Data Readiness: The Make-or-Break Factor for AI Success

The Dirty Secret of AI Projects

In our post-mortems of AI projects that didn't achieve their goals, the cause is overwhelmingly the same: data that was assumed to be clean, structured, and representative — but wasn't.

Models are commodities now. GPT-4, Claude, Gemini — these are table stakes. Your competitive advantage is your data. The question is whether it's ready.

The Data Readiness Framework

1. Completeness — Do you have enough labelled examples? For supervised learning: typically 10,000+ for production-grade models. For fine-tuning LLMs: as few as 200-500 high-quality examples, but they need to be very good.

2. Accuracy — Are your labels correct? Noisy labels are worse than fewer labels. Budget for a label quality audit before training begins.

3. Representativeness — Does your training data reflect the real distribution of inputs the model will see in production? A model trained on weekday data may fail on weekends. A model trained on normal operations will struggle with edge cases.

4. Freshness — How old is the data? In fast-moving domains (financial markets, social media, supply chains), a model trained on 2-year-old data may be systematically wrong about current patterns.

5. Accessibility — Is the data actually accessible to your AI team? Siloed in legacy systems, gated by privacy compliance, or locked in unstructured formats? This is often the longest-lead constraint.

The 6-Week Data Audit

Before any AI project kicks off, we run a 6-week data audit:

• Week 1–2: Data inventory and access mapping

• Week 3–4: Quality assessment and gap analysis

• Week 5–6: Data preparation roadmap and remediation prioritisation

This isn't overhead. It's the single highest-ROI investment you can make before writing any model code.

The Rule of Thumb

If you're less than 70% confident in your data quality, stop. Fix the data. The model can wait.

Good data + simple model = good outcomes.

Bad data + sophisticated model = expensive failure.