Aaron Cannon, co-founder and chief executive of the AI research platform Outset, made his third appearance on TBPN to launch Digital Twins — models trained on one named customer at a time rather than on a segment. He explained why he thinks research has only ever looked backwards, what it costs to ground a twin in a real person, which customers are worth simulating and which are not, and where the product goes once the model stops waiting to be asked. The news discussion the hosts ran earlier in the show is covered in a separate summary.
👤 Guest: Aaron Cannon, co-founder and chief executive of Outset, an AI research platform whose AI-moderated interviews now also train per-person digital twins
🎙️ Hosts: John Coogan and Jordi Hays, who present TBPN live on X and YouTube every weekday
📰 Published: 31 August 2026 on YouTube (TBPN)
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Show notes | ⏱️ 2 hr 24 min
Key Takeaways
Research is meant to predict the future and has only ever been able to look backwards
"And the idea is that you can now not only look backwards, but you can look forwards."
The twin is trained on one named person, not on a segment, and Cannon says that is what makes it trustworthy
"So, it's like a one to one."
Grounding a twin costs about seventy-five minutes of a real person's time and $200
"To be clear, it's worth more than the shoes."
The grounding interview is deliberately not about the client's product
Outset builds a "persona core" — values and priorities — then layers the client's operational data on top
Twins grade their own answers, and the platform reports how much to trust each one
Respondents mark their twin's output, and confidence scores flag which answers need more human data
The wedge is audiences a company cannot reach, not audiences it already hears from
Nobody needs to simulate whether shorter customer-support waits are popular
A food company is not short of consumer feedback; it is short of feedback from retailers
The second half of the pitch is internal access, not accuracy
A finance team testing new pricing cannot be turned loose on real consumers, and can be turned loose on twins
The destination is research that runs before the question is asked, and then acts
"We're gonna send agents to go talk to all the people in that part of the economy"
Three Appearances In, the Numbers Are Multiples and the Product Is New
Cannon opened on the business rather than the launch: since the last conversation, "I think tripled revenue, tripled the team size, and now launched a very cool you."
The new thing is a simulations lab. "And now we just launched Digital so we built a whole simulations lab and have now built" — the twins are what came out of it
Coogan set it up with the Y Combinator joke. Founders are told endlessly to talk to their customers; Cannon, who came out of YC, built the thing that means never having to take another customer call
"It's exactly right. Yes. I my I started my career as a researcher and I was like, I you know, I don't want to do that anymore."
Cannon then said, in all seriousness, that the real argument is different
Research Exists to Predict the Future, and Until Now It Could Only Look Backwards
Hays pushed first, and hard. He has run e-commerce companies; cohort profiles are useful and plenty of platforms already build them; talking to an individual customer is nice but you might get a weirdo who represents nobody. What does grounding a twin in one person actually buy?
Cannon's answer started with what research is for. "So here's the like, to take a step back, research the whole point of research is to try to predict the future." You are trying to make a decision for the business, and the data you have is behind you
"And the idea is that you can now not only look backwards, but you can look forwards. You can actually start simulating, like, if I do this new marketing campaign, how would this customer start reacting? What are the things they would say?"
The existing business is the foundation, not the product being replaced. "And we're very good at research. That's what we're doing. That's what we built our business around. But now what we can do is take that research and build out the simulation platform so that you can say, okay, I can actually see into the future of how this group would react to some very big business decision."
The Twin Is One Person, and Cannon Says That Is the Whole Trust Argument
"And so we train them by training digital twins on individuals. So, it's like a one to one."
"Like, can go in, talk to a digital twin, and you know that's based on John, and like, that's got all of John's experience and like all this data about John, and it can talk on behalf of you, right? And so that's like the mechanism that makes it super, super useful and also like trustworthy."
The cost of grounding one is a real payment to a real person. Coogan asked what would persuade anybody to sit through a long AI-moderated interview about, say, a pair of shoes
"I think the last run we did was these were actually like seventy five minutes, the last one, and I think we paid them $200."
"To be clear, it's worth more than the shoes. Right?"
The Grounding Interview Is Not About the Client's Product
Coogan asked whether a twin belongs to one client — whether Outset interviews someone about a single Excel feature, or about how they use technology generally, with the result useful to several customers at once.
"That's right. We developed our own grounding interview, so it's not just about their use of Excel."
"The grounding interview, it develops what we call persona core. It's like trying to get to like who you are, what are your values, what do you care"
The client's own data goes on top of that base. "So we can do that as the core and then we layer on other sets of data. So maybe Microsoft has Yeah. You know, operational data we can layer on to make it much more powerful."
