Every publisher and media company is asking the same question right now: How do we make AI work for our business? But in the rush to adopt, there’s a foundational question most teams aren’t asking loudly enough: Is your data actually ready for it?
AI doesn’t conjure insights from nothing. It works on what you give it. And if you give it a mess of inconsistent naming conventions, ungoverned exports, siloed systems and unreconciled discrepancies, AI will hand you something that looks authoritative but is, at its core, unreliable.
To be fair, AI can sometimes work with imperfect data, particularly on the buy side. Doing so is expected when it comes to modeling built on top of incomplete measurement. The model accounts for any gaps, and no one really demands absolute precision. Analysts have always estimated incomplete measurement data, and LLMs can do the same.
But revenue data is different and presents a unique challenge for publishers. When a number goes to finance or a client, being even 1% off is not a rounding issue. And you can expect an uncomfortable call asking why the invoice does not match the report.
Here’s what publishers should keep in mind to ensure their data infrastructure is ready for AI.
Auditability and explainability
There is a version of AI adoption that creates two parallel worlds inside your organization: the world where the AI does interesting things with data and the world where the finance team actually operates.
If those two worlds don’t connect, as is often the case with vibe-coded or quickly assembled AI implementations, the technology ends up creating additional work instead of eliminating it.
Finance, legal and billing functions require auditability. They need to know exactly where a number came from, what logic produced it and whether it will produce the same result if re-run tomorrow. An AI model cannot guarantee that. When billing runs at month-end, your team will still need to normalize, validate and reconcile data from scratch.
The deeper risk is fragmentation. Without a single source of truth, billing data, performance analytics data and client-facing reporting can all come from different pipelines with different logic. Every team is right in its own system, and no one agrees at the executive level. That is not a new problem that AI created, but rather an underlying data architecture problem that AI makes visible and urgent.
AI still has to pass procurement
Enterprise organizations have formal requirements for any software that touches business-critical data. Certifications and standards compliance, documented data handling processes, access controls, retention policies and clear vendor accountability are table stakes.
A vibe-coded internal tool or ad hoc pipeline assembled with a few AI prompts and a CSV export will not pass that review. If the output of that pipeline is used to produce numbers that go into billing, reporting or client deliverables, the compliance team’s answer will be “no,” and it will stay that way until the foundation is rebuilt properly.
They’re not Luddites. Compliance requirements exist for real reasons. AI adoption in enterprise contexts has to happen within those constraints, not around them.
AI can’t reconcile what you haven’t aligned
Publisher data lives across ad servers, OMSs, SSPs, verification vendors, audience platforms and bespoke internal tools, each with its own naming conventions, metrics definitions, time zones and attribution logic.
AI cannot reconcile those differences on its own. It will map what it can, make assumptions about the rest and produce output that looks unified but contains hidden discrepancies. An AI that smooths over a 10% variance between what your ad server reports and what your SSP invoices is making a business decision you likely did not intend to empower it to make.
The normalization and integration layer has to be built deliberately, with human oversight, before AI-powered analytics can be trusted. That work isn’t glamorous; it’s taxonomy alignment, field mapping, discrepancy thresholds and data contracts with your tech partners. But it’s the work that makes everything downstream reliable.
A practical checklist for AI adoption
Before investing time, resources or money in AI tooling for your data and analytics workflows, make sure the foundation is in place:
- Normalize and aggregate your data from all sources into a single, governed pipeline before routing it to any AI system
- Ensure your data is stored in a format and location that AI systems can access reliably, whether through an MCP server, a structured database or a well-documented API layer
- Establish data quality standards and implement automated checks upstream, so the AI never sees bad data in the first place
- Work with your BI, finance and legal teams to define what “audit-ready” means for your organization and build to that standard from day one
- Document your data logic explicitly, not just for compliance, but also to ensure that any AI-generated analysis can be traced back to a verifiable methodology
The promise of AI in advertising and media is real. AI is incredibly valuable, and we should collectively embrace its potential. The companies that will realize it aren’t the ones who moved fastest; they’re the ones who built on solid data foundations.
AI amplifies what’s underneath it. Make sure what’s underneath it is worth amplifying.
“The Sell Sider” is a column written by the sell side of the digital media community.
Follow Burt Intelligence and AdExchanger on LinkedIn.
For more articles featuring Shay Brog, click here.
