Context Is the Bottleneck: What AI Agents Really Need

AI-built analytics rarely fails on the data. It fails on the things nobody wrote down. Why context, not the model, is the real constraint.

AI does not fail at building analytics because the model is weak. It fails because it cannot see why a measure is defined the way it is, or which report nobody trusts. Context is the asset, not the model alone.

In practice, this is where our AI work comes in, and AI Does Not Fix Bad Data. It Amplifies It covers useful related ground.

When an AI-built dashboard comes out wrong, the instinct is to blame the model or the tool. Usually it is neither. The agent did exactly what it was told, against the data it could see. What it could not see was the context: the part of the answer that lives in people's heads, in old decisions, in the reasons behind the rules. That missing context is the real bottleneck, and it is the thing worth fixing.

What the agent cannot see.

A semantic model tells an agent what the data is. It does not tell it what the business means. There is a difference, and it is where most of the value and most of the risk sits.

Consider what an agent does not know unless you tell it. It does not know that "revenue" in finance excludes intercompany sales, while "revenue" in the sales report does not. It does not know that one historic dashboard is quietly wrong and nobody trusts it, so building something that agrees with it is a failure, not a success. It does not know that a measure is defined the way it is because of a decision made eighteen months ago that nobody has written down. It does not know this quarter's priorities, so it cannot tell which cut of the data really matters.

None of that is in the model. It is scattered across finance, operations, old email threads, and the memory of whoever has been there longest. The agent works blind to all of it, and a confident answer built blind is still blind.

The data tells an agent what is true. The context tells it what matters. Most teams have only written down the first one.

Why this is the thing that decides outcomes.

When context is missing, an agent does not stop. It guesses, or it brute-forces its way to something plausible, and plausible is exactly the answer you do not want, because it survives a quick look. Every gap in context is a place where the output can be confidently wrong and a place where you burn time, and tokens, rediscovering things the business already knew but never recorded.

The flip side is the opportunity. The teams that pull ahead with AI will not be the ones with the cleverest prompts. They will be the ones whose context is closest to hand, because their agents spend their effort answering the question instead of guessing at the rules.

Context is an asset you can build.

The good news is that this is fixable, and it is ordinary work rather than anything exotic. Write down what your core metrics mean, precisely, including the exclusions and the edge cases. Capture the decisions behind the definitions, so the "why" outlives the person who made the call. Flag the reports that are not trusted, so nothing gets built to match them. Keep this close to the model, not buried in a wiki nobody opens.

Done well, this is not documentation for its own sake. It is the thing that makes every AI-assisted build faster and safer, and it happens to make your human analysts better too, because they were working without that context as well.

What captured context looks like.

This does not have to be a grand programme. The most useful version is small and specific. For each important metric, a short entry: the plain-language meaning, the precise inclusions and exclusions, the grain it is measured at, and a line on why it is defined that way. Alongside the metrics, a short register of the decisions that shaped them, and a frank note on which existing reports are trusted and which are not. That is most of it. A handful of pages, kept current, sitting where the model and the people both reach it.

The value is that it serves two readers at once. A person picks it up and stops guessing. An agent is handed it and stops brute-forcing. The same document that makes a new analyst productive in their first week is the document that makes an AI build land correctly the first time.

The advantage compounds.

There is a competitive edge here that is easy to miss. As building gets cheap, the differentiator is no longer who can build, it is who can build the right thing fastest. The team whose context is written down and close to hand gets correct results in one pass, while a rival's agent is still guessing at the rules and running up the cost of getting it wrong. Context is slow to build and hard to copy, which is exactly what makes it worth building.

What we would do this quarter.

Write down the definitions of your ten most important metrics. Not the formulae, the meaning: what is in, what is out, and why. That single document is the highest-leverage thing you can give an agent, and the fastest way to find out how much of your business logic currently exists only in someone's head. It is also, conveniently, the start of treating context as the asset it is.

If any of this sounds familiar, talk to us about your data.

Context is the bottleneck, not the model