About
At Foundation, I invest in early-stage (seed / Series A) B2B software including companies…
Articles by Joanne
Activity
24K followers
-
Joanne Chen shared thisAgents need a new software stack - most of today's software was built for humans. The 3 biggest opportunities for founders today: 1. Simulating complete environments to test code. Before AI, code changes were slow, and manual. With agents, thousands of changes can be made a day to code, in parallel. The bottleneck moves from producing code to verifying it works. The opportunity is to build a system that spins up full test environments in seconds, so hundreds of agents can be verified in parallel. Every change will have to pass through these environments before it ships - and whoever owns that check becomes the gatekeeper to production. 2. Building a live, orchestration layer for documents. When people work on internal documents, they often rely on an informal system, a shared sense of which document is the real one. i.e., to find a more up to date document you might Slack someone. That system collapses when hundreds of agents create and edit the same work at once. The risk becomes an agent grabbing whatever it can find and building on stale work. The opportunity is a layer above documents that every agent checks before it acts. Is this the official version or a draft? Who approved it? Instead of being passive storage, the layer would orchestrate how work gets done. 3. Business process systems for agents SAP and Salesforce became giants by owning what happened. The next generation gets built by owning why. ERPs like SAP or Netsuite, are designed for human operators. They record outcomes, not reasoning. That judgment lives in Slack threads, and people’s heads - and then it evaporates. Without that context, AI features don’t work. They don’t have what they need to act autonomously. The opportunity is to build a system where agents are the primary operators. As agents do the work, the reasoning behind every decision gets saved automatically. These systems would own why decisions happened and are a big opportunity in verticals where edge cases dominate work. We go deeper on each - plus the four questions every builder needs to answer - in the blog https://lnkd.in/gYmMdc86
-
Joanne Chen shared thisIn 5 years, there will be trillions of agents crawling the web. Most websites are not prepared: They’re built for human users. If founders want their products to be visible in the next few years, the web needs to change. Agents get tripped up by things humans barely notice: JavaScript-heavy pages, bot walls, information that isn’t cleanly structured, and hard-to-interpret buttons. When a site doesn’t give an agent a direct way to accomplish something, the agent has to figure out how to use it from the page structure, the visual interface, or both. This adds a translation layer, and it can be slow, brittle, and prone to error. We’ve already seen efforts from first movers. Shopify is experimenting with WebMCP and Google and Microsoft have proposed a website standard. It enables sites to adapt their code so key flows (search, checkout, bookings) become readable and usable by agents. Websites can also now publish an llms.txt file, which gives agents a cleaner map of the site. But we’re still very early. At Foundation Capital, we assessed the readability of 20 leading AI companies, and the results were surprisingly spread. Only 7 of them had published an llms.txt. In a world where agents become integral to every B2B sale, it’s important to get this right. But this tension is not unique to websites - it runs through the entire software stack. A whole new stack needs to be built for agents. Leo Lu and I wrote about the 3 biggest opportunities here: https://lnkd.in/gYmMdc86 Big thanks to Savin Goyal and Nicholas Diao for their thoughts and feedback on this blog.
-
Joanne Chen shared thisThe demand for salespeople at AI startups is going through the roof. A few years ago, the bottleneck to growth was product engineering - coding, testing, review. AI is shrinking that bottleneck fast. With agents and tools, you can scale R&D without the equivalent growth in headcount. But a new bottleneck has emerged… trust. Personalized demos are easy nowadays. When every product looks similar, buyers need to trust you more. That means more face-to-face, in-person, emotional selling. Events and relationships become even more important. You can't remove this bottleneck with more agents. You just need more humans. It reminds me of ATMs. When they arrived, everyone assumed bank tellers would disappear. Instead, teller jobs grew - ATMs made banking cheaper, branches multiplied, and tellers shifted to higher-value work like selling financial products. I think something similar is happening here. AI is collapsing the cost of building product. That means more companies, more competition, and more need for the human work of convincing someone to change how they operate. At least in the short-medium term, we're going to need more salespeople in the age of AI, not fewer.
