Data Strategy Before AI Strategy

The Demo Was Great. Production Wasn't.

The pattern repeats in almost every organization. The board asks for AI. A pilot launches, the demo dazzles, the deck goes up the chain — and then production arrives. The model that answered beautifully in the sandbox starts contradicting the ERP. The customer-churn predictor turns out to have been trained on a customer list that three departments maintain differently. Legal asks where the training data came from, and the honest answer is a shrug.

The pilot didn't fail. It succeeded at revealing something the organization didn't want to know: the data was never ready, and no model — however good — can outperform its inputs.

Buying Intelligence for Data You Don't Trust

The instinct is to treat this as an AI problem — pick a better model, a better vendor, a better prompt. But the constraint was never model quality; frontier models are a commodity anyone can rent. The constraint is that AI is an amplifier. Point it at governed, consistent, well-owned data and it amplifies insight. Point it at the spreadsheet economy — five versions of "customer," metrics that change meaning between departments, exports nobody can trace — and it amplifies confusion, now with confidence and at scale.

An AI strategy without a data strategy is a plan to automate your disagreements.

The Four Foundations — Sliced Thin

Here's where most organizations overcorrect: they conclude they need a three-year enterprise data program before any AI can start. That's the opposite error — a program that fixes everything fixes nothing first. The right move is to pick the two or three AI use cases that matter, and build only the data foundations those use cases stand on:

• 1. Ownership. Every dataset feeding an AI use case gets a named business owner — not a platform team, but the executive whose capability produces the data. If nobody owns "customer," no model that consumes it has an accountable answer for being wrong. This is the same capability-based ownership that anchors architecture and cost visibility; data is its third application.
• 2. One definition for the entities that matter. Not a company-wide semantic layer — a reconciled definition of the three to five entities your chosen use cases depend on: this customer, this product, this order. Agree where each lives, which system is authoritative, and retire the competing copies as they're touched.
• 3. Quality measured at the source. Data quality enforced by downstream cleanup is a treadmill. Instead, define what "good" means for each feeding dataset — completeness, freshness, validity — and measure it where the data is produced, visibly, on the owner's scorecard. Models then inherit quality instead of laundering its absence.
• 4. Lineage and access you can defend. Before a model consumes data, you should be able to answer two questions in one sitting: where did this come from, and who is allowed to see what it produces? That's not bureaucracy — it's the difference between an AI answer you can act on and one you have to re-verify by hand, which is to say, one that saved nothing.

Thin-sliced this way, the data work for a first serious AI use case is a quarter, not a multi-year program — and every slice compounds, because the second use case inherits the entities, owners, and pipelines the first one governed.

The Sequencing Rule

This isn't "data before AI" as a purity test. It's a sequencing rule: for each use case, the data foundations ship first, sized to that use case only. Organizations that invert it — model first, data later — pay twice: once for the pilot that can't be trusted, and again for the credibility that AI loses internally when its first production answer is confidently wrong. The fastest route to AI that works is a short detour through the data it eats.

What Good Looks Like

The end state is quiet: an AI answer is challenged in a meeting, and someone traces it to a governed source in minutes instead of convening a task force. Each new use case starts with a data-readiness check that takes days, not a discovery project. And the AI roadmap stops being a list of demos and becomes a list of decisions the organization now makes faster — because the data underneath them finally agrees with itself.

Data foundations are exactly half of what our free AI Readiness Assessment measures — twelve questions across data ownership, quality, use-case discipline, and operating readiness, with specific next steps for your band.

John

Founder, Splendor Technologies

20+ years in AI, enterprise architecture, and application development. Helping organizations modernize technology and drive measurable business outcomes.