Data quality

Data quality decides success or failure before the first model exists: which criteria matter, how to measure data quality and what poor data actually costs.

AI pilots rarely fail on the model,
they fail on the data

Not every dataset can carry a model. Whether yours can is testable before any budget flows into modelling, and that test is what this page describes.

The evidence

Why AI projects fail on data quality

The MIT NANDA programme reported in August 2025 that 95 per cent of the organisations it examined see no measurable business effect despite their spending on generative AI. A year earlier, Gartner had named poor data quality as one of the reasons at least 30 per cent of GenAI projects are abandoned after the proof of concept. Both match what we see in mid-sized companies: AI pilots rarely fail on the model, they fail on data whose quality nobody checked beforehand.

The mechanism is unspectacular. A model learns from what is in the dataset. If values are missing systematically, units are inconsistent or the same transaction appears twice, the model learns those errors too and delivers an analysis that looks convincing and is wrong. The most expensive moment to notice is after the project has started.

Seven questions decide whether your AI project fails: our open audit

Why 95 per cent of GenAI pilots fail: the methodical causes

Criteria

Data quality criteria: the six that matter

Data quality is neither a gut feeling nor a norm. It is the degree to which a dataset is fit for a specific purpose, so it is judged criterion by criterion, each with its own question and its own metric. The same dataset can be good enough for reporting and useless for forecasting.

The criteria only become useful once each one is tied to a metric and a threshold, and that is exactly what the next section does.

Prozessdatenerfassung: wo die Rohdaten für KI entstehen

Completeness

Are the mandatory fields populated? Measured as the missing share per column, after placeholders such as "?" or 9999 have been declared as missing values. A missing value is rarely an accident and often a finding in itself.

Consistency

Are types, units and formats uniform within each column? Type drift, dates stored as text and changing decimal separators are the most common silent sources of error in systems that have grown over years.

Accuracy and plausibility

Are the values physically and commercially possible? Outliers are counted and classified: sensor fault, mixed-up unit or genuine rare event. They are a cleaning task, not a reason to stop.

Timeliness

How current and how gapless is the time series? Every gap needs an explanation such as a weekend, a public holiday or a documented outage. Unexplained gaps are critical.

Uniqueness

Is every transaction in the data exactly once? Duplicates inflate aggregates and teach models patterns that do not exist. Countable technical duplicates are harmless, undocumented business duplicates are not.

Historical depth

Does the covered period answer the question? Forecasting typically needs 12 to 24 months of history depending on seasonality. If the history does not cover the question, that is a reason to stop.

Measuring data quality

Measuring data quality: seven dimensions, one verdict

Measuring means giving every criterion a metric and a threshold, and the threshold depends on the use case. So that this does not end in manual labour, we have published an open-source audit tool that checks a dataset against seven data quality dimensions in ten minutes: completeness, schema consistency, uniqueness, value plausibility, temporal gaps, referential integrity and representativeness. Each dimension returns a traffic-light verdict and a numeric score.

The tool runs where your data lives: on your infrastructure, not on ours. No dataset leaves the house for the check. That is not a feature, it is how we work in projects too. How the check fits into a project is described in the data acquisition phase of our workflow, and how findings are cleaned up in the data preparation phase.

Workflow phase 2: data acquisition and quality assessment in a project

Workflow phase 3: data preparation and cleaning

When it needs to be binding: our fixed-price Data Quality Check examines your data with a verdict per source. The result is yours, even if the answer is no.
Costs

What poor data quality costs

The cost of poor data never appears on an invoice. It hides in four places. First, in preparation: every undetected inconsistency becomes extra effort in data preparation, which already carries the largest share of project time. Second, in abandonment: a pilot that fails on the data after months costs the full budget and the internal trust in the topic. Third, in wrong decisions: a convincing, wrong analysis is more expensive than none, because plans are built on it. Fourth, in operations: unchecked data sources drift silently and the model degrades without anyone noticing.

The counter-calculation is short. A quality check before the project costs a fixed, small amount and answers the one question all four cost items depend on: does this data carry the plan or not? An honest no at checking time is the cheapest result an AI project can have.

Data strategy: where data quality fits into the bigger picture

Frequently asked questions about data quality

What is data quality, in short?

The degree to which a dataset is fit for a specific purpose. Data quality is always relative to the question: the same dataset can be sufficient for management reporting and unusable for a demand forecast. That is why every assessment starts with the purpose, not with the data.

What are the data quality criteria?

In our pre-project assessments we work with six criteria: completeness, consistency, accuracy and plausibility, timeliness, uniqueness and historical depth. Each criterion has a question, a metric and a threshold that depends on the use case.

What are the dimensions of data quality?

Our open-source audit tool checks seven dimensions: completeness, schema consistency, uniqueness, value plausibility, temporal gaps, referential integrity and representativeness. The last three are optional and need extra inputs such as the time column or a key relationship. Each dimension returns a traffic light and a numeric score.

How do you measure data quality?

Per criterion, with a metric against a threshold: missing share per column for completeness, type drift for consistency, Tukey outliers for plausibility, gaps larger than three times the median step for timeliness, duplicates on key columns for uniqueness. Tool-supported, a first pass takes about ten minutes.

How do you judge whether data quality is good enough for an AI project?

With a verdict per data source against the concrete use case, not with an overall grade. If the target variable is missing or the history does not cover the period of the question, the verdict is no, and it comes before any modelling. Everything else is a cleaning task with a calculable effort.

What does poor data quality cost?

It feeds four cost items: extra effort in data preparation, abandoned pilot projects, decisions built on wrong analyses and silent model decay in operations. The published abandonment figures for GenAI projects name poor data quality explicitly. A fixed-price check before the start is the cheapest insurance against all four.

Service

Let us talk about your data

Free 30-minute initial call. Briefly describe your dataset and your plan; optionally you can state a preferred date.

Free 30-minute initial call
Fixed price at a clear scope
Your data stays with you
Result as a management report

These services do not replace legal data protection advice.