When Aliisa Rosenthal joined OpenAI, the company was pulling in a couple million dollars in revenue and had no enterprise product. By the time she left, that line item had grown to several billion. The path between those two points was not a triumph of salesmanship. It was, by her own description, a series of expensive mistakes that she now wants other founders to avoid.
“Build the machine first before you build the team,” Rosenthal said, speaking on the AI Engineer podcast. “Figure out what you can automate, figure out where the bottlenecks are and what’s broken, and then add humans on top of that to get through those bottlenecks.”
That framework — software before headcount — is the organizing principle of her entire playbook. It sounds obvious in an industry that worships automation. But Rosenthal’s authority comes from having watched the opposite sequence fail at one of the most valuable companies in the world.
The Nine-Month Silence That Handed Deals to Microsoft
ChatGPT launched at the end of 2022 with zero enterprise features. Rosenthal spent the next nine months trying to get the technical team to build basic capabilities — single sign-on, NDAs, invoicing. Meanwhile, she and four sales reps were drowning in roughly 10,000 inbound leads per day.
The backlog was not just unmanageable. It was actively destructive. When the enterprise product finally shipped, Rosenthal found that the delay had already converted potential customers into competitors’ wins.
“Nearly every company I approached was like, ‘You never got back to me, and I went out and I brought Microsoft Copilot,'” she said.
The lesson she draws is that silence is not neutral in sales. It is a signal to the buyer that you are not serious. Her retrospective fix is simple and cheap: even if you cannot respond substantively, send an automated acknowledgment. Tell the lead you hear them, the enterprise product is not ready, they are on the waitlist, and ask what they are looking for. The acknowledgment alone preserves the relationship.
Her broader prescription for the inbound problem is equally practical. Add fields to sign-up forms — phone number especially — because when you eventually build automation to follow up, you need the data. On the outbound side, she names tools like Clay and Nooks that handle research, sequencing, and dialing, freeing humans to focus on conversations that actually require them.
Self-Serve First, Enterprise Second
OpenAI did the launch sequence backwards, Rosenthal acknowledges. After nine months of building enterprise features, the company shipped a high-end, expensive product oriented toward the large enterprises that had been loudest in demanding it. Four months later, in January 2024, OpenAI released a self-serve version.
The result was immediate and unambiguous: “The self-serve motion completely cannibalized the enterprise business.”
Customers did not want to talk to a salesperson. They wanted to swipe a credit card and start using the product. The enterprise sales reps found themselves competing against a cheaper option that customers preferred.
“We should have launched with self-serve first, listened to feedback from our customers on what they weren’t getting from the self-serve product, and then figured out how to build a more expensive enterprise offering, and then hire the team to sell it,” Rosenthal said.
The generalized lesson runs through everything she advises: let customer behavior, not internal assumptions, dictate where human touch is needed. Automate everything. Add people only where the machine breaks.
The Price of Pricing by Cost, Not by Market
OpenAI’s pricing mistake is the most concrete cautionary tale in the episode. The company set ChatGPT Enterprise at $60 per user per month. That number was derived from what it cost OpenAI to serve the product, not from any analysis of what the market would bear. As the first mover, there was no benchmark.
Then Microsoft Copilot, Gemini, and Anthropic entered at lower price points. OpenAI realized it had overpriced.
The fix was a structural shift: a base license fee plus usage-based pricing. The effect was dramatic. At $60 per user per month, organizations bought licenses only for subsets — developers, investors, specific teams. At the lower threshold, “it spread like wildfire” across entire companies.
MetricValueInbound leads per day at OpenAI (2023)10,000Time from ChatGPT launch to enterprise feature approval9 monthsMonths between enterprise launch and self-serve launch4 monthsInitial ChatGPT Enterprise price$60 per user per monthOpenAI revenue when Rosenthal joinedA couple millionOpenAI enterprise revenue a few years laterSeveral billion
Rosenthal acknowledges the emerging backlash against usage-based pricing — companies fear unpredictable costs. Her mitigation is a dashboard with monthly spend caps and per-employee caps. Most companies never actually use them, she notes, but they want the reassurance that the control exists.
Don’t Give Buyers Homework
The single most repeated principle in Rosenthal’s playbook is friction reduction. “Just make it easy on your buyers,” she said. “Don’t give them homework. Don’t make them go and fill out a form… You do as much of the work for them. You hand-hold them through that.”
This manifests in specific practices. Never send a buyer away with a list of five things to do and wait for them to come back — you lose control of the sales cycle. Instead, run hackathons in their office, do live demos on Zoom, do the work alongside them. The buying process itself should have fewer pricing options, faster paperwork, simpler approval flows.
