Ashwin Srinivas | The a16z Show Summary

1. Key Themes

The Open Source Migration Is a Performance Story, Not a Cost Story

Decagon shifted 90% of its workflow to open source models, but the driver was latency and control for voice agents — not cost reduction. Cost savings were a "nice side effect." The deeper insight is that fine-tuned smaller models can actually outperform frontier models on narrow tasks.

"Even if you have a, quote, dumber model, you can get it to higher performance on that specific task. So when we fine-tune smaller, dumber models, it's that they're just not as general purpose, but on the specific task we want them to do, they actually outperform the large, smart, state-of-the-art models. We end up getting all three things. It is better at the task, it is cheaper, and it is faster." — Jesse Zhang 00:05:04

Decagon Built a Business Process Agent, Not a Customer Support Bot

The most strategically important framing in the episode: Decagon's core product is an agent that follows business process, not one that does customer support. This is why the product expanded naturally into inbound sales, operational workflows, and proactive outreach.

"The thing that we built, and we kind of built this intentionally from the start, was not an agent that does customer support well, but rather an agent that follows business process well. Executing on operational workflows, doing sales lead qualifications, answering customer support questions — at the end of the day, it's just an agent following a business process." — Jesse Zhang 00:48:59

The "Glass Box" vs. "Black Box" Deployment Model Is a Real Competitive Differentiator

Decagon's most recent customer switched from Sierra specifically because Sierra's FDE-heavy model created drag, opacity, and slow iteration. Decagon's productized, self-serve approach let the same customer spin up seven new journeys in a month, versus three over a full year with Sierra.

"Within basically a month, they spun up like seven new journeys on Decagon. And it'd taken a year to get three. So it's just kind of the speed of iteration. Some teams will really like this — hey, we have control of it, our teams, especially our non-technical people can come in and do things and they understand what's happening in the conversations." — Ashwin Srinivas 00:39:27

Decagon Labs Is a Model Factory, Not a Research Vanity Project

The R&D team's real function is compressing the cycle time between a new open source model release and a production-ready, fine-tuned, task-specific model. The model landscape changes so fast that this factory must run continuously.

"We find ourselves constantly training new models and deprecating old ones that are no longer relevant because maybe the open-source frontier has advanced a lot and the model out of the box can do a lot of things that it couldn't do before. Decagon Labs is in a way a model factory of sorts — we really built it to compress the time between new model coming out and useful fine-tuned-to-our-task model kind of popping out the other end." — Jesse Zhang 00:09:39

Forward Deployed Engineers Are a Transitional Role, Not a Permanent Strategy

Jesse draws a sharp line between FDEs as a product discovery mechanism (legitimate and necessary at early stage) versus FDEs as a permanent go-to-market motion (a path to becoming a glorified consulting firm). The key test: does their output become core product, or one-off custom work?

"Forward deployed engineers are necessary or newly necessary for early stage AI companies because the workflows are new. But long term, I think they should just be building product. Once you know what the workflow is, if you can productize it, you should productize it and then become a typical company with these scaling properties of a tech company. And if you can't do that, then you're just building a glorified consulting truck." — Jesse Zhang 00:21:12

"Forward deployed engineers eat pain and excrete product." — Jesse Zhang, quoting Shyam Sankar, CTO of Palantir 00:25:24

The AI Concierge Vision Reframes the Total Addressable Market

The long-term vision is that every customer interaction — reactive or proactive — should be handled by AI. This reframes Decagon from a cost-reduction tool into a revenue-generating interface that sits at the front door of a business.

"An AI agent should just be the front door of your business or your brand. And every interaction, whether it's reactive or proactive with a customer, should be handled by AI. We've already seen that AI is very good at that — customer service is like a huge pillar of that where it's all these inbound interactions, but why not have it also be able to do all these other things?" — Ashwin Srinivas 00:50:54

AI Creates Demand for Customer Support, It Doesn't Just Eliminate Supply

When the cost of support drops to near zero, companies make support more accessible, which creates more support interactions. This is Jevons Paradox applied to enterprise AI — the overall market expands rather than contracts.

