AI Engineer World's Fair 2026
Reverse-Engineering the AI Buyer — Aliisa Rosenthal, Acrew Capital
Loading slides…
Read the talk
Reverse-Engineering the AI Buyer
Aliisa Rosenthal draws on OpenAI’s enterprise growth to explain how product access, sales automation, pilots and pricing shape adoption—and where human help still earns its cost.
From a talk by Aliisa Rosenthal
At a glance
Ideas worth remembering
Start with self-service and automation; use their failures and customer feedback to decide which enterprise capabilities and human roles to add.
Acknowledge inbound demand immediately and help buyers complete difficult next steps together. Silence and homework both let the buying cycle drift.
Reserve supervised pilots for opportunities that justify their resource cost. References, limited evaluations and contracts with exit terms can satisfy validation needs differently.
A lower base fee plus usage can widen access, while company and employee spending caps address the risk of growing bills.
Human help still matters for trust, value selling and customer-specific integration. Initial design partners also need more attention than a repeatable sales funnel.
Build the machine before the team
OpenAI already had a working buying motion when Aliisa Rosenthal joined: developers found the self-serve API, swiped a credit card and started building. Revenue was a couple million dollars. Over the following years, Rosenthal helped grow an organization of several hundred people and enterprise revenue to several billion. Bringing this “somewhat alien technology” into companies exposed a recurring problem: the way a product gets bought can obstruct the way people want to use it.
The conventional sales playbook starts by hiring salespeople, revenue operations and sales engineers. Only afterward does the organization find bottlenecks and add automation. Rosenthal reverses that sequence. Automate what can work without a person, observe where the process breaks, and hire people to get through those specific bottlenecks. The resulting team has a reason to exist beyond reproducing the usual sales organization.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Enterprise demand hid the larger self-serve market
ChatGPT launched at the end of 2022 without enterprise features. Employees were already using it, but their organizations needed single sign-on, a nondisclosure agreement and an invoice. Those requests concerned the machinery around adoption: a useful product still needed a way to fit into company purchasing and administration.
After nine months of requests, the enterprise effort received approval. The largest companies had been the loudest and most successful at reaching Rosenthal, so their requirements shaped the offering. OpenAI built a fast, expensive product with extensive features and went aggressively upmarket. That selection process made large-enterprise demand unusually visible while leaving buyers who preferred a simpler purchase less represented.
The enterprise business performed well. Then, in January 2024, about four months after the enterprise version, a self-serve option arrived. Buyers could enter at a lower price without a heavy sales process. The mid-market and small-business motion grew much faster, and some customers preferred the cheaper option. Sales representatives now competed with their own company’s self-serve funnel. Rosenthal describes this as cannibalization; the account establishes the direction of the change without quantifying its size.
This example changes the order of product discovery. A self-serve launch first would have revealed what customers could accomplish independently and what they still lacked. Those missing capabilities could then justify a more expensive enterprise tier, followed by a team to sell it. Rosenthal says subsequent OpenAI product launches followed that sequence: start with self-service, learn where human help belongs, and add it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Ten thousand inbound requests still needed an answer
Early inbound demand overwhelmed Rosenthal and four sales representatives: roughly ten thousand requests arrived each day. A “champagne problem,” perhaps, but demand did not stay available indefinitely. OpenAI used Clay, which Rosenthal found useful, yet catching up with lead handling later could not restore information or relationships the initial process had failed to capture.
Two separate mechanisms would have helped:
- Capture information for follow-up: An optional, detailed sales form could collect requirements and a phone number from buyers willing to provide them. Once the team had capacity, it would have a way to reconnect.
- Acknowledge demand immediately: An automated email could explain that the enterprise product was not ready, confirm a place on the waiting list and ask what the buyer needed. That would keep the relationship active during the delay.
When the enterprise product finally launched, almost every company Rosenthal contacted told her it had received no response and had bought Microsoft Copilot. Silence had become part of the buying experience. The practical lesson is to make an answer possible before making a full sales conversation possible.