The pilot, in this framework, is the ultimate form of friction. Rosenthal calls “pilot hell” a common founder complaint — pilots that never convert, or that convert but require restarting the sales process from scratch. Her advice is blunt: pilots should be the exception, reserved for the highest-value deals, never the default.
“As soon as you give someone access to your product, you’re giving away a lot of power and leverage in the deal cycle,” she said.
She offers a menu of alternatives. If a buyer says they need to validate the product works in their environment, introduce them to an existing customer. If they need to test with their own data, run an eval over Zoom without handing over access. If procurement requires a trial, offer a 90-day opt-out clause in a signed contract — which shifts the burden of validation onto them because the clock is ticking. Or, if they insist on a longer evaluation, offer a one-year POC instead of the standard three-year contract. Every alternative preserves leverage by keeping the contract signed.
Security Is Where Deals Die
The least glamorous but most common deal-killer, according to Rosenthal, is security. “Security is where deals tend to stall out and die over time,” she said.
Her prescription is automation. Trust portals can auto-sign NDAs and distribute penetration test results and security documentation. AI products can auto-fill security questionnaires. When a prospect insists on a two-hour security call, push back: direct them to the trust portal and ask them to come back with specific gaps they could not find. The default should be automation, with human security calls as the exception.
This is not a minor point in her framework. Security review is one of the least differentiated parts of enterprise sales — every company asks the same questions, requests the same documents, and runs the same checks. Automating it is one of the highest-leverage ways to compress the sales cycle.
The Pull-Up Market Trap
Rosenthal’s most contrarian stance is her warning about large enterprise customers. “Avoid the pull-up market as long as you can,” she said. “Resist it, because once you get a big customer… they will take so many of your legal, security, product, engineering, sales resources.”
The allure is obvious: a single customer promising a million or ten million dollars. The cost, she argues, is disproportionate. That one customer will consume attention from every function in the company, pulling resources away from the self-serve engine that is actually growing faster.
When the time does come to hire salespeople, the trigger should be precise: you have gone as far as automation can take you, and companies are explicitly asking to talk to a human — for a demo, for trust, for a handshake. That happens organically as you move upmarket. When you hire, the profile matters: AI-native sellers who genuinely use the technology and can sell it authentically.
Rosenthal reveals an awkward fact about OpenAI’s own sales org: the company famously never implemented commission plans. She describes that choice as “kicking the can down the road” until the team grew too large to fix easily. She does not recommend the path generally — OpenAI got away with it because of equity appreciation. The first two or three sales hires can be motivated by equity and company-building; they are “wired a little bit differently.” Eventually, you need “coin-operated enterprise sellers” who require cash compensation. Her design principle: simpler is better, but preserve upside for outperformance.
The Revenge of the Steak Dinner
The deepest tension in Rosenthal’s framework is between automation and human touch. The machine she advocates — self-serve, no pilots, no security calls, no salesperson until forced — is precisely the frictionless experience buyers want. Yet she also predicts that as automation saturates the sales function, human contact will become more valuable, not less.
Her prediction: “As more of sales becomes automated, in-person human interactions (conferences, dinners) will become more valuable and effective for closing deals.”
She calls it “the revenge of the steak dinner.” Conferences, booths, and dinners are effective and not expensive. They still require humans to run. The resolution to the apparent tension is temporal: automate everything that is commodity, spend human capital only where it creates differentiation. Value selling — articulating what the product does for the buyer, not what the product is — is the one function AI has not cracked. It is the justification for the dinner.
“You’re not selling your tool. You’re not selling your product. You’re selling what that tool or that product does to your champion or your buyer,” Rosenthal said.
All-You-Can-Eat AI, But Not Away From the Table
Rosenthal’s playbook lands in a market that is simultaneously expanding and consolidating. Her former employer is now preparing for an IPO as soon as 2027, with Chief Financial Officer Sarah Friar telling staff that the listing is just “another fundraise.” Enterprise revenue — the line Rosenthal helped build — is growing 50% quarter to date, faster than the company’s blended 35% growth.
Meanwhile, the competitive field has thickened. Anthropic’s sales doubled in the same quarter OpenAI’s grew 18%, and Google is pushing into vertical-specific enterprise AI with Gemini Enterprise for Financial Services and Legal. The enterprise deals Rosenthal describes are now being fought over by every major lab, each with its own approach to the same problem: how to sell AI to companies that are still figuring out what to buy.
The discipline she advocates — self-serve first, automation everywhere, humans only at the bottlenecks — is not just a nicety for early-stage startups. In a market where OpenAI lost customers to Microsoft Copilot because nobody answered the inbound queue for nine months, the go-to-market motion itself is a competitive weapon. The companies that build the machine first will not just grow faster. They will be harder to displace when the steak dinner finally happens.