightfield, I’m a closer.”
The difference became concrete in a single weekend. “In one two-hour working session, Lightfield’s agent helped me revive over 40 opportunities and leads I had fallen behind on for six months,” Postle says. “Within two days, 10 of those became active opportunities that moved to proof of concept. This wasn’t possible in HubSpot.”
It wasn’t possible because HubSpot — like most traditional software — is a system of record. It stores what you put into it. AI agents are systems of action. They do the work, then record that it happened.
The Six-Person Team That Disappeared
The old go-to-market playbook required layers of specialization.
“You needed an SDR, an account executive, an account manager, a customer success person, a sales engineer, and maybe a product marketer supporting them,” Peiris says. “That’s at least six people before you’ve really started.”
Six mid-level hires in a major metro run $600,000 or more annually before benefits. Founders using AI agents to handle meeting prep, follow-up emails, CRM updates, and pipeline management don’t need that headcount.
“What I’m seeing with the companies using AI-native tools is one person doing the work of that whole team,” Peiris says. “Not because they’re working unreasonable hours, but because the software is actually doing the grunt work.”
The founder shows up to a call and the context is already there. The follow-up email drafts itself. The CRM fills itself in. Dropped leads resurface automatically.
Radu Spineanu, co-founder of Humble Ops, points to a specific example. “The killer feature is asking, ‘who haven’t I followed up with?’ Most deals die from neglect, not rejection,” he says. “Lightfield catches these dropped threads and can draft and send the follow-up immediately. That’s prevented at least three deals from going cold this quarter.”
His response times dropped from weeks to one or two days. Prospects noticed.
Why Legacy Vendors Can’t Catch Up
Salesforce has announced $16 billion in AI investments. HubSpot and others are rushing to add AI features. But the architectural problem may be insurmountable.
Traditional CRMs were designed for humans to navigate — dashboards, dropdown menus, search bars. AI agents don’t need any of that. When the interface is built for human eyes, and the user is a machine, every click becomes friction.
“Salesforce and HubSpot are built for a different era,” Spineanu says. “They assume you have dedicated ops people to configure workflows and maintain data hygiene. We don’t, and neither do most early-stage companies.”
The deeper issue is the data entry model itself. Traditional CRMs require consistent manual input. “The moment you skip a day, it becomes useless,” Spineanu says. “I’d rather have a system that captures 80% automatically than one that theoretically captures 100% but relies on perfect human behavior.”
The gap may be too wide to bridge. An AI-native CRM built from scratch looks fundamentally different than a relational database with bolt-on AI features. “Software has changed rapidly from a tool for humans to enter data to a tool that can complete work,” Rosenberg says. “If you don’t adapt quickly to become AI-native across all areas of the economy, including SaaS, you’re taking on significant risk.”
Postle is blunt about where this leads: “HubSpot and Salesforce are going to lose because they aren’t AI-native, no matter how much they try to pretend to be. The fact that they have integrations already isn’t a moat.”
Beyond CRM
The pattern playing out in CRM seems like it’s about to repeat across every category of business software.
Any tool that primarily serves as a system of record — storing data that humans enter and retrieve — is vulnerable. Expense management. Project management. HR systems. Support ticketing. The moment an AI agent can do the underlying work, the software layer becomes unnecessary overhead.
Dan Rose, partner at Coatue, sees this in how companies are already using Lightfield beyond sales. “It can be helpful to product and engineering teams who can now query the database to inform their roadmaps,” he says. “Lightfield is more than just a sales database — it’s a customer intelligence layer.”
That framing — intelligence layer, not software tool — may be the key distinction. The winning products won’t be better versions of existing categories. They’ll be systems that understand context deeply enough to act autonomously, collapsing multiple tools and multiple roles into a single capability.
The New Math
The startups forming habits around AI agents today are building fundamentally different cost structures than the companies that came before them.
They’re not just saving on software subscriptions. They’re saving on the administrators who configured the software, the analysts who pulled reports from it, and the junior staff who fed data into it. They’re reaching profitability faster, stretching runway further, and competing with larger players without the overhead that used to define growth-stage startups.
“A five-person company with the right tools can genuinely outperform a bloated fifty-person sales organization,” Peiris says, “because they’re moving faster, they’re closer to their customers, and they’re not losing information in the handoffs between twelve different systems.”
Peiris offers founders a simple test to know if they’re on the wrong side of this shift: “When was the last time your CRM helped you understand something about your customers that you didn’t already know? For most people, the answer is never. They’re putting data in, manually, after every call, and they’re not getting anything back.”
If the tool creates work instead of doing work, the price is too high — regardless of what the invoice says.
For anyone still assembling the old stack, the math is worth revisiting.
Read the full article here