---
title: "What 375 AI Builders Actually Ship"
description: "Production AI is more mature \u0026 hybrid than expected : 70% use open source, agents connect to databases over chat, RLFT delivers 16-30% lift, \u0026 context remains the #1 unsolved problem."
categories: ["AI"]
date: 2025-11-16
lastmod: 2026-07-17
canonical_url: https://www.tomtunguz.com/ai-builders-survey-2025/
author: "Tomasz Tunguz"
---


70% of production AI teams use open source models. 72.5% connect agents to databases, not chat interfaces. This is what 375 technical builders actually ship - & it looks nothing like Twitter AI.

{{< email_image src="subhhh3l1crt49axxrt5" alt="350 out of 413 teams use open source models" width="540" height="947" >}}

70% of teams use open source models in some capacity. 48% describe their strategy as mostly open. 22% commit to only open. Just 11% stay purely proprietary.

{{< email_image src="pmd3rt98lw1greupxqav" alt="Agents access deep systems: databases, web search, memory, file systems" width="540" height="366" >}}

Agents in the field are systems operators, not chat interfaces. We thought agents would mostly call APIs. Instead, 72.5% connect to databases. 61% to web search. 56% to memory systems & file systems. 47% to code interpreters.

The center of gravity is data & execution, not conversation. Sophisticated teams build MCPs to access their own internal systems (58%) & external APIs (54%).

{{< email_image src="axmcfw24hllbdmruvygi" alt="85% use synthetic data for generating evals vs fine-tuning" width="540" height="483" >}}

Synthetic data powers evaluation more than training. 65% use synthetic data for eval generation versus 24% for fine-tuning. This points to a near-term surge in eval-data marketplaces, scenario libraries, & failure-mode corpora before synthetic training data scales up.

The timing reveals where the stack is heading. Teams need to verify correctness before they can scale production.

{{< email_image src="i5qpmittzj8ytn4b05v1" alt="Automated methods for improving context: prompt optimization, ablations, manual" width="540" height="370" >}}

88% use automated methods for improving context. Yet it remains the #1 pain point in deploying AI products. This gap between tooling adoption & problem resolution points to a fundamental challenge.

The tools exist. The problem is harder than better retrieval or smarter chunking can solve.

Teams need systems that verify correctness before they can scale production. The tools exist. The problem is harder than better retrieval can solve.

Context remains the true challenge & the biggest opportunity for the next generation of AI infrastructure.

[Explore the full interactive dataset here](https://survey.theoryvc.com/) or [read Lauren's complete analysis](https://theoryvc.com/blog-posts/ai-in-practice-survey).