Outbound has a similar division of labor. Researching prospects, sequencing outreach and dialing can be automated so representatives spend their time on qualified leads and valuable conversations. Clay appears here as a tool OpenAI used; Nooks appears as a tool Rosenthal had heard good things about. The useful distinction is between preparing a conversation and spending a person’s attention on it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Keep the buyer moving—and avoid selling the same deal twice
A sales conversation can stall the moment the buyer leaves with homework. Filling out a form, integrating something or finding another person’s answer creates work outside the seller’s immediate reach. Instead of asking a buyer to complete five tasks and return, help complete them together: run a live Zoom session, organize a hackathon or work in the customer’s office. Shared work keeps the next step moving rather than waiting for the buyer to restart it.
The same principle applies to purchasing. Fewer pricing choices, fast paperwork and manageable approvals reduce the work required to say yes. A pilot can introduce the opposite pattern: sell the pilot, invest heavily in supervising it, then run another sales process to secure the actual purchase. “Pilot hell” consumes resources even when the trial succeeds.
Reserve that effort for the largest, highest-value opportunities. Buyers still need to validate that a product works in their environment, but a supported trial is only one way to satisfy that need:
- Customer reference: Introduce another customer who can discuss the product’s use.
- Limited evaluation: Evaluate a portion of the buyer’s data or demonstrate the product with custom data over Zoom, without handing over ongoing access.
- Contract with an exit: Sign the agreement with a ninety-day opt-out if the product does not meet the customer’s needs.
- Shorter commitment: Offer a one-year proof of concept against a normal three-year contract.
Why does the ninety-day opt-out change the cycle? In a conventional supported pilot, the vendor invests in validation before winning the purchase. With a signed contract and an exit window, the customer must validate the product while its opportunity to opt out is still open. The diagram shows that change in timing. It moves commitment earlier and puts a deadline on the customer’s evaluation. Rosenthal presents these flexible terms as tactics available to startups, rather than terms a giant company would routinely offer.
Where commitment falls in the evaluation cycle
Read the diagram as text
- Sell the pilot. Supported pilot — Win agreement to begin a supported trial.
- Manage validation. Supported pilot — Invest team resources in running the pilot.
- Sell the purchase. Supported pilot — Run another sales process after a successful pilot.
- Sign the contract. Ninety-day opt-out — Include an exit if the product does not meet the customer's needs.
- Customer validates. Ninety-day opt-out — The customer evaluates while the opt-out clock runs.
- Sell the pilot → Manage validation: Begin supported trial.
- Manage validation → Sell the purchase: Success starts another sale.
- Sign the contract → Customer validates: Ninety days to opt out.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Remove security delays and lower the cost of entry
Security review is another place where deals stall. A trust portal lets customers sign an NDA and retrieve penetration-test reports and security documentation themselves. AI tools can also fill out security questionnaires. These mechanisms automate the repeated exchange of information around security review. Before scheduling a lengthy call, direct the buyer to the portal and ask which unanswered question still requires a conversation.
Pricing created a different obstacle. Rosenthal recalls an initial ChatGPT Enterprise price of sixty dollars per user per month, set largely around the cost of serving the high-end product. Lower-priced competing offerings exposed the problem. Her analogy is “building the Porsche before the Prius”: the expensive offering arrived before a more accessible entry point.
Follow the buying decision through its change. At sixty dollars per user per month, organizations restricted purchases to a subset of employees, such as developers or investors. Each additional employee required another substantial fixed commitment. OpenAI later moved to a lower base license fee plus usage charges. More companies signed agreements, access spread more widely within companies and usage increased over time. The entry decision became easier, while spending could grow with actual use.
