StratoData Logo
    Content/Blog
    BlogAI Analytics

    From BI to AI Analytics: The Questions Nobody Asked

    Dashboards answer the questions you already knew to ask. AI analytics changes who gets to ask, how many questions get answered and how deep the answers go. What shifts, and what has to be in place before it works.

    Martin Faith··7 min read

    20–30×

    More data questions within weeks of Ramp opening its AI agent to everyone

    ~50×

    More questions after six months

    16% vs 8%

    What the data said vs. what the room guessed at Fivetran

    1

    Data leader in a year who thought AI could replace most of their team

    The Short Answer

    What Changes When You Move from BI to AI Analytics?

    Dashboards answer the questions you already knew to ask. AI analytics removes the queue in front of every other question. When Ramp opened an AI research agent to the whole company, the number of data questions grew 20 to 30 times within weeks and about 50 times within six months. Those questions existed all along. Nobody asked them, because getting an answer took too long.

    The catch is the foundation. An AI agent is only as good as the data and the business context it works with. Here is what shifts, and what has to be in place before it works.

    Source

    The figures and quotes in this article come from The Analytics Engineering Podcast (dbt Labs), “Open Data Infrastructure in the age of AI agents”, 15 October 2026. Guests: George Fraser (CEO of Fivetran and dbt Labs), Tristan Handy (co-founder of dbt Labs and host) and Martin Casado (general partner at Andreessen Horowitz). The interpretation and the five steps are ours.

    The Queue

    Why Does a Dashboard Only Answer Yesterday’s Question?

    Traditional BI works like this. Someone has a question, the data team builds a report, and the report keeps answering that same question every morning. It is reliable, it is useful, and it is frozen. It is great at answering “what is happening?”

    The moment the question changes, you are back in the queue. Write a ticket, wait for an analyst, get an answer next week, at best. Every data team we know has more questions than analysts. That queue is not a sign of a bad team. It is how the model works.

    AI analytics does not replace the dashboard. It removes the queue: you get the answer when the question comes up. If you are deciding which reports to keep, read Power BI vs AI Analytics: Will Talking to Your Data Replace Dashboards?

    The Difference

    How Is an AI Analytics Agent Different from BI?

    Best at
    Classic BIKnown, recurring questions: what is happening?
    AI analytics agentNew questions: why did it happen, what else is going on?
    Time to an answer
    Classic BIA ticket, then days or weeks
    AI analytics agentWhen the question comes up
    Hypotheses checked
    Classic BIThe two or three most likely
    AI analytics agentThe long tail, one by one
    You pay in
    Classic BIAnalyst hours and headcount
    AI analytics agentTokens
    Business context
    Classic BIIn the analyst’s head
    AI analytics agentMust be written down (context engineering)
    Number of questions
    Classic BICapped by the size of the data team
    AI analytics agentGrows with demand: 50× at Ramp

    The Long Tail

    Why Does an AI Agent Check More Causes than an Analyst?

    Revenue dropped in one market. A good analyst looks at the two or three most likely causes, finds something plausible, and hands it to the business. The rest of the search space stays untouched. Not because the analyst is lazy, but because there is no time.

    The analyst checks

    Three causes

    The agent checks

    Thirty

    An agent can keep going. It can test hypotheses where nothing points in any particular direction and every explanation is roughly as likely as the next. That long tail is where the surprises live: the one shipping option, the one product category, the one campaign quietly eating margin.

    The same search catches emerging trends, positive or negative. Catching them early matters. You can double down on an opportunity fast, and fix a threat before it becomes a real problem with a bigger price tag. Martin Casado calls this work “spelunking” through a long tail of tedious problems, and now you can have as many agents doing it as you need.

    The Economics

    What Does an AI Data Analyst Cost, and What Does It Need to Know?

    An agent is available at any time. You pay for it in tokens, not in headcount. That changes the economics of asking a question from “is this worth an analyst’s afternoon?” to “why not just ask?”

    There is a catch. An agent starts every conversation from zero. It does not know that “revenue” in your company means net of returns, or that the Czech shop switched ERP in March. A human analyst carries that tribal knowledge in their head. An agent needs it written down.

    Context engineering

    Write down what your best analyst knows.

    Context engineering is the practice of systematically describing your business to an AI agent: metric definitions, business rules, known data quirks. Then each analysis does not start from zero, it builds on what is already known.

    The longer that context is collected and kept clean, the better the agent understands your business. How we do it →

    Latent Demand

    Why Do Data Questions Grow 20 to 50 Times?

    Speed is not the only change. Demand for analysis grows too.

