Frontier model capability isn’t the bottleneck for agentic AI anymore. The bottleneck is your data — and nobody’s going to fund fixing it.

GPT-5.4 shipped this week with native computer use, a 1-million-token context window, and 33% fewer factual errors than its predecessor. Gemini 3.1 Flash-Lite processes 363 tokens per second at $0.25 per million input tokens. Qwen 3.5’s 9-billion-parameter model now performs close to models several times its size, under an Apache 2.0 license that means anyone can run it. The models are outrunning the ability of most organizations to actually use them.

Here’s the harder-to-quote part: a survey of chief data officers published this week found that 50% of organizations already deploying agentic AI cite data quality and retrieval issues as their primary deployment barrier. Not model capability. Not compute costs. Not integration complexity. Data.

This is structural, not accidental.

Enterprise data is a disaster by default. The CRM has duplicate records, missing fields, and entries last touched three product generations ago. The internal knowledge base exists in seventeen formats, none consistently tagged. Documentation was accurate once, then the team shipped four updates and nobody had time to revise it. These aren’t edge cases. They’re the median state of organizational knowledge — accumulated fast, maintained inconsistently, never cleaned up because cleaning doesn’t ship a feature.

No model improvement addresses any of that.

What’s interesting is that the models themselves seem to have learned this. GPT-5.4’s most practically useful new feature isn’t the computer use mode or the expanded context — it’s “tool search,” which reduced token usage by 47% on a 250-task benchmark by being more disciplined about when and what to retrieve. The model is doing retrieval hygiene because it has to. It learned that garbage retrieval is the failure mode. The capability investment went toward working around the problem, because the problem itself isn’t something a model release can fix.

This is the demo problem. In any system where funding follows visibility, the work that produces a demonstrable artifact gets resourced. The foundational work that makes everything else possible — but has no deliverable of its own — gets deferred until failure makes it impossible to ignore anymore.

You can demo computer use. You can show an agent navigating a desktop, filling a form, extracting data from a screen. That’s a compelling five-minute video. You cannot demo a well-structured, consistently maintained enterprise knowledge base. There’s no five-minute video for “we spent six months cleaning our retrieval layer.” There’s no benchmark. There’s no press release. There’s no moment where the CEO’s eyes light up.

So the investment goes to capability — which does demo — and the data infrastructure sits there accumulating more technical debt, waiting.

The companies shipping agentic AI that actually works aren’t the ones with the biggest model budgets. They’re the ones that made a boring, unglamorous, internally invisible investment in their data before they touched a model API. They know what’s in their knowledge base. They’ve removed the duplicates. They have metadata standards that someone actually enforces. The AI part was almost an afterthought.

The models are ready. They’ve been ready for a while, for most practical work. The question is whether the infrastructure underneath them ever gets the same attention. Based on where the incentives point: probably not until the failure pattern is too obvious to explain away.

The unglamorous always determines the ceiling. We just keep funding the glamorous anyway.