Data Quality for AI: Stop Blaming the LLM
    Data GovernanceSingle Source of TruthAnalytics EngineeringLLM

    Data Quality for AI: Stop Blaming the LLM

    Your AI search project is failing because of poor data quality, not the LLM. A CTO's guide to fixing the foundation before building the skyscraper.

    Executive Summary

    Pain

    You're a CTO at a growing online marketplace. The board is keen to use AI for search and analytics, but your first attempts are giving strange or conflicting answers. This is usually because the underlying data isn't quite consistent.

    Risk

    Putting more money into LLM tools before sorting out the data is just a way to get wrong answers, faster. Each time an AI project doesn't work out, it can reduce the business's confidence in the data team and hold back real progress.

    Fix

    The answer isn't a better LLM or more data scientists. It's usually a sign that the data's structure needs attention. You first need to build a governed Single Source of Truth before you can get any real intelligence from it.


    Why an LLM can't tidy up your data

    There's a tempting idea doing the rounds in boardrooms: that you can buy a clever Large Language Model, point it at your data warehouse, and it will sort out years of accumulated untidiness. The sales pitch is appealing, promising instant answers from plain English questions, automated insights, and a lead on your competitors.

    In my experience, this rarely works. It’s a bit like hiring a world-class translator to make sense of a meeting where everyone is shouting in a different language. The translator is not the problem, the chaos is.

    An LLM will amplify existing data problems

    An LLM is a powerful tool for finding patterns and generating language. It doesn't, however, have your business context or the ability to guess which of your three conflicting definitions of "Active User" is the right one. When you give an LLM ambiguous or contradictory data, it won't fix it. It will simply amplify the confusion, but with great confidence and clarity.

    You have probably moved to the cloud and hired good engineers. But often, all that's happened is that the same mess has been moved to a faster system. Automating a broken process just means you produce bad data more quickly. The problem isn't that your warehouse is slow. It's that your Metric Definition for a core concept like Gross Merchandise Value (GMV) hasn't been agreed on and exists in ten different SQL scripts.

    I see this pattern quite a lot in scale-ups, especially in businesses with lots of transactions, like online marketplaces. They've invested in the Modern Data Stack: Snowflake, dbt, Fivetran, Looker. But they've often under-invested in the less glamorous, but essential, work of agreeing on what their data actually means. This is a common problem in Marketplace Analytics; you can't balance supply and demand if you can't define them consistently.

    Whiteboard

    How to build a reliable data foundation

    My advice is usually to stop looking for better AI tools for a moment. The work that's easiest to put off is often where the real value is. The fix is to treat your data as a product, which needs some discipline and structure, not just more technology.

    This is often a Data Governance problem that looks like a technology problem. The solution is to build a Semantic Layer: a central, governed place for your business logic. This is where all your core metrics, dimensions, and relationships are defined in code, just once.

    This isn't really a task for a data scientist; it's the work of Analytics Engineering. It involves a few steps:

  1. Metric Consolidation: Getting the commercial and technical teams to sit down and agree on one single definition for every important KPI. This is often more about people than it is about code.
  2. Logic Centralisation: Taking the business logic out of individual dashboards and reports and putting it into a version-controlled dbt or LookML model.
  3. Ownership and Documentation: Making it clear who owns which data, and creating a place where anyone can find out what a term means.
  4. Only when this foundation is in place can an LLM do its job properly. When the model asks, "What was our user churn last month?", it gets one clear answer because the logic is already defined and governed. It's no longer guessing; it's asking a trusted source.

    The outcome of getting the foundation right

    By focusing on the data structure first, the whole situation changes. You move from reactively explaining why the AI's answers are off, to working on strategy. Your data team stops being a helpdesk for broken reports and can start helping with the direction of the business. The business gets what it always wanted: not AI for its own sake, but reliable answers to its most important questions. You stop spending money on trials that don't go anywhere and start building a reliable data setup that gives you a real edge.

    Ready to Transform Your Data?

    Book your free clarity call today and discover how NorthStar Analytics can help you build a single source of truth.