"So the idea is that like we have extracted, you know" as much of the person as possible
The twins are refreshed, and marked by the person they are modeled on. "And then what we do is we go back to that person and refresh that data regularly."
"So we'll even have them grade their own twins' answers. Like, how do our twin do?"
Cannon confirmed the loop is reinforcement learning in shape — "They're kind of RL."
The output is labeled with how far to trust it. "And then we also put confidence scores. So you'd be like, obviously not all of our probabilistic, out your output here is going to be the same. So we can tell you which ones you should, like, really trust and which ones need a lot more human data to kind of bring to it."
The Wedge Is Audiences a Company Cannot Reach
Coogan asked where the market pull is strongest, since in theory every business wants to forecast customer behavior.
"The biggest pull right now is that companies have really hard to reach audiences." He gave buyers of customer-relationship software and other narrow slices of the population as examples
"And so the incremental value there is massive because they can't actually go out and do constant research with that group. So they have, like, very few signals today."
That is also why those buyers move first. The incremental value is large enough that they adopt early, while the market still has questions about the approach
Where Simulation Does Not Help, and Why Nestlé's Retailers Are the Interesting Group
Coogan asked the inverse — which company is already drowning in customer feedback.
Some decisions do not need forecasting at all. "Like a business doesn't need to say like, oh, what if we made customer support wait times shorter?"
"Like, no one is sitting there being like, wish when I called my internet service provider that I wish it took an hour Yeah. Instead of thirty minutes."
The example he offered, flagged as illustrative rather than live, was Nestlé. Consumers are easy and cheap to reach; the company is not short of their opinions
"But if you were actually trying to do if you're at Nestle and you're trying to do research on retailers and what they think about your brand, That's actually a much harder group to go find and do research with."
"You're not drowning in feedback from your retailers."
Hays raised the obvious objection to simulating a food business. He said he would not trust an AI to tell him which milk chocolate tastes best, because the taste data is not in the model, and that a lot of consumer research on those products means watching someone eat the thing
Cannon did not argue. The Nestlé pitch, he said, was the existing product — AI-moderated interviews with real people, "which AIs are just doing the interviewing. You guys are asking the questions."
"And that's our core thing, and that's what we've had for the last couple of years. But simulations is an extension to that"
The second half of the pitch is not accuracy, it is internal access. "Number two is we say that, like, anybody in your company can access them."
"It's like if your finance person's working on new pricing model, you don't want to like give your finance person like the ability to go research with real consumers necessarily. But now they can actually like, test ideas, test new pricing and see how things would adapt"
The Destination Is Research That Runs Before Anyone Asks, and Then Acts
Coogan's last question was about the gap companies have between feedback received and changes made — whether Outset ends up being prescriptive, telling a client to change the color from red to green.
Cannon reframed it as removing the human trigger. Today a person decides to go and run research. "What we are building towards, and that's our original product, right? We're building towards is something where the model can go do research for you before you've even asked the question, right?"
"And so, it is constantly trying to figure out the answers to the things that you don't even know matter"
The worked example is a decline nobody has spotted yet. "It's like, hey, there's a you know, your revenue's down in that sector. You don't even know that yet. We're gonna send agents to go talk to all the people in that part of the economy and, like, go figure out what's going on, talk to some simulations if we already have them built, and then go take action."
"And that action is where I think where you like where we can ultimately have more impact than just collecting and synthesizing feedback."
Coogan carried it one step further — that the output should become a specification handed to a coding company such as Cognition to go and build. Cannon agreed, and said that is already happening
Cannon's bottom line is that the value of a digital twin is not that it replaces talking to customers but that it makes a customer available to everyone in the company at any hour, including the people a business would never let near a real one.
Products, Companies & Tools Mentioned
Outset (Cannon's AI research platform, which says it has tripled revenue and team size since his last appearance and has now launched Digital Twins out of a new simulations lab)
Microsoft, Uber, Google and Coinbase (The client list Coogan read out — his "murderers row" — and the reason he asked whether a twin serves one customer or several)
Nestlé (Cannon's illustration, offered as an example rather than a live simulation: the consumers are easy to reach, the retailers are not)
Cognition (Coogan's endpoint for the product — hand the research output over as a specification and let the coding agents build it. Cannon said that is already happening)
Y Combinator (Where Cannon came from, and the source of the "talk to your customers" advice the launch is a joke about)
Get the latest market chatter as it happens:
If this saved you the listen, send it to one person who would rather have the time back.