-
Joanne Chen shared thisThe software stack for agents will be built this year. We’ve outlined 2 tests at Foundation Capital to show founders where to build: Nearly every system in the software stack was designed around a human sitting in front of a screen. But increasingly, work is being done by agents. Agents that work continuously, in parallel, at machine speed. Capturing that capability means redesigning the systems themselves. But where is the opportunity greatest, and where should founders focus their energy? There are 2 questions that identify where the need for change is most urgent: Test 1: The agent accessibility test “How easily can agents use the system as it exists today?” For systems like databases, the opportunity is small. From day one, they were built for software-to-software communication - able to be queried a thousand times at once by software. Everything a database can do is already reachable by an agent. But for systems that cater poorly to agents, the opportunity for innovation is high. Take the visual workflows you would find on a website. Those require someone to click through five screens, interpret ambiguous controls, and approve each step. That leaves far more room to redesign. Test 2: The agent usage test “How much will the system’s usage multiply once agents take over the work?” Agents try many more paths than humans to produce a finished result. In the time a person completes one pass, an agent can run hundreds of searches or test dozens of implementations. The scale of the increase depends on when the system gets used: Some systems are only touched when work is finalized. Say an agent is hired to find a company's next software vendor. It compares dozens of options, negotiates terms, and lands on one. All that work produces a single signed contract. The contract system is hit once. Growth is real, but not explosive. Systems touched on every intermediate attempt - like code testing and validation - see usage explode. Every trial and parallel experiment hits the system. Hundreds of agents can generate thousands of these attempts at once. When the answer to both questions is high, the biggest opportunities for early-stage builders appear. It’s also where incumbents get displaced. The system has to be rebuilt for agents, and incumbents can't follow - the rebuild removes the human seats, screens, and workflows they monetize. Taken together, these two tests give founders a way to judge what’s worth building. Leo Lu and I applied them across the stack. More in our blog: https://lnkd.in/gYmMdc86
-
Joanne Chen shared thisIntroducing MachineView. It scores how readable your website is to an AI agent, from 0 to 100 (for free). Agents are about to become the web’s heaviest users, and most websites are not ready. In 5 years, there will be trillions of agents doing tasks on our behalf. Gartner projects that by 2028, 90% of B2B buying will run through AI agents. Sites that agents can't read will get skipped or misrepresented. We built MachineView to strip away the pixels and see a site the way a machine does. It tests 5 things: Can the agent: (1) access the site, (2) retrieve its content, (3) parse its structure, (4) understand its meaning, and (5) navigate through it Leo Lu and I ran 20 leading AI companies through MachineView, and the spread of scores was surprisingly wide. Score your site: machineview.vercel.app Websites are just the beginning. Nearly all software was designed around a human sitting at a screen. As agents become the primary users, existing software stacks will need to change - some incrementally, others from the ground up. In my latest p.o.v. with Leo Lu, we walk through the 3 biggest opportunities for rebuilding the software stack for agents and much more. https://lnkd.in/gYmMdc86 Special thanks to Savin for pushing on our ideas and helping make this piece stronger.
-
Joanne Chen shared thisHumans& $4bn valuation, pre-product. Thinking Machines Lab $12bn valuation, pre-product. Safe Superintelligence $32bn valuation, pre-product. How can valuations be so high? Startups like humans& with only a handful of founders can raise hundreds of millions of dollars at multibillion valuations. Out of the gate - pre-revenue and pre-customer. How can this be? Well, OpenAI and Anthropic have showed that model labs can be trillion dollar businesses, so investors are willing to pay up early hoping the neo labs follow the same trajectory. But there is danger here. The market is inundated with options, many with no commercial traction yet. So how do you decide which to invest in? It's not dissimilar from recruiting strong talent. You invest in the person first. You look for exceptionalism, founders with big ideas and a strong growth mindset. And the persistence and tenacity to grow as the role changes from founder to CEO. Thanks again Edward Ludlow for having me on Bloomberg Tech.
-
Joanne Chen shared thisAs an AI application layer company, there are 3 ways you can deliver value right now: 1. Own model implementation Application companies can enable large enterprises to successfully use models internally. Helping with the implementation, governance, compliance, and vertical context of how that model should be used. Working within a customer’s existing budgets for tokens and models they chose to use. Jasper is a great example - today, it’s selling to large enterprises teams, doing both workflow automation for marketers as well as content creation. 2. Guarantee the delivery of a business outcome using AI Think a law firm that does the law work for you, versus selling software to a law firm. The value here is taking away risk from the customer and taking ownership of the outcome. 3. Optimize model cost Right now, everyone is buying a bunch of tokens. Over time, CIOs are going to want predictability of cost and strong ROI on spend. The model companies can't offer it. They're incentivized to grow revenue by selling more tokens. Everything they do is predicated on that. They're never going to tell you to buy less. If I buy food from Safeway and I’m looking to go on a diet to lose weight, I would not want to hire my nutrition consultant from Safeway. Startups can offer that budget oversight and control.
