Your Agent Doesn't Have a Data Problem. It Has a Context Problem.
And you've seen this failure before.
a16z published a piece today on data and analytics agents. The headline finding: agents are essentially useless without the right context. They can’t decipher business definitions. They can’t tease apart vague questions. They can’t reason across disparate data effectively.
The framing was new. The failure pattern isn’t.
The Revenue Question
Here’s the scenario they walked through. An organization builds an internal data agent. Connected to the right sources. Hooked into a clean UI. Ready to field questions from the business.
First real query: “What was revenue growth last quarter?”
The agent stalls.
Not because it lacks access to data. Not because the model is weak. Because nobody told it how the business actually defines revenue. Is it ARR? Run rate? Does the fiscal quarter align with the calendar? Which of the three materialized views in the warehouse is the one the finance team actually uses?
The semantic layer that was supposed to answer these questions turns out to be stale. Written by someone who left last year. Not updated when two new product lines launched.
So the agent produces something. And it’s wrong. Or confidently approximate. Which is often worse.
This is not a data infrastructure failure. It is a context failure.
You’ve Seen This Before
I’ve been writing about the specification gap for a while now. The distance between what you can clearly define as “correct” and what the system actually needs to deliver.
The a16z framing of the context problem is the specification gap showing up at the organizational layer.
When an individual operator tries to get useful output from an AI model and the result is generic garbage, the problem is usually that they never defined what success looks like. No tone guidance. No format constraints. No negative rules. The model gave them the average of everything it’s ever seen because they didn’t tell it what they actually needed.
The data agent failure is structurally identical. The organization never built the equivalent of a well-specified prompt at the systems level. The “prompt” for an enterprise agent isn’t a few lines of text. It’s the entire corpus of business logic, tribal knowledge, canonical definitions, and operational context that makes a question like “what was revenue last quarter?” answerable at all.
Without that corpus, the agent is pattern-matching against a void.
What the Context Layer Actually Is
The a16z piece maps out what a working context layer requires. It’s worth sitting with the full picture because most organizations are wildly underestimating the scope:
The first layer is accessible, consolidated data. Table stakes, and still harder than it sounds.
The second layer is automated context construction. Using LLMs to scan query history, identify the most-referenced tables, pull business metric definitions from existing data modeling work. High-signal inference from what already exists.
But here’s the layer that actually closes the gap: human refinement. Automated construction can get you most of the way. But the most important context in any organization is implicit, conditional, and historically contingent. It lives in someone’s head. It’s the tribal knowledge that explains why you use Affinity for North American deals from 2025 forward but Salesforce for everything global before that. No scraper finds that. No LLM infers it. Someone who knows has to write it down.
This is where context construction stops being a technical problem and becomes an organizational one.
The Deeper Pattern
Here’s what ties this to the broader failure mode I keep writing about.
Organizations adopt AI on top of infrastructure that was never designed to support it. The modern data stack solved one problem. It centralized data and made it queryable. But it didn’t encode business logic. It didn’t preserve tribal knowledge. It didn’t build a system that understood how the organization thinks.
So agents hit the wall not because the models are weak but because the surrounding architecture has no memory, no context, no way to translate a natural language question into the right answer for this company, with its specific definitions, its history, its way of doing things.
The McKinsey data I wrote about earlier this year said only 6 percent of companies see consistent, durable value from AI. The a16z framing helps explain why. The other 94 percent are deploying intelligence into systems that were never designed to hold it. Agents can’t compensate for that. They surface the gap instead of closing it.
What This Means for Operators
Context engineering isn’t just a skill for individual AI users. It is the foundational design challenge for AI-capable organizations.
Building the context layer is unglamorous work. It requires data teams and business leaders to agree on definitions that have been fuzzy for years. It requires tribal knowledge holders to articulate things they’ve never had to put in writing. It requires ongoing maintenance as the business changes, because a context layer that isn’t updated becomes a liability faster than it becomes an asset.
Most organizations won’t do this work. They’ll keep trying to build agents on a foundation that can’t support them, then wonder why the technology isn’t delivering.
The organizations that will actually extract value from agents are the ones willing to do the structural work first. Define the problem. Build the corpus. Document what the organization actually knows and how it actually operates.
That’s not a data project. That’s an architectural shift.
And the companies that treat it as one will find that the question “what was revenue last quarter?” isn’t hard to answer at all.
Source referenced: — a16z, Jason Cui and Jennifer Li, March 2026
