In early 2023 it seemed that every executive wanted an AI strategy. Most of them were starting with the solution and working backward to the problem. I suggested taking a breath and doing the groundwork first. Most didn’t.
Some are now discovering what everyone who’s ever deployed enterprise software already knows: the demo is not the product. The demo is the holiday brochure. The product is the actual vacation, complete with lost luggage, food poisoning, and a hotel room that looks nothing like the photos.
Mind the Gap
The gap between what generative AI can do in a demo and what it can do in production is enormous because the conditions in the demo bear no resemblance to the conditions in a real enterprise. In the demo, the data is clean. In your organization, the data is a decade of mergers, migrations, and manual workarounds. In the demo, the model answers questions about a tidy, well-structured knowledge base. In your organization, the knowledge lives in 400 SharePoint sites, a retired Confluence instance nobody decommissioned, three different CRMs, and Dave’s head. Dave, by the way, is on sabbatical.
In the demo, the model is right. In production, it’s right often enough to be dangerous. I watched a professional services firm pilot a contract summarization tool that worked brilliantly on their test documents. Clean contracts, standard clauses, consistent formatting. Then they pointed it at actual client contracts. Amendments stapled to amendments. Handwritten margin notes scanned as images. Clauses that contradicted other clauses because nobody cleaned up the redlines from 2019. The model summarized with the same confidence it always has. It was just wrong more often, and nobody could tell without reading the original document anyway.
That’s the hallucination problem in a nutshell. It’s not that the model necessarily makes things up out of nowhere, although it can, but it presents everything with the same polished certainty, whether it’s right or spectacularly wrong. And in a business context, confidence without accuracy is worse than no answer at all.
About that Data Problem
I wrote in December that the unsexy work of data quality just became your most important AI initiative. Six months later, this has gone from prediction to lived experience for nearly every organization that started this journey.
It looks like this: Leadership greenlights an AI pilot. The team picks a promising use case. They connect the model to internal data. And then they spend three months discovering that the data is a mess. Not because nobody cared about data governance before, but because the bar just changed. Data that was good enough for reporting and dashboards is not good enough for a model that’s going to generate customer-facing answers from it.
One social services client wanted to use AI to help potential beneficiaries find relevant services to enroll in, based on their circumstances. Reasonable use case. High value. The pilot stalled because the protocol documents existed in multiple versions across different systems, with no reliable way to determine which was current. The AI wasn’t the bottleneck. The twenty years of document management decisions that preceded it were.
This is the part that doesn’t make it into the vendor pitch. The model is the easy part. The data underneath it is the hard part. And most organizations have been underinvesting in data quality for years because there was never a forcing function strong enough to justify the expense. Now there is.
The Governance Vacuum
While organizations are figuring out data quality, they’re simultaneously discovering that they have no governance framework for AI. And in their defense, up until six months ago, there was nothing to govern. Fast forward, employees are using ChatGPT and a half-dozen other tools with company data. Every day. Without policies, without oversight, and mostly without malicious intent. They’re summarizing meeting notes, drafting communications, analyzing spreadsheets. They’re being productive. They’re also feeding proprietary information into systems the organization doesn’t control.
This is one of those cases where a little goes a long way. It’s easy to kill innovation with overly comprehensive AI policies. So start with a simple outline. One: what data can touch external AI services, and what can’t. Two: what decisions can be informed by AI output, and what requires human judgment. Everything else is refinement. The traditional approach is to form committees to study the issue. By the time the committee reports, the workforce will have established patterns that are much harder to change than they would be next week.
Who Owns AI?
Few organizations are asking yet where AI fits into the existing org chart. It’s not an IT capability in the usual sense. It’s not truly a data team function, though the data team is critical. It’s not completely a business unit responsibility, though the business units are where the use cases live. It’s more of a horizontal capability that cuts across everything, and organizations are terrible at horizontal capabilities.
There’s an interesting jurisdictional dance playing out in slow motion. The CTO thinks it’s a technology decision. The CDO thinks it’s a data decision. The business unit leaders think it’s their decision because they own the use cases. Nobody’s wrong. Everyone’s incomplete. And while they negotiate, the shadow AI usage grows.
What We Learned
Organizations making real progress share a common trait: they started small and specific. Not “implement AI” but “use AI to reduce the time our analysts spend on initial contract review from four hours to one.” Bounded scope. Measurable outcome. A use case boring enough that nobody puts it on a conference keynote, but valuable enough that the business actually cares if it works. Perhaps we shouldn’t even talk about use cases, only value cases.
They also share an uncomfortable realization: the AI initiative surfaced problems that existed long before AI showed up. AI illuminated and amplified them, which, painful as it is, is a useful realization all by itself.
