Everyone Builds, Ships, and Sells. Winners Do It Differently.

In this interview, Kimberly Tan explains why there's such a big gap between a flashy AI demo and a real production application, and what it actually takes to close it. She shares the inside story of how Decagon and Prepared became two of a16z's most successful applied AI bets, and why she believes an unglamorous back-office workflow was one of the best opportunities in tech. She also lays out exactly where she thinks AI automation should stop, and why.

I'm Kimberly Tan. I'm an investing partner here at Andreessen Horowitz. I've been at the firm now for over six years, all in early-stage B2B software investing, today focused a lot on applied AI. I've had the privilege of working with a lot of companies, including Decagon, Prepared, Mem, and Sola.

I joined when I was 23 years old, not very familiar with the world of tech, so it was actually very hard in the first couple of months. I just tried to learn everything I could about venture. Having had very little exposure to the industry, I read all the classic books people tell you to read about venture. In every single meeting I would write down every single thing that they said that I didn't understand and then look it up that night.

I draw a lot of inspiration about what trends might be on the horizon from what founders on the ground are telling me, because they know best. They're seeing things that nobody else is seeing. In 2020 and 2021, a lot of these founders and builders were telling me, you don't understand, AI is going to be in everything. You guys have to understand AI. I think we just saw very organically how AI was going to change a lot of the fields that we were already spending time in, and really decided to spend more time as a result.

One of my very strong beliefs about investing in AI is that there's a huge gap between a fancy demo and a real production application, and that's true for a number of reasons. Primary of which is that it's really hard to build a really good AI product. It's easier to show an interesting demo when you know exactly what the underlying data source is and exactly what outcome you want to produce. But it's actually very hard to build AI products because AI is non-deterministic by nature: something that 95% of the time might do this thing, but 5% of the time does something totally different.

It's just a totally different paradigm of building than enterprise software, which is 100% deterministic: you click this button, this thing will happen. And so it requires a very different muscle and a very different way of building products. That has several implications.

First, I think it means that seeing a product actually work in production is very important, that this product will actually do and solve the problem you're intending it to solve, in a way that a demo just can't really show anymore because it's nondeterministic. It also means there's a lot more education, onboarding, and implementation needed to make sure this product will do what they want.

So what we're seeing today is a really big trend that wasn't necessarily true prior to this wave of AI: the forward-deploy motion that a lot of people have started talking about. You'll have somebody who goes to the client side and actually helps set up these AI products, making sure it integrates into the right solutions, has the right business context, and understands the guardrails of what you do and do not want to do, taking that last mile very seriously to make sure their product delivers value. I think that's become more important than ever.

Today, vertical AI or applied AI that serves a very specific industry, whether it's logistics, healthcare, etc., is one of the most fruitful areas to build in. It requires a lot of work to translate an industry's rules, regulations, compliance, and general business logic and culture into code in a way that an AI agent can actually do something against.

I think what people don't understand is that it's very difficult to work with these models. The capabilities out of the box are not the same as the capabilities needed to actually show value to an end enterprise buyer, and there's so much work between the base model and the end client that needs to get done. I think a lot of people underestimated just how much work there needed to be done.

You don't just call the most advanced model to solve everything, because it probably has more latency, it's probably much more expensive, and it's probably overkill for what you actually need. A more sophisticated agent query understands the intent of a query that comes in and knows which model to route it to, which models are best at which things, and which has the best latency and cost-effectiveness trade-off. It's actually a very sophisticated chaining of many models together in order to get to your actual output. I think a second reason is that enterprises have a lot of business context that is not immediately apparent through one prompt to a model.

Not only because people are pretty bad at prompting models, and it's actually very difficult to write a good prompt, but also because there's so much knowledge that exists in people's heads that isn't on paper. It's not ingestible by the models. You just need someone to sit down and talk to the customer, understand what it is they're trying to do, and then map that out. That sort of work is not something that a model can do out of the box. The only people who will understand it are the founders or the employees of startups who go to the customer site and take the time to actually learn these things.