"Ticket volume, the amount of customer support inquiries that we get per month, was like 50,000 a month. Once they started using us, they said, wow, turns out our customers have a lot of problems. They said, let us make support more easily accessible — so instead of it being buried within a support panel, let's put support on every page. People are more likely to get stuck, let's allow immediate support for even free users rather than paying users." — Jesse Zhang 00:16:30

Hiring Is Still the Binding Constraint, Not Model Capability

Despite being among the most sophisticated AI model users in the world, Decagon's binding constraint is human talent, not token availability. AI tools accelerate throughput but expand the roadmap proportionally — so every company keeps hiring.

"All the AI coding startups are hiring like crazy. If our competitors are going to use them and build more things, we need to build more things. If somebody else said, here's our roadmap, and now we can get through it in a third of the time and then they just stop hiring, we would just take that to mean, wow, we can get through it in a third of the time. Great. Let's build three times as much stuff." — Jesse Zhang 00:53:11

2. Contrarian Perspectives

Fine-Tuned Small Models Outperform Frontier Models on Specific Tasks — The Tradeoff Is False

Conventional wisdom says open source means sacrificing intelligence for cost. Jesse explicitly rejects this.

"Oftentimes when you see these debates being had on Twitter, the tradeoff tends to be, oh, do we want the smartest model that is very expensive or can we dumb it down a little bit and get it cheaper? I actually think that is a false tradeoff. Because what we've seen in practice is even if you have a, quote, dumber model, you can get it to higher performance on that specific task." — Jesse Zhang 00:05:04

This is backed by practice: 90% of Decagon's production workflows run on fine-tuned open source, not frontier models.

Open Source's Share of Enterprise Inference Is Actually Declining Right Now

Despite the enormous hype for open source, Ashwin notes enterprise open source inference share is currently shrinking — not because open source is losing, but because enterprises are spinning up net new use cases on frontier models first.

"If you look at enterprise right now, even though there's a lot of hype for open-source, the sort of share of open-source inference is actually going down right now, because people are spinning up all these new use cases, and if you're spinning up new use cases, of course you're going to use the frontier models until they're working." — Ashwin Srinivas 00:08:11

The implication: the open source wave is real but lagged, and will arrive as those use cases mature into production.

The Labs Won't Be the Last Startups — Because AGI Itself Will Still Need Software Infrastructure

The "labs eat everything" narrative ignores that agents, like humans, will need databases, CRMs, and workflow software. Human beings are effectively AGI and they still use software.

"We human beings are kind of AGI, right? And human beings have needed to use software for lots of things. You need databases to put stuff in. You need CRMs to track things. And I think even once you have AGI, all our AGI agents are going to need somewhere to store work and pull information from and reason about things. So I don't think software as a whole in any meaningful way is going away." — Jesse Zhang 00:19:30

AI Will Kill Jobs But Not Careers

Rather than the standard "AI augments workers" framing, Ashwin makes a sharper claim: the jobs being eliminated should not be done by humans because they are mundane and high-volume. But there is near-infinite valuable work to be done for customers, so people shift up.

"AI will kill jobs but not careers in a way. Because those jobs that are being done currently should not be done by humans — they're very mundane and menial. It's a super high volume use case. But there is actually a near-infinite amount of things that people could be doing to make their customers happier and take care of them more. And so people end up doing those things." — Ashwin Srinivas 00:17:25

X Influencers Have Replaced Billionaires as the True Priest Class

In a brief but striking aside, Ashwin referenced a Jeremy Giffon thesis that reframes power dynamics in modern society. The insight has direct implications for how founders should think about platform strategy.