-
Joanne Chen shared this“Model companies are overselling tokens.” "Agents are burning tokens like crazy." Well what did we expect? That’s what the model companies are incentivized to do. Anthropic and the other model labs grow revenue by selling more tokens. Everything they do is predicated on that. And they are under intense pressure to hit aggressive growth targets and justify massive valuations. That pressure makes their incentive very clear - sell as many tokens as possible. They are never going to tell you to buy fewer, and they are certainly not rewarded for helping customers dramatically cut token spend. It’s like an oil company also being your energy-efficiency consultant. They get paid every time you burn more fuel. So they will never push hard for solutions that dramatically cut your consumption. If the same companies selling you tokens are the ones selling you “token optimization” services, they have an inherent conflict of interest. There has to be a separation. An independent third party whose job is to extract more value per token - or decide when to switch models entirely. That is a huge opportunity right now for the AI application layer.
-
Joanne Chen reposted thisJoanne Chen reposted thisJoin me at UC Berkeley and help find and commercialize the next breakthrough discoveries in biotech from one of the most prolific life sciences ecosystems on the planet. (Am I hiring *two* Executive Directors positions three weeks apart? I am, if you recently saw a similar post for ATEC, this is different.) If this sounds familiar, yes, the Life Sciences Entrepreneurship Center is the virtual center I started five years ago. It works closely with Bakar Bio Labs, Innovative Genomics Institute, Energy & Biosciences Institute, Berkeley SkyDeck, the Bakar Fellows Program, the Molecular Therapeutics Initiative, and others. (If you've attended the huge Berkeley Bio Startup Showcase @ JPM, that group should look familiar.) LSEC has had some big wins, and now I'm looking for someone to carry the torch. Tons of information at the link below, and if you could please share with anyone who may be interested. 🙏 #YouSeeBerkeley #Entrepreneurship Laura Hassner, Ariel Notcovich, Katarina Klett (she/her), Tobias Schmid, Kaspar Mossman, Gino Segre, Jeremy Alberga, David Schaffer, Daniel Nomura, Julia Schaletzky, Jason Barron, Sibyl Chen, Caroline Winnett, Yi Liu, Susan Jenkins
-
Joanne Chen liked thisJoanne Chen liked thisPersonal update: I ended my summer break early to join Dyna Robotics as Head of Sales. Most of my career has been spent selling software. This is a little different. AI is already changing how we code, answer support tickets, do legal work, and more. But there’s an even bigger world of work that happens outside a computer: manufacturing, logistics, construction, infrastructure. Work that still requires hands and machines. That's the bet we're making at Dyna: robots that can own the whole job, not just automate one task. I got to know Lindon Gao through the First Round Capital community this year. I was impressed by how he thinks about where this space is going, but even more by how obsessed he is with making customers successful. I'm a few weeks in, learning how to train a model, and trying to keep up with everything this team has already built. There's a lot coming next! After years of building the commercial engine at the earliest, hardest stage of a company's life, I'm excited to do it again in a completely new category. This time, the software has arms. Let's go 🦾
-
Joanne Chen liked thisJoanne Chen liked thisWe launched a firm-wide newsletter this summer called The Foundation where we share ideas from the earliest stages of building. It’s monthly. It’s smart. It’s thoughtfully designed. This month Joanne and Leo share their take on AI agents as software’s newest users, along with a tool they built called Machineview that lets you see your site the way AI agents do. Subscribe at the link below and let us know what you think!
