Demos never fail. Ask a chatbot ten carefully curated questions, and it answers with the polish of someone who rehearsed all night. But point that exact same system at 3,000 customer records scattered across a ten-year-old CRM, an orphaned spreadsheet, and a forgotten 2019 support queue, and watch the wheels come off. The confident tone survives, but the facts start to drift.
That is exactly when reality hits, and a company has to call for backup. What starts as a hunt for AI development companies to come in and “fix the model” quickly turns into a search for someone who can untangle the messy data underneath it. For executives, this data gap usually registers first as a minor nuisance. Then it becomes a glaring pattern. Finally, it becomes the exact reason their expensive pilot never makes it out of the sandbox.
What gets labeled an AI problem is, more often, an information problem wearing a costume. In almost every kickoff call, a firm that builds machine learning systems for a living says the same thing: the model is rarely the hardest part anymore. Pull a thread on any stalled pilot, and it leads back to data nobody structured, labeled, or agreed on how to define across departments. Marketing calls a customer a “lead.” Sales prefers “opportunity.” To finance, the same person is simply a line on a contract. None of the three definitions line up in the database, and no algorithm, however capable, can reconcile them on its own.
When the Model Sounds Right and Isn’t
Confidence is not the same thing as correctness, and generative systems have never learned the difference on their own. A support bot trained on a company’s public help pages might invent a return policy that never existed, not out of malice, but because nothing told it the policy changed eighteen months ago. Multiply that by every stale spreadsheet, every duplicate customer record, every product description three different teams edited three different ways, and the hallucination stops looking like a model flaw. It starts looking like an average.
By the middle of 2025, the pattern had a number attached to it. A widely discussed study from MIT’s NANDA initiative, based on interviews with company leaders and an analysis of hundreds of enterprise AI programs, found that 95% of generative AI pilots produced no measurable financial return. Not a small return. None. The businesses spending the most on pilots were rarely short on talent; some had hired the best AI development companies money could buy. What most were missing sat one layer down, in the pipes and tables nobody photographs for the pitch deck.
None of this shows up in a sales demo. It shows up three months later, when a support agent trusts the system’s confident answer over their own memory, and a customer gets a stale promise that legal now has to unwind.
What “Ready” Actually Means
Ask a data engineer what “AI-ready” means, and the vague word tends to get replaced fast with something specific. Gartner put a number on the gap in February 2025. A survey of more than 1,000 data leaders found that 63% of organizations either lacked the right data management practices for AI or weren’t sure they had them, and the firm now expects 6 in 10 AI projects built on unprepared data to be abandoned by the end of 2026.
A separate survey of chief data officers backs this up from a different angle. Data quality and readiness ranked as the single biggest obstacle to enterprise AI, named by 43% of respondents, ahead of budget, talent, or leadership buy-in. Only a small share of the same group described their data as clean and accessible enough for serious AI work.
Put those two findings side by side, and a picture starts to form. Data leaders tend to describe an AI-ready foundation with the same handful of traits, no matter the industry:
- Data pulled into one place instead of trapped across a dozen disconnected systems, so a model isn’t left guessing which version of the truth to trust
- Clear ownership of each dataset, someone accountable when a field is wrong, not just a ticket queue nobody closes
- Consistent definitions across departments, the “lead versus opportunity versus contract” problem solved once and written down
- Pipelines that catch bad data automatically, before it reaches a model, rather than after a customer notices
- A record of where data came from and who touched it along the way, which regulators increasingly expect and models quietly depend on
Building the Foundation Before the Model
Fixing this rarely starts with buying a bigger model. It starts with an audit nobody wants to pay for: a few weeks spent mapping where data actually lives, who owns it, and how far it sits from something a machine learning system could trust. Companies that skip this step tend to discover the gap the expensive way, mid-pilot, when the demo that worked in a controlled sandbox falls apart against production data.
A well-built data lake gives an organization one place to land structured and unstructured information without forcing everything into a rigid schema too early. Governance structures decide who can touch what, and under what conditions, before any model gets trained on it. Together, the two turn a pile of files into raw material that a data science team can work with. Many AI development companies now open every engagement with a data audit rather than a model-selection call, because skipping straight to the algorithm on unprepared data tends to fail, expensively and predictably.
N-iX, a software engineering firm that has spent much of the past two years expanding its data and AI practice, describes the same pattern from the delivery side: projects that start with a data strategy phase move faster later, and projects that skip it circle back to data work anyway, just later and pricier. That order rarely reverses.
None of this makes data engineering more exciting than the model sitting on top of it. It makes the model possible. A business that treats the data phase as a prerequisite, not an afterthought, reaches production in months, not years, with a system that keeps working once the demo ends.
Summing Up
The lesson underneath all of this is not that generative AI fails. It’s that generative AI reveals, faster and more visibly than any earlier technology, exactly how organized a company’s data was to begin with. A model trained on chaos answers with chaos, dressed in complete sentences. Before the next pilot gets funded, the more useful question is not which model to buy, but whether the data underneath it could survive contact with production. Get that answer honestly, and the next pilot has a real shot at becoming something more than a slide in next year’s deck.
Connie Cardillonero has opinions about investment trends in commerce. Informed ones, backed by real experience — but opinions nonetheless, and they doesn't try to disguise them as neutral observation. They thinks a lot of what gets written about Investment Trends in Commerce, Strategies for Profitability, E-Commerce Finance Insights is either too cautious to be useful or too confident to be credible, and they's work tends to sit deliberately in the space between those two failure modes.
Reading Connie's pieces, you get the sense of someone who has thought about this stuff seriously and arrived at actual conclusions — not just collected a range of perspectives and declined to pick one. That can be uncomfortable when they lands on something you disagree with. It's also why the writing is worth engaging with. Connie isn't interested in telling people what they want to hear. They is interested in telling them what they actually thinks, with enough reasoning behind it that you can push back if you want to. That kind of intellectual honesty is rarer than it should be.
What Connie is best at is the moment when a familiar topic reveals something unexpected — when the conventional wisdom turns out to be slightly off, or when a small shift in framing changes everything. They finds those moments consistently, which is why they's work tends to generate real discussion rather than just passive agreement.