"There's this guy Jeremy Giffon... he made some claim about how it used to be that the big status symbols in the world is like everyone wants to be a billionaire because he calls it like the priest class. But nowadays actually it's when people become billionaires, now they want to become X influencers because those people hold the real power because they can influence what the whole world is thinking about." — Ashwin Srinivas 00:11:45

3. Companies Identified

Decagon AI agent platform for enterprise customer interactions; co-founded by Jesse Zhang and Ashwin Srinivas. Mentioned throughout as the subject company. Runs 90% of its workflow on fine-tuned open source models, has expanded from customer support into inbound sales and operational workflows, and competes directly with Sierra. Has offices in San Francisco, New York, London, and Australia.

"The thing that we built was not an agent that does customer support well, but rather an agent that follows business process well." — Jesse Zhang 00:48:59

Sierra AI agent platform, direct competitor to Decagon. Mentioned as having a heavier FDE-driven, black-box deployment model. Lost at least one major customer to Decagon specifically because of this approach.

"When they worked with Sierra, it was mostly FDs and it just felt like a black box where the FDs were good, but they had to go through the FDs for everything. And over the course of the year, they may have built out — I think they said three journeys." — Ashwin Srinivas 00:38:27

Anthropic Frontier AI lab. Mentioned as a primary closed-source model vendor Decagon began with and still uses for 10% of workflows, particularly new/experimental products. Reasoning models from Anthropic power Decagon's Duet product.

"The reasoning models mostly for like the Claude Codes of the world, but they are also really good for Duet, for example. And that was kind of like a moment where it's like, oh wow, you can just see the improvement over time of the models." — Ashwin Srinivas 00:31:48

OpenAI Frontier AI lab. Co-mentioned with Anthropic as one of the original model vendors Decagon used, and as the basis of the "labs as last startups" narrative being debated.

"This narrative dominated the first half of 2026, which is that Anthropic, OpenAI, they're the last startups, they're going to take over everything." — Sarah Wang 00:07:13

Palantir Enterprise software and defense technology company. Cited as the originator of the forward deployed engineer model. Jesse Zhang previously worked there as a deployment strategist.

"Shyam, who's the CTO of Palantir today, had a phrase internally — forward deployed engineers eat pain and excrete product." — Jesse Zhang 00:25:24

Salesforce Enterprise CRM platform. Cited as evidence that horizontal enterprise software wins over vertical solutions, and as a model for Decagon's go-to-market strategy.

"Salesforce is very horizontal, Zendesk is very horizontal — all these solutions are very horizontal because you just gain more from having that scale and having a very robust and deep product than you do from having very vertical-specific features." — Ashwin Srinivas 00:10:02

Zendesk Enterprise customer service software. Cited alongside Salesforce as a horizontal platform that validates Decagon's horizontal build strategy.

"Salesforce is very horizontal, Zendesk is very horizontal." — Ashwin Srinivas 01:02:51

VMware Enterprise software company. Mentioned in passing — Ragnarom (Raghu Raghuram), former CEO of VMware, was brought on by a16z to help build out international capabilities.

"We brought on Ragnarom last year, obviously former CEO of VMware, and in addition to investing, he's helping us build out a ton of our international capabilities." — Sarah Wang 00:59:48

4. People Identified

Jesse Zhang Co-founder and CEO of Decagon. Former deployment strategist at Palantir. Wrote a viral piece on open source vs. closed source models. Technically deep and go-to-market oriented. Built a personal AI agent that maintains running business context to improve decision-making quality.

"I actually spent a while building agents for myself to capture all the business context well so it just kind of looks over my shoulder all the time and is constantly compiling context — here are the people that we've hired, here are the deals that we're working on, here are the current challenges that we have." — Jesse Zhang 01:07:17

Ashwin Srinivas Co-founder of Decagon. Spends approximately 80% of his time on sales, including managing relationships with some of the largest banks, airlines, and telcos in the world. Previously a solo founder. Featured in the New York Times in connection with Decagon's open source work.

