|
Huang is right: agents will use software tools, not replace them. But which layer of the stack captures value is shifting. For two decades, SaaS commanded premium valuations because it owned the application layer: the interface, the dashboard, the workflow. AI agents are increasingly becoming that layer themselves, and when the agent is the interface, the traditional software application loses its hold on the user and much of its pricing power.
Beneath the application layer sits the infrastructure layer: proprietary data, regulated processes, domain-specific judgment, and money movement rails — the back-end that makes an agent’s output accurate, compliant, and actionable. AI compresses the value of the application layer. It compounds the value of the infrastructure layer. Value is not leaving software. It is migrating from the surface to the foundation.
The market is treating this as a story about replacement. We see it as a story about reorganization. The application layer is being absorbed by AI interfaces that perform the same functions faster and cheaper. But those interfaces still need something to call into: clean proprietary data, auditable processes, domain logic refined against real-world edge cases. Running a payroll, closing a ledger, adjudicating an insurance claim: these are not tasks where a probabilistic answer is good enough. These companies are not the casualties of this shift. They are the infrastructure layer the new application layer cannot function without.
Where We Focus
At MissionOG, we have been spending considerable time on this question: in a world where code generation costs approach zero, what is actually defensible?
Our framework is straightforward. Better code generation compresses the advantage of building software, but increases the advantage of owning unique data and being embedded in critical workflows. In the internet era, distribution was the moat. In the AI era, operational data is the moat. We look for one or more of the following:
Proprietary, opinionated data. Companies that generate proprietary datasets through normal customer operations create feedback loops that competitors cannot quickly replicate. But data alone is not enough. The companies that also encode best practices, expert judgment, and institutional knowledge into their products – what practitioners call tradecraft – create a layer of intelligence that general-purpose models cannot approximate from public training data alone.
Embedded workflows in high-consequence domains. In domains that require precision, the last mile is the moat. The gap between a working prototype and a production system is filled with approvals, exceptions, audit trails, and deep integration into cores, ledgers, and legacy workflows. In these domains, almost right is entirely wrong – the cost of being off by a fraction is regulatory, financial, or reputational. AI makes the first mile easier. It does not shorten the last.
Regulatory moats and money movement infrastructure. Financial services represents one of the largest addressable markets in the world, yet compliance is not a feature. It is the product. The liability sits with the institution, not the vendor, creating structural defensibility that no agent can shortcut. And anything that touches money movement is never pure code. The industry still runs on legacy infrastructure and manual workflows. Moving money requires bank partnerships, regulatory licensing, and operational trust built through years of processing real transactions.
Our Thesis in Practice
This framework has been the core of our investment thesis across four funds and 40 portfolio companies since 2013.
Sayari — Over 10.6 billion records, 500 million companies, and 2 billion entity relationships. Sayari integrates global corporate ownership, supply chain, and trade transaction data from over 250 jurisdictions to surface counterparty risk. The differentiation is the continuously expanding dataset and compliance intelligence, not the application layer.
ValidiFI — Predictive bank account and payment intelligence across 51 million consumers and 63 million bank accounts. ValidiFI detects fraud, identifies synthetic identities, and provides compliance infrastructure to financial institutions. The moat is the proprietary view into real-time account behavior, built through direct integrations with banks and payment networks over years.
Brightfield — Aggregates roughly $640 billion of contingent workforce spend through a give-to-get data exchange with enterprise customers. The product improves as participation increases, and the dataset is not reproducible by an individual enterprise or by an AI model trained on public data. This is a classic network effect moat.
Roots — Has trained its models on more than 40 million insurance claims. The value is not the code base. It is the accumulated operational learning embedded in the model and underwriting workflows, the kind of domain-specific intelligence that general-purpose AI cannot replicate from public data.
Where We See Opportunity
Every platform shift follows the same pattern: development costs fall, more companies enter, and the gap between winners and losers widens. The cloud transition proved this. As hosting costs dropped and SaaS proliferated, the durable winners were companies that combined proprietary data with deep workflow integration, not the ones with the best UI. The mobile era followed the same arc: app development costs collapsed, millions of apps launched, and the lasting value accrued to platforms with network effects and data gravity, not to the apps themselves.
The AI transition is following this pattern with unusual speed and severity. Valuations for public horizontal SaaS have compressed to levels not seen in a decade. Private SaaS companies are raising down rounds or delaying IPOs. Strategic acquirers are buying data assets and risk platforms rather than generic productivity software.
For investors focused on the operational layer, the current environment is creating real opportunity. Competition for deals is thinner. Valuations are more attractive. And the thesis is being validated in real time by the very market forces driving the selloff.
Our view is that AI widens the moat. It does not fill it in. The same forces compressing the value of commodity software are simultaneously increasing the value of proprietary data, regulatory infrastructure, and embedded workflows. We have been investing against this thesis for over a decade. The current market is accelerating the relevance of everything we look for. |