I think Prepared is such an incredible story. They're an AI assistant platform for emergency response, literally for 911 centers: a perfect example of a vertical applied AI company. They knew the 911 market better than anybody. The CEO knows everybody in this industry, and everybody knows him. He goes to all the conferences, people love him, and he has a ton of customer empathy. He knew the needs of these customers in a way that a lot of other people didn't, and could build purpose-built AI solutions that worked just for that market.

I don't think 911 centers were particularly known for being incredible software buyers historically, but AI provides such differentiated value that people can understand: if you can help triage these calls to know what is an emergency and what is not, all those things add up to delivering the outcome, which is getting a person who needs help, help faster. They understand that value prop, and so there was a lot of market pull for Prepared. Being the industry-focused solution gives you a ton of leverage in knowing exactly what you need to build for them. To us, that's an enduring moat in a way that a lot of other sectors potentially don't have as much. A lot of our investments that have been doing incredibly well have taken a very specific industry and just been the premier AI solution for that industry.

The founding story of Decagon is, I think, one of the most fortunate things as an investor to have gotten to witness. We were there from the early days, before they had the name Decagon, before they even had an idea, and watched them build the company from the ground up. They were very clear from the beginning that they wanted to build in enterprise AI because they thought there was clear demand in the enterprise, but a lot of those enterprises don't know how to actually get that value. They thought their skill sets were particularly suited to building very good AI products, but their commercial instincts also told them they would be very good at the enterprise sale.

So they would go company to company, talk to a lot of tech-native companies at the time, and just ask them what their biggest pain points were: very straightforward, exactly what you would expect people to do, where they would just ask them for their biggest problems. They would ask them what the ROI was, and customers consistently told them that support was their biggest problem, and that if you could solve that, they would pay a lot of money for it. They were like, we have so many people who do this, consumers are always mad at support channels: imagine all the times you call somebody for a customer support query, and people are always angry. So they knew AI would be a very good solution to this problem and that they could build a solution that solved it.

It's been a really incredible journey to watch them go from iterating around having no idea exactly what they wanted to do, to very quickly landing on this, defining a world-class product, and then getting product-market fit almost instantaneously after that. You can really build an excellent solution, but you still need to find someone who will buy it. A lot of people who are technologists understand the value of the technology and why it's amazing and why there's going to be ROI, but that doesn't matter from a business standpoint if you can't explain it to a customer, they can't see the value, and you can't implement it in a way where they're happy about the product.

People need to be very smart about how they design their pilot so it can get into production relatively quickly and show a clear ROI metric. I think one of the big things everyone's talking about in enterprise AI today is what the actual ROI is on delivering an AI solution. In certain categories, like coding and support, I think people understand the ROI either intuitively, because you see your coders producing more, or quantitatively, because you can see for support that your resolution rate has gone up and your NPS score has gone up.

One of the most amazing things about Decagon is it's one of the only use cases today where there's actually very clear, quantifiable ROI. The market clearly understands that AI will respond 24/7: no more hold times, no more wait times, your CSAT score will go up, you'll answer more tickets, it'll be cheaper. It's just a clear, high-ROI use case.

But in a lot of other categories, where people are building AI solutions that automate one part of a flow but not the whole flow, or that are more of an augmentation solution, it's hard to really show the clear value you're delivering, and that's been a challenge for a lot of AI companies today. I would strongly urge everyone to not only make sure they have all the sponsorship they need in an organization to be set up for success, but also to have a very clear, identifiable ROI metric to point to at the end of the pilot, to be able to explain to people why they should adopt their solution.

There's a lot of mundane manual work that happens in the back offices of many large enterprises today, think data entry, claims processing, things like that. We thought there was a huge opportunity to actually automate a lot of this back-office mundane work, so we were very excited to invest in a company called Sola. They essentially allow the business users who have the context on the process, which is very hard to get out of their heads and onto paper, to record their own process. Then Sola's agentic automation framework will contextually understand the process: it'll understand that this is a login step, that this is a data-extraction step, and be able to build a bot that can dynamically handle those solutions.

Things that seem mundane and tedious that people had to do before, hopefully soon AI will be able to take over, and then those people can work on much more strategic work that is much more long-term value-creative to the businesses they work at.