"Probably most of my time, like 80% maybe. A lot of it is just pushing speed. And as one of the founders, you just have to be in calls — people want to meet the founders. But also it's just how can I be the sort of main force that's pushing our team to go faster." — Ashwin Srinivas 00:44:44

Shyam Sankar CTO of Palantir. Originated the internal phrase "forward deployed engineers eat pain and excrete product," now widely cited in Silicon Valley. Jesse Zhang attributed the phrase to him directly.

"Shyam, who's the CTO of Palantir today, had a phrase internally, I think now it's been written about a ton — he would say, forward deployed engineers eat pain and excrete product." — Jesse Zhang 00:25:24

Brian Chesky CEO of Airbnb. Mentioned as a cautionary example of a high-profile founder who faced public backlash for posting AI-generated content ("AI slop"), in the context of the discussion about authentic founder voice on social media.

"I have to bring up the whole AI slop fiasco — maybe fiasco is a strong word — that Brian Chesky just went through." — Sarah Wang 00:54:19

Jeremy Giffon Investor and podcast guest, appeared on Patrick O'Shaughnessy's podcast. Cited for his thesis that X influencers have replaced billionaires as the modern "priest class" — the people who truly shape societal thinking.

"There's this guy Jeremy Giffon... he made some claim about how it used to be that the big status symbols in the world is like everyone wants to be a billionaire because he calls it like the priest class. But nowadays actually, it's when people become billionaires, now they want to become X influencers because those people hold the real power." — Ashwin Srinivas 00:11:45

Raghu Raghuram (Ragnarom) Former CEO of VMware. Joined a16z and is helping portfolio companies build international go-to-market capabilities, specifically cited in the context of AI companies expanding internationally much earlier than expected.

"We brought on Ragnarom last year, obviously former CEO of VMware, and in addition to investing, he's helping us build out a ton of our international capabilities." — Sarah Wang 00:59:48

Patrick O'Shaughnessy Host of the Invest Like the Best podcast. Mentioned as the interviewer for both the Jeremy Giffon episode and a separate Jesse Zhang appearance where he discussed BPO displacement.

"On a podcast maybe Patrick O'Shaughnessy's podcast a couple months ago, you had mentioned that even your customers when they've been using BPOs for customer support, you haven't actually seen layoffs at the BPO." — one of the hosts 01:18:03

Mark Andreessen Co-founder of a16z. Mentioned for sharing his Claude system prompt publicly — a prompt designed to make the model aggressively disagree, which Jesse Zhang adopted and found highly effective.

"Mark at one point had posted the prompt that he used for Claude where it's like, you know, disagree with me, be very contrary — things like that. So I actually used that for a while. It was great. The funny story on this was I really enjoyed it because it would disagree with me very aggressively." — Jesse Zhang 01:09:16

5. Operating Insights

Build Evals That Measure the Entire System Outcome, Not Individual Model Tasks

Most companies evaluating AI models stop at task-level metrics. Decagon measures whether the entire model ensemble is delivering the end customer outcome. This systems-level eval framework is what makes fine-tuning actually work in production.

"We're not just saying, is this model good at this task? We're saying, is this model working in concert with all these other models delivering the end customer outcome that we care about? And because that is so unique to our setup, we found that in practice we've needed to build a lot of the infrastructure that we need to train these models and evaluate them." — Jesse Zhang 00:11:18

Map the Enterprise Deployment Journey Before the Product Demo

The technical capability of the product is table stakes for large enterprise deals. What actually accelerates close rates is being able to walk a buyer through the full path from first meeting to 100% live deployment — including their model risk governance, testing process, phased rollout, and error remediation.

