The question your board will ask about AI

Sooner or later, every board that has approved an AI program asks a version of the same question: how do we know it is right?

The first answers are usually about the model. Which one, how it was evaluated, what guardrails it has, what the hallucination rate was on the test set. Those answers matter. They also skip the question underneath, the one the audit committee will eventually put in writing: how do we know the data it is reasoning over is right?

For most large companies, that question leads straight to the warehouse. And the warehouse is usually in the middle of being rewritten.

Why does AI readiness start in the warehouse?

Because AI inherits the warehouse. An assistant that answers "what was net revenue in the northeast last quarter" does not compute net revenue. It retrieves a number that a chain of stored procedures, ETL packages and scheduled jobs produced. Those rules decide what counts as revenue, which returns are netted, which region a customer belongs to. Many of them were written years ago, few are documented, and they are trusted because they have always run.

An agent goes further: it may rerun a load, backfill a partition or act on a threshold. Now the warehouse logic is not just being read. It is being operated.

If that logic is being migrated to Databricks, Microsoft Fabric or Snowflake, which for most enterprises it is, the AI program is standing on code that is being rewritten. If nobody can prove the rewritten version produces the same numbers, the AI is confidently reporting figures nobody has verified, at a scale and speed no analyst could match.

How big is the hidden debt underneath AI programs?

Larger than most business cases assume. McKinsey's research found that CIOs estimate tech debt amounts to 20 to 40 percent of the value of their entire technology estate before depreciation, and that 60 percent of CIOs surveyed felt their organization's tech debt had risen over the previous three years. Gartner expects 80% of technical debt to be architectural by 2027 (Magic Quadrant for Technical Debt Management Tools, May 2026).

In a data estate, much of that debt lives in the logic. We call it logic debt: the thousands of procedures that must be moved, understood and proven before the numbers an agent reads can be called trustworthy. It is the line item almost no AI business case contains.

What will the board actually ask?

In practice, four questions, in roughly this order.

What does a good answer sound like?

That is a governance answer an audit committee can test. It is also, not by coincidence, a description of a migration done deterministically and proven object by object.

• We know what the logic is. Every object in the estate is inventoried, banded by complexity and mapped by lineage. Dead code has been retired. We can show where any reported number comes from.
• We can prove it moved intact. Every migrated object has a reconciliation report against the original on the same inputs: schema, row counts, column values and the checks our finance team supplied. The reports are in the audit file.
• We can see if it drifts. Production is compared against what we proved on a schedule, so a change that breaks equivalence surfaces in a report, not in a restatement.
• Agents work inside our controls. The pipelines agents call are exposed with the same access rules, approvals and logs as a person. Every action an agent takes is recorded and attributable.

What does a weak answer sound like?

"The migration team tested it." "The vendor's model is very accurate." "We have not seen any issues so far." "The dashboards reconcile at the summary level."

Each of these is true right up until the first restated number. None of them can be handed to an auditor. And the last one is the most dangerous, because summary-level agreement routinely hides object-level differences that cancel out, until one quarter they do not.

Where should an AI program start on data readiness?

The first proven domain is the first piece of evidence the board can hold. It is worth more than any slide about model accuracy.

1. List the AI program's first use cases and the tables they read. Usually a handful of domains: revenue, customer, claims, inventory.
2. Trace those tables back through the warehouse logic. That is the scope that matters most.
3. Prove one of those domains end to end. Convert it, reconcile every object, put it live, with a threshold written down in advance.
4. Expose its pipelines to agents under your controls. Then let the AI program build on evidence.
5. Move the rest of the estate the same way, in waves.