The gap between AI demos and AI revenue is now a $50 billion problem, and venture capitalists are betting that warm bodies, not better models, will close it.
The Summary
- Insight Partners is launching a dedicated forward-deployed engineering practice to embed engineers directly with portfolio companies, solving the "billions spent, nothing shipped" problem plaguing enterprise AI
- Enterprise AI agents lack organizational knowledge, the contextual understanding that turns raw data into decisions that actually make sense for specific companies
- The real bottleneck isn't model capability anymore. It's translation: getting AI to understand what data means in your warehouse, your approval chain, your edge cases
The Signal
Companies are pouring money into AI like it's 2021 again, but actually getting AI into production remains a major challenge. The symptoms are everywhere: pilot programs that never scale, proof-of-concepts gathering dust, executive dashboards showing "AI investment" but no line on revenue impact. Insight Partners sees this clearly enough that they're building an entire practice around fixing it.
The diagnosis from Insight Operating Partner Pablo Dominguez: you need forward-deployed engineers. Not consultants who fly in quarterly. Engineers who sit in your Slack, attend your standups, learn your weird legacy systems. The firm is betting that embedding directly with customers is how you turn AI from science project to actual working product.
"The gap between AI demos and AI revenue is now a $50 billion problem."
Here's what both sources agree on: the problem isn't the models. GPT-4, Claude, Gemini, they're all good enough for most enterprise use cases. The problem is knowledge. Not data, knowledge. The difference matters.
Data is numbers in a database. Knowledge is understanding that when the inventory system says "14 units" it actually means "12 sellable units plus 2 damaged ones we haven't written off yet because Bob in accounting is on vacation." AI agents need this understanding to reason about situations and make decisions that don't immediately break things.
This is the unglamorous infrastructure layer of the agent economy. Before your AI agent can negotiate vendor contracts or optimize supply chains, it needs to know:
- What each field in your ERP actually tracks (and which ones are lying)
- Who has approval authority for what, and when that changes seasonally
- Which customers always pay late but never cancel, versus which ones are flight risks
- The unwritten rules that keep your operation running
MIT Tech Review frames this as connecting AI agents to enterprise knowledge, which sounds abstract until you've watched a $200k AI implementation confidently make recommendations based on a "customer satisfaction" field that hasn't been updated since 2019.
The Implication
The forward-deployed engineer model signals something bigger: AI implementation is becoming a services business disguised as a software business. The margins look different. The scaling looks different. If Insight Partners is building this capability at the portfolio level, expect every major VC to follow. Which means two things for founders: one, you're about to have access to implementation talent you couldn't afford to hire. Two, your investors now have visibility into exactly how bad your AI integration really is.
For enterprises, this is actually good news. The "buy software, figure it out yourself" model was breaking under AI complexity. Having engineers who've seen the same integration problems across 47 portfolio companies means you might actually ship something. Watch for this pattern: VCs competing not on check size but on implementation bench strength.