"When we walk in to one of these enterprises, we can walk them through in very granular detail how we go from this first meeting today to going live at 100%. At this stage for a company like you, this is what your model risk process is likely to be. This is how your testing process should look like. This is how we should do the initial rollout. The product and technology part of what we sell is important, but for these large companies, equally important is us helping them think through the process to actually get this deployed and at scale." — Jesse Zhang 00:42:43

Productize Pain Before It Becomes a Support Burden

Every Decagon FDE task that requires repeated manual effort gets immediately flagged for productization. Agent Operating Procedures (AOPs) themselves were invented because writing procedures in code was taking too long. Duet Autopilot was built because reviewing conversations was taking too long. The operating cadence is: FDE does it manually → identify the pattern → ship it as product.

"Every single thing that Jesse just talked about was the result of forward deployed people doing things and us figuring out how to productize it. Before AOPs, you would have to write all these procedures in code. We found that it's taking a lot of forward-deployed engineering work. Wouldn't it be so much easier if we could productize it by writing it in plain text?" — Jesse Zhang 00:32:54

Use Cross-Functional Teams as a Competitive Velocity Weapon

Decagon deliberately blurs org boundaries — engineers on early sales calls, salespeople debugging products, APMs spanning both ends. This is not cultural chaos; it is a deliberate mechanism to shorten the feedback loop between customer reality and product decisions.

"You will very commonly see engineers on early stage sales calls. You will see salespeople like debugging parts of the product. Because it's so many different teams working together, it's kind of a — we're all in this together to cross the line." — Jesse Zhang 00:56:58

Build a Persistent Business Context Agent for Executive Decision-Making

Jesse built a personal AI agent that passively monitors his activities and continuously compiles running context on hiring, deals, challenges, and decisions. This eliminates the bottleneck of re-explaining context on every query and dramatically raises the quality of AI-assisted decisions.

"I actually spent a while building agents for myself to capture all the business context well so it just kind of looks over my shoulder all the time. Later on I can just go to it and say, hey, there's this new person that we're thinking of hiring, what do you think — and now it's able to automatically say, well, we had two candidates that were very similar and these candidates had these kind of drawbacks, so this is probably not the best person." — Jesse Zhang 01:07:17

6. Overlooked Insights

The Real Moat Is Enterprise Deployment Infrastructure, Not the AI Model Itself

This was stated almost in passing and the hosts didn't press on it, but it's the most important strategic claim in the episode. Jesse argues that even if models were perfect today, the software infrastructure to make them deployable inside a large enterprise — access controls, collaboration tooling, testing against regulatory lines, legacy system integration, conversation analytics — doesn't exist as a commodity yet. Whoever builds and owns that layer controls the enterprise AI market regardless of which foundation model wins.

"To make models like this deployable within the enterprise, you need to say, okay, I need a way to be able to tell the model what it can and cannot do. Then I need a way to make sure that hundreds of people within the enterprise can collaborate to make sure that the agent is behaving as expected. Then I need a way to be able to test this model and make sure that it doesn't cross any regulatory lines. Then I need a way to look over millions of conversations that happen for me to extract insights for the rest of my teams. So there's a lot of just infrastructure and software that you need to build around these models to make them deployable within an enterprise." — Jesse Zhang 00:35:25

The investment implication: the companies building this enterprise AI deployment layer — compliance tooling, model governance, multi-stakeholder agent management, conversation intelligence at scale — may be the most durable businesses in the AI stack, more durable even than the application companies themselves.

Voice-to-Voice Models Are the Next Latency Frontier That Will Force Another Stack Rebuild

Mentioned in a single sentence and immediately dropped, but significant: Ashwin flagged voice-to-voice models as a technology actively being watched. Given that Decagon already rebuilt 90% of its stack around latency requirements for its existing voice agent, a step-change in native voice-to-voice capability would likely trigger another major architectural shift — and the companies that build fine-tuning and deployment infrastructure for voice-to-voice models first will have a substantial head start.

"Voice-to-voice models is an interesting frontier that there's still research happening on — the getting smaller models to be smarter out of the box. There are going to be still developments that we are watching closely that we care about." — Ashwin Srinivas 00:54:19