Data Swamps: Where AI Agents Sink.
Understanding the pitfalls of proprietary data moats in due diligence.
It seems every AI startup today pitches a massive, proprietary data or algorithmic moat. Yet, during technical diligence, we usually find one of two things: the moat doesn’t exist, or unlocking it requires extensive, custom consulting work before a (typically free) pilot can even kick off.
When a software startup requires heavy services just to get a pilot live, they aren’t a SaaS business—they are a services firm in disguise.
Let’s dig into why this usually happens and how to spot it in diligence.
57 Ways to Spell Philadelphia
Steven Rich, an investigative data reporter for the Washington Post, famously noted that in 2020 PPP loan dataset, there were 57 different ways to spell “Philadelphia” – and 31 different ways to spell Chicago (!).
It’s no wonder, then, that models trained on unprocessed, raw data succumb to the “garbage in, garbage out” paradigm.
Clean Data Doesn’t Exist
Having worked with hundreds of companies as a consultant, advisor, and investor, I can count on one hand the number of organizations tracking customer activity seamlessly across every business touchpoint. Clean data simply doesn’t exist.
When a startup skips the thankless, brutal task of building a single source of truth across siloed systems and leaps straight to deploying “autonomous AI agents,” they are building a skyscraper on quicksand and hoping a tube of caulk holds the foundation together..
Why Gartner’s Failure Rate Predictions Are Too Optimistic
Gartner recently predicted that over 40% of AI Agents will be scrapped by 2027 because they can’t “autonomously achieve complex business goals or follow nuanced instructions over time. That number is way too low!
It ignores the underlying mountain of data debt. You can’t automate a process with an AI agent if the inputs it relies on have typos, duplicates, or just don’t exist!
The Diligence Playbook: How to Spot the Quicksand
