AI-ready data is modeled data
There's a comforting story making the rounds: you don't need to fix your data anymore, because the model will read it the way a smart human would. Chunk it, embed it, retrieve it — semantics included, free of charge. I've come to think of this as the plug-and-play AI fallacy, and it's quietly responsible for a large share of the AI project failure statistics.
Analytics-ready was a lower bar
Analytics-ready data had a merciful property: a human analyst sat between the data and the decision. Ambiguous column names, three definitions of "customer," a revenue figure that means different things in different marts — the analyst absorbed all of it, asked around, and compensated. Dashboards could survive semantic ambiguity because interpretation was still a human job.
An agent removes that buffer. When an LLM queries your data to act — not to render a chart someone will squint at, but to trigger a refund, draft the email, reorder the stock — every unresolved ambiguity flows straight through to the action. The four barriers that show up in every failed initiative are prosaic: inconsistent formats, ambiguous semantics, missing lineage, and unresolved entity identity. None of them are model problems. All of them are modeling problems.
Modeling is the act of deciding what things mean
The old three-layer discipline — conceptual, logical, physical — was never really about databases. It was a forcing function for questions that still have no other place to get answered: What are our entities? Which relationships are real? What does one row mean? Which key identifies a customer across systems? The classic traps — fan traps, chasm traps, under- and over-normalization — are all failures to decide meaning precisely. Embeddings don't answer these questions. They blur them. A vector is a lossy summary of whatever text happened to be nearby; it cannot tell an agent which of your three customer tables is authoritative, or that "margin" in the finance mart already includes freight.
This is why I hold the maybe-unfashionable position: the path to AI-ready runs through analytics-ready, not around it. Organizations that skipped semantic discipline and jumped straight to RAG built what the maturity literature calls the black-box lab — impressive prototypes on unaccountable data, stuck at the demo stage for reasons nobody can quite articulate. The reasons are articulable. The data never knew what it meant.
What AI-ready adds on top
AI-ready is analytics-ready plus machine-consumable semantics: definitions that live with the data rather than in a wiki; contracts that make schema change a negotiated event instead of a silent breakage; lineage an agent can cite when asked "why do you believe this?"; and entity resolution done in the platform, not re-derived per prompt. Do that work and something pleasant happens downstream — agent memory, RAG, semantic layers all get easier, because they inherit meaning instead of reconstructing it.
The uncomfortable summary for anyone hoping to skip ahead: if your data wasn't good enough for reliable analytics, it isn't good enough for autonomous action on your behalf. The model reads faster than your analysts. It does not read minds.