Agentic Scaffolding, IDEs & Demand Gen
Capital Efficient #4
Weekly Radar
Vertical AI vs. The Foundation Models
Every enterprise-focused app is now in quiet co-opetition with the foundation models, even if they haven’t realized it yet. In horizontal markets, this is already a given: OpenAI launched a browser today. They will undoubtedly launch a CRM in due time, they will go after Google’s ad business, and they will continue to build out large, mass-market applications for core business tasks. Imagine what Google did with G-Suite but on steroids.
Recent remarks I heard at a talk by Anthropic’s Chief Customer Officer were enlightening on what this means for the vertical software players. In his talk, he explained that for very large, regulated orgs (e.g., hospital systems, insurance carriers), they are currently sending teams of what are effectively FDEs to work with these huge enterprises and get them consuming Anthropic tokens. Anthropic is positioning itself as a safer, more regulation-friendly alternative to OpenAI with the aim of winning over more risk-averse institutions.
In my last post, I wrote about Custom Automation Platforms and how they represent an alternative way to build vertical AI for the enterprise. The foundation model labs are clearly planning to spin up their own version of this, as a way to directly capture compute spend and cut out the software-provider middleman.
The rub: many vertical AI startups pitch investors by explaining how they will start in the SMB or mid-market of insurance or healthcare, and eventually they will earn the right to move up market and win huge enterprise contracts. To do so, they will need to compete not only with the CAPs, but also with the foundation models that drive the core of their products.
IDE for X
AI agents are here, and they work, kind of. But there is increasing awareness that they will need more contextual infrastructure and guidance than initially anticipated. While AGI is reportedly on the horizon, many practical gaps remain.
Coding is well suited for an agentic workforce. There’s an established IDE and there’s fairly clear rules about what “good” looks like. Deploying agents in other verticals is more complicated. For most roles, tasks are multimodal, only loosely codified, and touch on many different working environments.
For example, an insurance broker may find themselves needing to log into and update their own CRM, a commission-tracking platform, email, and an insurance carrier portal, all before lunch time. The agent-meets-world challenge can be seen in the rush of capital into RL-environment startups and the immense sums the foundation labs are spending on RL training. This can also be seen in the glut of browser agent startups looking to give AI agents the ability to reliably navigate the browser-based workflows that make up most of today’s white-collar work.
Founders face a fundamental choice: try to build AI agents on top of existing legacy systems vs. take the time to rebuild the scaffolding their agents need to thrive.
The tradeoff depends on the industry. In slower, legacy categories, you often have to play the existing game; in faster, tech-forward sectors, there’s an opportunity to rip-and-replace the stack and rebuild for agents. Startups taking the rip-and-replace approach tend to focus on AI-native companies and provide them with new infrastructure. For example, Campfire and Rillet’s ERP customers are Series A–D startups, not traditional businesses. Serving legacy firms is harder and requires more integration, but there are far more of them to go after.
News Round Up
Serval: The demand for “AI for IT” continues apace, with Serval announcing a $47MM combined Seed/A round led by General Catalyst. Instead of laying automation onto legacy systems, Serval’s bet is to build an AI-native IT automation from the ground up. Most of their public customers are startups (e.g., Clay, Mercor) where there’s less legacy ops debt to tackle.
Starting fresh helps avoid some of the challenges I outlined above - it’s much easier to build something like an IDE for IT Ops, and then build an agent on it vs. letting it run wild on your existing, human-centric systems.
OpenEvidence: Medicine is proving to be a field where AI - with just its current capabilities - can have a big, real-world impact. Yesterday, OpenEvidence announced a $200MM Series C at a $6B valuation spearheaded by GV, Coatue, and Thrive. The business aims to supercharge doctors and nurses by allowing them to turn patient symptoms into clear diagnoses via natural language search, and helping them with workflow automation around things like prior-auth requests.
I’ve had some of the highest-paid medical specialists in NYC click off Epic and turn to WebMD - at a minimum, OpenEvidence is an improvement on that.
What I Am Reading
Karpathy x Dwarkesh - This fits more under “What I Am Watching” but Andrej Karpathy sat down with Dwarkesh Patel for a discussion on the state of AI that went briefly viral. I’m still digesting it, but depending on who you ask about the interview, it either means that the AI bubble is popping or that the opportunity has never been greater. Decide for yourself.
Why Vertical SaaS is Riding the Waves of AI to New Heights - Great piece by Shomik Ghosh giving his take on how recent advances in AI capabilities will accrue to vertical software co’s.
BVP Demand Gen Guide - Bessemer put out a solid overview of B2B demand generation tactics for the modern startup. Worth a read for anyone trying to crack mid-market / enterprise buyers.