There's a persistent belief that better models fix bad data. They don't. GPT-4 processing garbage produces eloquent garbage. The most expensive and least visible cost in enterprise AI isn't inference or infrastructure. It's the downstream impact of poor-quality training and reference data.
What You Need to Know
- Data quality is the single biggest predictor of AI project success. Not model choice, not compute budget, not team size. If your data is inconsistent, incomplete, or outdated, no model will save you.
- "Garbage in, garbage out" understates the problem. With AI, garbage in produces confident, plausible-sounding garbage out. It's worse than no output because people trust it.
- Most organisations underestimate their data quality issues until they try to use the data for AI. Legacy systems, inconsistent formatting, duplicate records, and undocumented schemas are the norm.
- Data quality investment has a compounding return. Clean data for one AI capability makes every subsequent capability cheaper and faster to build.
73%
of enterprise AI project delays are caused by data preparation and quality issues
Source: Gartner, Data Quality for AI Survey, 2023
The Confidence Problem
Traditional software breaks visibly when data is bad. A null value throws an error. A missing field fails validation. The system tells you something is wrong.
AI is different. Feed a language model poorly structured data, and it still produces output. Fluent, confident output. The answer might be wrong, or partially wrong, or correct but based on the wrong source, but it reads well. It sounds authoritative.
This is the hidden cost: not that AI fails on bad data, but that it fails in ways that are hard to detect. A claims processing AI that extracts the wrong policy number from a poorly formatted document doesn't throw an error. It returns a confident extraction of a number that happens to be from the wrong section.
We've seen this pattern across multiple enterprise deployments. The AI performs brilliantly in testing (clean test data), adequately in staging (curated sample data), and poorly in production (real-world messy data). The gap between test and production is almost always data quality.
Five Data Quality Failures That Kill AI Projects
1. Inconsistent Formatting
The same information stored differently across systems. Dates as DD/MM/YYYY in one system, MM/DD/YYYY in another. Names as "Last, First" vs "First Last." Addresses with and without unit numbers. Currency with and without symbols.
For humans, these are minor annoyances. For AI, they're confusion signals that degrade extraction and matching accuracy.
2. Stale Reference Data
The knowledge base hasn't been updated in 18 months. The policy documents are three versions old. The product catalogue has discontinued items. The org chart reflects last year's structure.
RAG systems are only as good as the documents they retrieve. If your best match for a query is an outdated policy, the AI will confidently cite outdated information.
3. Duplicate and Conflicting Records
Customer records duplicated across CRM, billing, and support systems. When the records conflict (different email, different address, different account status), which one does the AI trust?
Without explicit deduplication and source-of-truth designation, the AI picks whichever record the retrieval system surfaces first. That's not a strategy. It's a lottery.
4. Missing Context
Data without metadata. Documents without dates, authors, or version numbers. Records without creation timestamps or modification history. The AI can't assess recency or authority because the data doesn't carry that information.
5. Implicit Knowledge
"Everyone knows" that column X in the legacy system actually means Y when condition Z is true. This tribal knowledge is never documented. The AI doesn't know it, and nobody thinks to tell it.
40%
of data scientists' time is spent on data cleaning and preparation, not model development
Source: Anaconda, State of Data Science Report, 2023
What Good Looks Like
Organisations that get data quality right for AI follow a consistent pattern:
Audit before you build. Before starting any AI project, run a data quality assessment on the specific data sources the AI will use. Not a general data audit. A targeted assessment of completeness, consistency, recency, and format.
Define quality metrics. Completeness (what percentage of records have all required fields), accuracy (what percentage of values are correct), consistency (do related records agree), and timeliness (is the data current).
Clean incrementally. You don't need perfect data to start. You need good-enough data for your first use case, with a plan to improve. Each AI capability you build reveals new data quality issues, and fixing them benefits every subsequent capability.
Build quality into pipelines. Don't clean data once. Build validation, normalisation, and quality checks into your data pipelines so quality is maintained automatically.
The model is a commodity - your data is the differentiator. The uncomfortable truth is that most organisations' data isn't as clean as they think it is, and the AI will show you exactly where.
Dr Tania Wolfgramm
Chief Research Officer
Every data quality issue you fix for one AI capability benefits every future capability - that's the compounding return. The first AI project is expensive because you're paying the data quality debt; the fifth one is dramatically cheaper.
John Li
Chief Technology Officer