Outset Founder Aaron Cannon: We Invented AI Users, One-to-One Synthetic Twins Anchored to Real People
Aaron Cannon, founder and CEO of Outset AI, announced the launch of Digital Twins. He stated that in 2023, they created AI interviewers, and by 2026, they aim to develop AI users: one-to-one synthetic twins anchored to real people, rather than having the base model repeat crowd averages.
The company will first conduct large-scale AI-hosted research, then generate interrogable twins using real human data. The page specifies: twins are not average personas; they come from a single AI-hosted interview and can be updated with follow-up visits; each answer carries high, medium, or low confidence levels and links back to the supporting real person's original words, rated by the individual. Questions can be posed to one or a thousand twins within minutes.
Cannon mentioned that they have collected millions of deep multimodal responses globally, supporting voice, video, and text in over forty languages, and can showcase prototypes, websites, or physical items from respondents' homes. Their client list includes Microsoft, Google, Uber, Nestlé, and Bain. The company is part of Y Combinator's Summer 2023 batch. He stated that the future of insights lies in understanding and simulating human experiences in parallel.
The product's selling point is to keep hard-to-reach demographics in a library that can be queried at any time, filling the decision-making gaps left by budget cuts in research. Unlike synthetic respondents that only generate crowd statistics, it emphasizes traceability to a specific real person. The launch itself does not disclose revenue or valuation.
The buyers are product and insights teams looking to determine pricing and copy before release, while the sellers are platforms that turn interview data into reusable simulators. Funding flows from corporate research budgets into interview minutes and twin calls. Beneficiaries are suppliers who can turn a single interview into infinite queries; those under pressure are paid per session human discussions. This is a product launch statement, not an independent third-party audit of simulation accuracy.
Source: Public Information
ABAB AI Insight
Outset transforms interviews from "scheduling meetings" to "collecting data first, then nurturing twins." In 2023, using models as hosts addresses scale; in 2026, using the same group to nurture digital avatars addresses the marginal cost of repeated questioning. One-to-one anchoring is aimed at countering the base model's average illusion: without original word links and personal ratings, simulation is merely stylistic imitation. Consulting firms and tech companies on the client list seek to hold meetings with "specific user types" even when research sessions are cut from budgets.
The capital path is to first capture the interview pipeline, then turn that output into a subscription-based simulation layer. The motivation is to shift insights from project-based to operating system: a person only accepts one long interview, after which anyone in the company can question the twin. Risks lie in compliance—multimodal home scenarios and "always interrogable" may encounter consent boundaries and deletion rights. If accuracy relies solely on self-assessment, buyers will eventually demand double-blind comparisons with real people.
Similar to quantitative trading recording transactions for playback, or games creating player bots, it also resembles digital customer service treating script libraries as employees. Market research is transitioning from sampling expansion to controlled simulation: samples are expensive, follow-up questions are pricier, and twins drive the cost of second questions close to zero. Whoever controls a confidence-rated original word library will be more like a research infrastructure than models that merely prompt "pretend you are a 25-year-old user."
The essence is technological substitution. The mechanism is to amortize the sunk cost of one-time interviews over infinite synthetic Q&A. Understanding still requires real people, but simulation begins without needing to bring people back. When the second question can be asked of a twin, the pricing power of the first question shifts from the interview room to the platform.
ABAB News · Cognitive Laws
Interviews are done only once; follow-up questions can be asked of twins.
Simulations without original word links are merely averages.
Once users are made into twins, insights transition from projects to systems.