-
Joanne Chen liked thisJoanne Chen liked thiswe spent 30 years making the web easy for humans to read. so i built a small tool to check how it reads to AI agents called machineview: strip away the pixels and see what an agent sees in md format. i ran a handful of VC and startup sites through it - what looks polished on screen often turns to noise once the pixels are gone. fun experiment (link to the full leaderboard below...you can also run the scoring on ur own website and let it generates the llms.txt file for you), but it points at something bigger: agents need a different software stack than human users do. nearly every system we use carries the same assumption: a human sitting in front of a screen, clicking, typing, reviewing, approving. agents remove those constraints - and they'll use these systems far more heavily than we ever did. not everything needs to be rebuilt. three opportunities stand out to us right now: → execution environments: agents make writing code nearly free - the bottleneck shifts to verifying a change actually works. a clean sandbox proves nothing, because the code depends on the database state, the payment webhooks, the feature flags around it. agents need to fork that entire world: snapshot, branch, test, merge or discard... and these environments become the testing and release infrastructure for everything agents build. → shared artifact systems: humans coordinate through file names, folders, and remembering which doc is the "real" one. that collapses when hundreds of agents modify the same work in parallel. the opportunity isn't another storage engine - it's a layer where every artifact carries its identity, lineage, and status wherever it moves. that layer becomes the shared memory of the company: it decides what agents can trust. → business process systems: ERPs, EHRs, and supply chain systems record outcomes, not reasoning. the ERP knows the discount was approved - not who made the exception or why. that judgment lives in Slack threads and people's heads, and it evaporates, which is why bolting AI onto these systems doesn't work. Salesforce, Workday, and SAP became giants by owning what happened. the next generation gets built by owning why it happened. the last platform shift rewarded whoever made software easiest for humans to use. this next one will reward whoever makes work possible for agents to do. full piece in the comment section below ⬇️ if you're building one of these primitives, i'd love to talk! co-authored with Joanne Chen and special thanks to Savin Goyal Nicholas Diao for you inputs!
Experience & Education
Volunteer Experience
-
Speaker
TEDxMarin
2019: https://www.ted.com/talks/joanne_chen_confessions_of_an_ai_investor/up-next
2023: https://m.youtube.com/watch?v=Eq2ivRatIl0
Publications
-
Anomalous Radio Quiet Zones in X-ray Bright Young Supernova Remnants G54.1+0.3 and 3C58
American Astronomical Society
Languages
-
Chinese
-
-
French
-
Other similar profiles
Explore more posts
-
Tom Lazay
An emerging VC manager's fundraising lessons... the first two funds are a grind. Now, we’re on our third fund, it feels like we’re almost over the hump, but fundraising never gets easy for most of us. I want to congratulate the emerging VC firms presenting at this year’s RAISE Global conference. As former RAISE presenters, and (soon-to-graduate) emerging managers, we thought we’d share this LP Translator, a lighthearted guide to decoding what LPs really mean during the fundraising process. Fellow GPs, which ones did I miss? 👇 The LP Translator 📣 "Let’s stay in touch.” Translation: We’re not interested. “We want to see your track record develop.” Translation: Either we don't believe in your strategy, or we’re focused on managers with more buzz. “We’re not allocating to new managers right now.” Translation: We’re not allocating to you right now. "Show us your deals so we can get to know you.” Translation: We’d like free co-invests if you get something hot. “We need to see more DPI before we commit.” Translation: We don’t really understand VC, but we’re pretending to. “We’re fully allocated for this year; check in early next year.” Translation: Next year we’ll still be fully allocated (just not to you). “Call us before final close.” Translation: I’m too polite to say no at this time, so I’m kicking the can down the road. “Your fund is too small.” Translation: Okay, that one might actually be true (for some LPs). “We went through your data room and want to meet face-to-face.” Translation: We’re genuinely interested, keep going! “Can you send us your LPA for signature?” Translation: Let’s go! 🚀 -------------------------- Fundraising is a long game, longer than we ever expected. We're now seeing how LP relationships are built across several funds, not several months. If they’re investing time to learn about you and your strategy, that’s your best signal of real interest. #emergingmanager #venturecapital #LP #RAISEGLOBAL
1 Comment -
Daniel Dart
🚨NEW EPISODE: Recorded live at FUTURE TITANS 2026 - Jeff Perry of Carta sat down with the iconic Seth Levine, co-founder of Foundry. Seth has been in venture for 25 years, built Foundry from scratch as an emerging manager himself, and has backed about 50 emerging manager funds through his fund of funds. He has genuinely seen every side of this table. They went deep on building Foundry, why VCs are in the influence business, not the decision business, and why the concentration problem in venture is not only bad for LPs, but also for the innovation ecosystem overall. And why Seth's new book, Capital Evolution, is so important for the future of America. 🎧 Links to listen... Apple: https://lnkd.in/ehQUQ2EM Spotify: https://lnkd.in/eU4FExpg
1 Comment
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content