Building the AI-Native Analytics Function
Trust, accountability, and independence in an AI-native analytics function.
Product analytics has always existed to help a company make better decisions. Queries, dashboards, and reports are how that work got done, but in my experience leadership judges the function by three things: whether they can rely on the numbers, whether someone is accountable for them, and whether the analysis is independent enough from the outcome to be believed.
In The AI-Native Analytics Organization, I argued that agents create an opportunity to redesign the analytics function around the decisions it supports. This essay steps back to the question underneath that one. For founders and analytics leaders, it is about what should and should not change as agents take on more of the work. For practitioners, it is about where the value they bring has always been.
What the function is valued for
Trustworthy data starts with shared definitions that hold up when a decision depends on them. If revenue, active users, and retention mean slightly different things to finance, product, and marketing, every meeting begins with reconciling numbers rather than deciding what to do.
Accountability is the second. When a number turns out to be wrong, someone needs to explain why, correct it, and make sure the same mistake doesn’t reach the next decision. Trust in data is less a property of a table than of the people and practices standing behind it.
Independence is the third. Product managers want their launches to work, engineers answer the question they were asked, and every team naturally favors the reading that supports its plans. None of that is bad faith. It’s why a function whose main stake is getting the answer right is useful: it can say that a pricing change didn’t do what was hoped, or that growth came from somewhere other than the campaign getting credit for it.
The best analytics partners also don’t wait to be asked. They know which decisions are coming, size opportunities before they’re prioritized, and help set goals and forecasts the company can realistically be held to. They notice when something in the business is moving that nobody is discussing yet, and bring evidence early enough to change the direction of the conversation.
What agents change
Agents have changed the cost of producing analysis. An agent can find the relevant tables, write and debug queries, try several breakdowns, and draft an explanation in the time it once took to scope the work. They contribute to the reasoning too, proposing explanations and testing them against the data. A modern analytics function should be built around that capability rather than treating it as a faster way to do the old job.
Cheaper answers don’t remove the need for trust, accountability, and independence, and in some ways they make all three harder to maintain. Imagine a company that raises the price of a mid-tier plan and asks whether it worked. An agent can compare conversion before and after the change and write a clear summary. But if the price change coincided with a marketing push, a seasonal shift, or a change in how sign-ups are logged, that summary can be confidently wrong, and nothing in the writing signals the problem. Someone has to decide beforehand which comparison could separate those explanations, check the result against numbers the company already trusts, and say plainly how confident the conclusion is.
When anyone can ask an agent, people also arrive at meetings with different numbers built on slightly different definitions. And agents depend on context: metric definitions, known data quirks, business rules, and past corrections. That context has to be built, kept current, and checked, which is a responsibility in its own right. Without someone who owns it, a company can lose the shared view of its business that good decisions depend on.
Agents also tend to answer the questions they are given. They can monitor far more of the business than any person could, and that is a real opportunity, but broad monitoring produces mostly noise. Deciding which movements matter, which questions leadership should be asking, and when to raise them depends on knowing the company’s priorities and the decisions in front of it.
What the function could look like
An AI-native analytics function could cover more of the business with fewer people per decision area. A new function might start with one accountable owner; an established team would move toward owners of broader decision areas. Either way, the owner would:
- define and maintain the core metrics and goals
- build and govern the agent workflows, including the context they use and the checks applied to their output
- help shape what the company instruments
- stay close to leadership, sizing opportunities and raising questions before decisions are made
Routine questions could move to scaled tools and workflows as those prove reliable, while consequential ones would get closer attention.
Where to start depends on where a company is. A company building the function for the first time can ask which decisions over the next year matter most, and what evidence leadership would need to rely on for them. A company reshaping an established team can separate the work of producing analysis, which agents can increasingly take on, from the work that makes it trustworthy, accountable, and independent, which deserves more room. For practitioners, the same distinction applies. The queries and reports were always the means; the value was numbers people could rely on, someone willing to stand behind them, and evidence that arrived in time to change a decision.