    Ramp launched its research agent in Slack and opened it to anyone. Tristan Handy expected about as many questions as the data team used to get, maybe double. Within weeks it was 20 to 30 times as many. After six months it was about 50 times.

    “That’s the latent demand in the organization to know things.”

    Tristan Handy, co-founder of dbt Labs

    Every company has this demand. Sales managers who never asked why one region underperforms. Category managers who guessed instead of checking. The questions were never asked, so nobody knew they were missing. The path to the answer used to be too long.

    The top few answers to a problem usually explain a large part of it. Pareto tends to work. But if you want to understand the whole business situation, the devil is in the details. People rarely have the time or patience to go all the way. As long as there are tokens, an AI analyst does.

    Gut Feeling vs. Data

    How Wrong Is Gut Feeling Compared with One Query?

    George Fraser described a meeting at Fivetran. The team discussed what share of Global 2000 companies use their product, and what share use one particular SKU. The guesses in the room were 8% and 4%. A query against the data returned 16% and less than 1%.

    Global 2000 using the product

    Guess 8% → data 16%

    Using one specific SKU

    Guess 4% → data <1%

    Neither guess was close, one off by half and the other by more than four times. And these were experienced people talking about their own customers.

    BI’s strength

    Known questions, answered reliably

    Its blind spot

    Everything nobody thought to ask

    The Foundation

    Why Is the Data Foundation Harder than the AI Agent?

    Martin Casado makes a point worth repeating: building an AI analyst is less a product story and more a “do the work” story. Agents are good at what human analysts used to do. The chat window is not the hard part.

    The hard part is everything underneath: clean, connected data, consistent definitions, documented business rules and secure access. An agent working on top of five disconnected systems will give you five confident, different answers. Faster than before, but not better. That foundation is what a data platform is for.

    The Team

    Will AI Analytics Shrink Your Data Team?

    The obvious fear is that AI replaces the data team. The people closest to it see the opposite. Tristan Handy said that over the past year exactly one data leader told him they thought they could lay off most of their data team. Every other data leader called it a boon.

    Casado, who sits on boards with visibility into hundreds of companies, describes it as a productivity story, not a headcount story. Teams that adopt AI tend to grow, because the agent keeps surfacing new problems worth solving. Someone still has to own the data, the definitions and the decisions.

    Getting Started

    How Do You Open Your Data to AI in Five Steps?

    1

    Find the queue. Which questions wait longest for an answer, and who is asking them?

    2

    Build one source of truth. Connect the systems behind those questions into one data foundation. Without it, the agent multiplies inconsistencies.

    3

    Write down the tribal knowledge. Metric definitions, business rules, known data quirks. This is the context an agent needs and a human analyst never wrote down.

    4

    Open the data securely. Give Claude, ChatGPT or Copilot governed access to the warehouse, not to raw exports in someone’s inbox.

    5

    Start with one team and measure. Count the questions before and after. If the Ramp pattern holds, the number will surprise you.

    From Reports to Answers

    What Should You Take Away?

    BI gave companies a reliable view of the questions they already had. AI analytics gives them a way to ask the rest. But an agent is only as good as the data it stands on. Get the foundation right first, and the questions nobody asked start getting answered.

    To see how this works on real retail data, read the Altisport case study or see how we set up Talk to Your Data.

    FAQ

    Frequently Asked Questions

    What is AI analytics?

    AI analytics means asking questions about your business in plain language and getting answers computed from your own data. An AI model such as Claude, ChatGPT or Copilot turns the question into queries, runs them on your data warehouse and explains the result. Unlike a BI dashboard, it is not limited to questions someone prepared in advance.

    Why do data questions increase after a company launches an AI analytics agent?

    Because of latent demand. People always had more questions than the data team could answer, so they lived without the answers. When Ramp opened its research agent to everyone in Slack, data questions grew 20 to 30 times within weeks and about 50 times within six months.

    Will an AI data analyst replace the data team?

    Not in the companies that adopted it so far. Teams that use AI agents tend to grow, because the agent surfaces more problems worth solving. Someone still has to own the data, the metric definitions and the decisions.

    What does an AI analytics agent need to give correct answers?

    Clean, connected data in one place, written definitions of your metrics and business rules, and secure, governed access to the data warehouse. Without that context the agent guesses, and on top of disconnected systems it gives confident but different answers.

    How do you start with AI analytics?

    Find the questions that wait longest for an answer, connect the systems behind them into one source of truth, write down your metric definitions, give an AI model governed access to the warehouse, and start with one team. Count the questions before and after.

    If your data team has a queue, we are worth a conversation.