I remember sitting there and realizing I wasn’t thinking about the model at all. I was thinking about the data feeding it.
One discussion stands out in particular. We were evaluating how predictive analytics could improve sales forecasting for a national portfolio of opportunities. Leadership wanted greater confidence in projected outcomes so resources could be prioritized earlier in the sales cycle. As conversations turned toward model accuracy, we discovered something more important. Different teams weren’t consistently recording opportunity stages, probability scores and client attributes. The model wasn’t struggling because it lacked sophistication. It was learning from business processes that had never been standardized in the first place. That meeting changed how I approached every AI initiative that followed.
Throughout my career leading enterprise business intelligence initiatives, I’ve repeatedly watched organizations blame the algorithm when the real issue was inconsistent data, fragmented ownership across departments and business definitions that meant different things to different teams. AI doesn’t distinguish between disciplined and inconsistent business processes. It learns from both with equal confidence.
I’d built and defended executive dashboards for years before that meeting, and dashboards had trained me to believe imperfect data was a manageable, even routine problem. Experienced leaders read a dashboard with context. They know which numbers to trust, which ones need a caveat and which gaps to mentally fill in based on what they already know about the business. Predictive AI doesn’t have that judgment. Machine learning assumes the historical data it’s trained on represents reality as it actually is. If two departments define “active customer” differently, or if a critical field has been silently incomplete for two fiscal years, the model doesn’t notice or compensate. It learns the inconsistency as ground truth, and it repeats that mistake at scale, with confidence, every single time it runs.
That moment fundamentally changed how I approach every AI initiative. I stopped starting with technology and started with data integrity instead.
Today, I rarely begin AI discussions by talking about technology. Before any conversation about platforms or vendors, I ask five questions of the leadership team. Can we explain, in plain language, where this data actually comes from? Do the business leaders in the room agree on what our core definitions mean, or does “revenue” or “active account” shift depending on who’s presenting? Would we rely on this data to make a multimillion-dollar decision without a human manually double-checking it first? Is there a specific, named person accountable for every critical dataset, or does ownership dissolve the moment something goes wrong? And underneath all of it, are we actually solving a business problem, or are we chasing a technology because it’s the thing everyone else is talking about this quarter?
I remember one initiative where these questions prevented us from moving too quickly. During an early assessment, we discovered that two operational systems treated the same customer differently because each had evolved around separate business processes. Executive reports appeared consistent because manual reconciliation had become part of the monthly reporting routine. Once we identified the inconsistency, the project d while business stakeholders agreed on common definitions and ownership. That decision delayed the AI initiative by only a few weeks, but it likely prevented months of troubleshooting after deployment. More importantly, it strengthened confidence in every analytics initiative that followed.
These conversations reveal far more about whether an organization is genuinely ready for AI than any vendor demonstration ever will. A polished proof-of-concept can make almost any dataset look production-ready for the ten minutes it’s on screen. These five questions don’t have that luxury. They tend to surface, quickly and uncomfortably, where an organization’s data confidence actually breaks down, and that’s the information leadership needs before committing budget and reputation to a rollout.
This lines up with what the NIST AI Risk Management Framework has argued for a while now: governance and accountability belong at the foundation of an AI initiative, not layered in after a model is already in production. Governance built in retroactively tends to be theater, built to explain a failure that’s already happened rather than to prevent one.
The organizations I’ve seen actually succeed with AI invest first in governance, ownership and shared business definitions, and only then in the platform itself. They clean up master data before they scale a model against it. They remove duplication in customer and product records. They assign accountability for datasets the same way they’d assign accountability for a budget line, with a name attached and consequences if it slips. This work rarely shows up in a demo, which is probably why it gets skipped so often in the rush toward deployment.
One lesson I’ve seen repeatedly is that assigning ownership changes behavior almost immediately. Once business leaders understood they were accountable for the quality of specific datasets, not just the reports generated from them, conversations shifted. Instead of asking why dashboards looked different, teams began discussing why the underlying business process produced inconsistent information. Governance stopped being viewed as documentation and became part of everyday decision-making. The improvements weren’t dramatic overnight, but they were sustainable, and that consistency ultimately mattered more than any individual technology upgrade.
I saw this firsthand during an executive reporting initiative where multiple leadership teams relied on the same performance dashboard but interpreted one KPI differently, because ownership had never been clearly assigned. Once the business designated a single owner for the metric and standardized its definition across reporting systems, disagreements disappeared almost overnight. More importantly, that same governance work later allowed predictive analytics to be introduced with confidence, because everyone was working from the same version of the truth.
I’ve learned that AI projects rarely fail in the data science team. They fail months earlier, when leadership assumes the organization already understands its own data.
McKinsey’s research on scaling AI reinforces this pattern at scale: the organizations that generate lasting value from AI are consistently the ones that pair the technology with real changes to their operating model and governance, rather than simply layering AI on top of how things already worked. That finding matches what I’ve observed leading enterprise analytics initiatives directly. The technology was rarely the constraint. The organization’s relationship with its own data was.
It’s tempting to frame AI adoption as an engineering problem with a leadership footnote, when in practice it’s closer to the reverse. CIOs reporting on rebuilding an AI-ready data strategy make a related point: treating data ownership as a purely IT issue stops working once business units, product teams and AI platforms are all generating and transforming data continuously, which is exactly why accountability has to sit with named business leaders, not a technical team working in isolation. A related piece on building an AI-ready data culture puts it more bluntly: an organization can’t scale AI without first scaling trust in its own data, and that trust starts with culture and ownership, not tooling.
I no longer ask whether an organization is AI-ready. I ask whether its leaders would bet on their own data without a human checking behind the model first. If the honest answer is no, the next investment shouldn’t be another AI platform or a more sophisticated model. It should be a stronger data foundation, built deliberately, with clear ownership, before a single additional AI use case gets greenlit.
If another executive asked me for one piece of advice before approving a major AI investment, I’d tell them this: spend one day interrogating your data before spending another dollar on your model. What that conversation reveals will tell you more about your organization’s readiness than any vendor demonstration ever could. Organizations rarely fail because their AI isn’t intelligent enough. They struggle because they ask AI to learn from data that was never prepared to support intelligent decisions in the first place. The organizations that lead in this next era won’t be the ones with the most advanced models. They’ll be the ones that got their own house in order first, and knew it.
**This article is published as part of the Foundry Expert Contributor Network.**Want to join?