Data as Business Asset — Why Your Data Matters More Than the Model

Every competitor in your industry has access to the same foundation models. OpenAI, Anthropic, Google, and Meta's models are available to any organisation at comparable pricing. Two companies deploying the same model on the same task produce similar outputs — unless one has proprietary data that makes its model version better on that task. The strategic question for every AI investment: what proprietary data does this initiative create, capture, or improve — and can competitors replicate it by signing a different vendor contract?

Volume matters — but it is the minimum necessary condition, not sufficient. More training examples generally improve model performance, but the relationship is logarithmic: doubling data volume after a certain threshold produces diminishing returns. A model trained on 1 million examples of a specific task typically performs similarly to one trained on 10 million — if the additional 9 million examples are of similar quality and variety. Volume alone is not an AI data strategy. Assess all five dimensions before concluding your data position is strong or weak.

A data audit has four components: inventory, quality assessment, gap analysis, and liability review. Inventory: what data exists, where it lives, in what format, and who owns it. Quality assessment: what is the accuracy, completeness, and consistency of each dataset? Gap analysis: what data would the target AI models require that is currently unavailable? Liability review: what personal data, sensitive data, or legally constrained data exists in potential training sets? Commission a data audit as the first deliverable in any AI initiative with a budget above your materiality threshold. The audit cost is typically 5–10% of total project cost and reduces failure risk by substantially more.

Data quality problems are primarily process failures, not technology failures. Duplicate customer records are created by onboarding processes that do not check for existing accounts. Incomplete product data results from catalogue management processes that allow partial entries. Inconsistent employee records reflect HR systems that accept non-standard job title inputs without validation. Map data quality problems to the business processes that create them. Fix the process first; invest in tooling to maintain the fixed process.

The flywheel loop has four stages: usage, data capture, model improvement, and better product. Users interact with the product → their interactions generate training signal (clicks, corrections, outcomes, preferences) → the signal retrains the model → the improved model makes the product better → the better product attracts more users. The flywheel design question is: what signal from user interactions is captured and fed back into model improvement? If the answer is 'none', you are not building a flywheel — you are building a static tool that competitors can match.

Three ownership questions must be resolved in every AI vendor engagement. First: who owns the training data provided to the vendor for fine-tuning or model customisation? The answer should be: your organisation. Second: can the vendor use your interaction data to improve their model for other clients? The answer should be: no, without explicit written consent. Third: if you terminate, can you export your interaction data and any model weights trained on your data? The answer should be: yes, with a defined export format and timeline. Add these three questions to your AI vendor onboarding checklist as non-negotiable minimum requirements. Vendors who refuse all three should be eliminated from consideration.

External data comes in three commercial forms: purchased datasets, licensed feeds, and data partnerships. Purchased datasets are one-time transactions: you buy historical data from a data broker or specialist provider. Useful for bootstrapping training sets but carries no ongoing freshness. Licensed feeds provide continuous updates at recurring cost — appropriate for time-sensitive applications like market data, weather, or regulatory updates. Data partnerships are bilateral agreements where two organisations exchange data to mutual benefit — often the richest source of unique, non-commodity data. Match the data acquisition form to the freshness and exclusivity requirements of the target AI application.

A data strategy for AI has five components that business leaders must own. One: data inventory and governance — what do you have, where is it, and who is accountable for its quality? Two: data quality investment — the roadmap to get the data from its current state to AI-ready state. Three: flywheel design — how does your AI product design capture training signal from usage? Four: data rights management — vendor contracts, partner agreements, and ownership frameworks. Five: data acquisition — what gaps require external purchase, licensing, or partnership? Use these five components as a data strategy assessment framework in your next AI strategy review. Score each component one to five — the lowest scores are the highest priority investments.