There's a phrase every data scientist has repeated so many times it's become a cliché: "garbage in, garbage out." But it's repeated constantly because it's true, and because clients keep discovering it the hard way — mid-project, after the AI vendor has already been paid for months of work.
The pattern is familiar. An organization decides to build an AI solution. They select a vendor, sign a contract, and proceed to discovery. Three weeks in, the AI team is asking uncomfortable questions: why does the CRM have 40,000 records where the "close date" field is empty? Why do customer IDs in the billing system not match customer IDs in the support system? Why are two years of transaction data stored as PDF exports rather than structured records?
These aren't unusual situations. They're the norm. Most organizations accumulate data over years without consistent standards, and the systems that were designed to capture data for operational purposes weren't designed for machine learning use cases. The result is that data preparation — cleaning, reconciling, standardizing, labeling — consumes between 60-80% of the time in a typical AI project.
The organizations that get AI deployed faster and cheaper are the ones that start fixing their data problems before they start the AI project.
What "Good Data" Actually Means for AI
Data quality requirements vary by use case, but there are universal dimensions that matter for almost every AI application:
Completeness: Are the fields your model needs actually populated? A churn prediction model needs historical customer behavior data. If 30% of your customer records have missing fields for key behavioral signals, your model trains on a biased subset and performs poorly on the full population.
Consistency: Is the same thing described the same way across records? "British Columbia," "BC," and "B.C." are the same thing to a human and different strings to a database. Inconsistent categorical values, date formats, and identifier schemes create noise that ML models learn from incorrectly.
Accuracy: Does the data reflect reality? This is harder to assess systematically but critical — a model trained on data where salespeople systematically overestimate deal sizes in CRM will learn to make systematically wrong predictions.
Timeliness: Is the data current enough to be relevant? A model trained on customer behavior data from 2022 may not capture how behavior has shifted. Stale training data produces models that are poorly calibrated to the current environment.
Volume: Is there enough data to train from? This is use-case specific. A simple classification model might train adequately on a few thousand labeled examples. A complex NLP model needs tens of thousands. If you have six months of transaction history but need two years for seasonal patterns to emerge, the model will miss cyclical behavior.
Labeling: For supervised learning, do your historical examples have the correct outcome labels? If you want to predict which leads will close, you need historical leads with known outcomes — and those outcomes need to be accurately recorded, not just "status = closed" which might have been set for administrative reasons rather than actual deal closure.
A Practical Data Assessment Before You Start
Before engaging an AI vendor, conduct an internal data assessment. This doesn't require data science expertise — a structured audit of your key data sources will reveal the biggest issues.
Step 1: Map your data sources. What systems capture the data relevant to your AI use case? For a customer churn model, this might be your CRM, billing system, support ticketing system, and product usage logs. List each source, what it captures, how far back it goes, and who owns it.
Step 2: Profile the key fields. For each data source, examine the fields your AI will likely need. Check: What percentage of records have this field populated? What's the distribution of values? Are there obvious data entry errors (dates in the future, negative quantities, free-text where categorical values are expected)? Most database tools (even basic SQL or Excel) can generate these statistics.
Step 3: Test your join keys. If your AI needs to combine data across systems, test whether the join keys actually work. Match a sample of records from system A to system B using your proposed join key. What percentage match? What percentage fail? A low match rate — even 80% — means 20% of your data is effectively unusable for cross-system analysis.
Step 4: Assess your labeling situation. If you're building a supervised model, examine your historical outcome data. Are outcomes recorded consistently? Is the historical data complete? Are there periods of missing data that would create gaps in training?
The Three Fixes That Matter Most
Not all data quality problems are worth fixing before an AI project. Focus on the ones that will most directly limit model performance.
Fix your join keys. If customer records in your CRM can't be matched to their billing records, fix this before you start. Create and maintain a unified customer identifier across systems. This single fix unlocks cross-system analysis that was previously impossible and dramatically improves the data available for AI training.
Standardize your categoricals. Run deduplication on your categorical fields — industry codes, geographic fields, product categories, status values. Create a controlled vocabulary and update historical records to use it. This work improves not just AI model quality but the reliability of your existing business reporting.
Fix your missing values systematically. For fields that are critical to your AI use case, implement processes to ensure they're captured going forward. For historical missing values, decide on an imputation strategy (fill with median, mode, or a dedicated "missing" category) and document it. Inconsistent handling of missing values is a common source of silent model degradation.
The Realistic Timeline
Data remediation is not instant. For most organizations, a focused data quality program takes 4-8 weeks of dedicated effort before an AI project begins. This feels frustrating — it delays the exciting part — but it dramatically improves AI project outcomes.
Think of it as building the foundation before the house. Organizations that skip data preparation and start building AI systems are building on sand. They discover the problems during the AI project, which costs more to fix than if they'd addressed them upfront, and they often end up with AI systems that don't perform well enough to deploy — or that degrade quickly after deployment as the data pipeline quality issues cascade into model outputs.
The AI vendor who tells you your data is fine and you can start immediately is not doing you a favor. The vendor who does a rigorous data assessment and gives you an honest picture of what needs to be fixed before you can build effectively — that's the partner you want.