Usage pricing introduces its own objection: the bill can climb as adoption grows. A dashboard with a monthly spending cap, or caps for particular employees, gives buyers a way to contain that exposure. Rosenthal observes that many companies like knowing the control exists even when they do not use it. Usage ideally corresponds to value, but the pricing mechanism alone does not establish that every additional unit of use is worthwhile.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The revenge of the steak dinner
The signal to hire salespeople is a repeated limit in an otherwise working process. Buyers need a demo, a conversation or someone they can look in the eye before trusting the purchase. Larger companies are more likely to require that interaction. Hiring then addresses a demonstrated obstacle rather than an assumed need for a large team.
Large deals can also distort priorities. An enterprise offering a million or ten million dollars is tempting, but serving it draws on legal, security, product, engineering and sales resources. Be selective about the first large partners. The contract’s revenue is only one side of the decision; the company must also live with the work and product direction that customer requires.
The preferred salesperson uses AI themselves, understands it and can sell it credibly. Automation can handle outbound, demos and security checklists while leaving buyers hungry for time with a person. Conferences, booths and dinners create those conversations. Rosenthal calls this “the revenge of the steak dinner”: human contact retains a job even as more surrounding work becomes automated.
Compensation should match the kind of seller the company needs. OpenAI grew a large sales organization without commission plans during Rosenthal’s account, helped by equity packages and the company’s increasing value. That experience does not make commission-free hiring a general model. The first two or three sales hires may be builders motivated by equity and the chance to shape a company. Experienced enterprise sellers are harder to recruit without variable cash compensation. Keep the eventual plan simple and give outperformers upside.
The remaining human task is value selling: connect the product to what changes for its buyer or internal champion. A feature list describes the tool; a useful sales conversation explains what the tool lets that person accomplish. Rosenthal considers humans better at making that connection than AI at the time of the talk. That is the substantive reason to keep the dinners and conversations, beyond familiarity with traditional sales.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Self-serve trials and deployed engineering have different costs
The Q&A sharpens what avoiding pilots means. Product-led growth lets customers enter and use a product themselves. Access can be limited by consumption, time or available features. The objection concerns a contracted, services-heavy proof of concept that needs a salesperson or sales engineer to supervise it. Both let customers try the product, but only the latter commits substantial team attention to each evaluation.
Forward-deployed engineers add another services tradeoff. They are expensive and difficult to hire, yet helping integrate a product into a customer’s environment can make it harder to leave. The customer gains a customized tool that works for its particular organization. Rosenthal would prefer a product that works without this help, but sees good deployed engineers as an investment that can pay through a longer customer relationship. Her hope that continual learning might eventually remove this integration burden remains a forecast.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The first ten customers need a different approach
The final question places an important limit on the automation-first advice. A startup seeking its first ten customers should treat them as design partners who need much more hands-on attention. Automated outbound belongs later, once the company has moved beyond those initial partnerships, has some product-market fit and is starting to scale. Early customer work must generate feedback about what the product should become.
Choose partners who will collaborate well and understand the product’s actual maturity, including its enterprise features. A recognizable company can be useful without placing the experiment in its main production system. Working on an internal or less critical product may make that customer willing to accept more risk and a lighter security process. The relationship can provide a recognizable customer logo and open the door to a larger agreement later.
Existing relationships are ideal for these early experiments: former customers or colleagues, investor introductions and friends inside companies. Friendly partners make room for a product that is still developing. The closing distinction is about stage. Spend personal attention where the company needs to learn, then automate the repeatable buying process and add people where customers still get stuck.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
Related talks
- GTM Engineering: The Technical Bits — Everett Berry, Clay
Develops the account data, tool orchestration, persistent agents and outreach coordination behind an automated go-to-market process.
- Mastering AI Pricing — Mayank Pant, Stripe
Explores subscription-plus-usage pricing, variable compute costs and billing units that reflect customer value.
- Forward Deployed Engineering at Cursor — Pauline Brunet
Explains when customer maturity and customization call for embedded engineers, advisory support or self-service.
Full transcript
Loading transcript…