There's a big difference between automating a support ticket, which I think can be done relatively end-to-end today depending on the complexity of the career, versus fully automating, let's say, a lawyer, fully automating a doctor, fully automating an engineer. One dimension to think about is what the actual domain is that you're spending time in, and how fully you can actually automate a solution from a technical standpoint. The second dimension is that there are a lot of cultural, social, and regulatory dynamics that allow you to fully automate something or not. For example, maybe with a lawyer, you can do a bulk of the work they do today for relatively straightforward legal work, but you still need a lawyer to sign off on it at the end of the day because there's liability associated with it.

So for a lot of work today, I think you still want a human there, and I also think that for many people it's an important step in actually gaining trust that the AI is doing what it intends to do. In most cases, some level of human oversight is important. Even in the support case, which I'd say is one of the more straightforward fully-replacing solutions, there are still paths to escalate to a human agent when you need to, and there are still human managers overseeing the AI agents to make sure it's doing what's intended. I think that dynamic will probably be true for quite a long time.

In venture, you have to understand that the job has risks. You're investing in very early-stage companies, often in markets that are still nascent, with founders who have a long road to build. There is risk involved, and there's a lot of responsibility in managing that amount of capital for LPs, making sure you can make the best decisions you can and then steer the companies you partner with to the best outcomes they can reach.

But on a day-to-day basis, you try not to think about it too much, because day-to-day it's just working with the founders you work with, meeting the folks you want to meet. You know that when you make an investment, many times the distribution of outcomes may not be the best outcome possible, and you just have to make peace with that. There's a lot of value in not only working with the companies and founders that are doing incredibly well, beyond what you could have even hoped, but also working with the founders who are going through a rough patch, which everyone goes through at one point, and making sure you're there for them too. So on a day-to-day basis, I don't think about the number as much. I just think about each individual decision you make when you partner with a company, and then helping them the best you can to steer them to whatever outcome they're hoping for.

Companies take a really long time. Some of them work immediately and then run into challenges. Some take a really long time to hit their spark and then work incredibly. And some just have many years of struggle. In those moments, I think a lot of founders probably do want tactical help from their investors, and we offer that too: I'm constantly closing candidates for people, we make lots of customer intros, we help them think about their runway and cash balance, etc. But I think often what they need most is just a patient investor, one who understands that this is a long journey.

We often have companies going through rough periods at any point in time, and just knowing that they can call us, that they can be honest, that they can tell us what's really happening in their company, and know that we're on the same side and we're going to try to help them, not yell at them, not blame them, because these founders are doing the best they can, I think that actually goes a long way. I spend a lot of time trying to nurture friendships with a lot of my portfolio founders. We're still investors at the end of the day, but to the extent that you have a real personal relationship with them, they know they can call you, and many of my portfolio companies have called me during stressful moments in their company building.

Oftentimes there's some big junction point in the business, so there's some offer on the table, maybe some really important executive who's going through something, maybe a fundraise didn't go as expected. In those moments, there's a time to take a lot of action and strategize what is correct. But often founders, in those moments, don't want to make any decisions that are too rash in the heat of the moment when something is happening.

You don't want to make any decisions that are so irreversible. So often we'll ask founders to think it through for a day. Many times I've gone and met with the founder in person. I had a portfolio company that at one point went through a relatively difficult fundraise, and we were actually in the office at 6:00 a.m. on a Friday doing last-minute pitch prep. In those moments, just knowing that you'll actually be there for them and look out for the best interests of the company goes a really long way.

This is a very competitive ecosystem we're in now. Everybody knows AI provides a lot of value, and everybody knows there are a lot of opportunities to go after, and you could have an early advantage. Maybe you're first, maybe you're generally earlier, but if you don't press the gas and continue to have momentum and build on that, the world might just move too fast. I really think now is the time to lock in and stay super focused, because a lot of great companies will be built in this time.

But the ones that do will have founders who are really empathetic to their customers, who are super technical and AI-native, and who also just move faster and work harder than anybody else.