Why good data matters for AI projects

A language model can summarise text and classify information. This does not automatically tell it which of several conflicting records your business considers authoritative. A more capable model alone cannot turn three conflicting customer addresses into one reliable address.

Fictional example: A customer has two names in the CRM, a spreadsheet contains the new address and the business system still holds the old one. The solution starts with a common customer identifier and an authoritative source, before AI uses the records to prepare a quotation.

An error in collection or classification can later affect retrieval, an answer and a decision. Research calls these effects data cascades. A qualitative study of 53 AI practitioners explores the issue in high-stakes applications; it does not establish a general failure rate for projects in German mid-sized companies.

Fictional example: A manual is assigned to the wrong product group. Search retrieves a similar machine, the answer names an unsuitable spare part and that answer is copied into a new instruction. Confident wording cannot compensate for an incorrect data foundation.

Start with the actual process and describe observed errors instead of declaring the entire dataset poor quality.

• Duplicates: A customer, product or transaction appears several times without a reliable common identifier.
• Inconsistent terminology: Departments use the same term differently or different terms for the same entity.
• Outdated versions: Superseded price lists and instructions remain available as valid sources.
• Unclear provenance: The source system, version, date or responsible role is missing.
• Missing ownership: Nobody can decide the business meaning of a field or how it should be maintained.
• Unsuitable context: An invoice list contains amounts but no cancellations, credit notes or disputed claims. It is then insufficient for evaluating payment transactions.

The Government Data Quality Framework describes quality as fitness for purpose. Its dimensions help you define your own checks; they are not automatic evidence of the quality of an AI system.

An internal reference system may have different requirements from decisions about people. Record which errors are acceptable and when a source must not be used.

• Completeness: Are the fields and documents required for the process present?
• Uniqueness: Can duplicate records be identified and reconciled?
• Consistency: Do values and terms agree across authoritative sources?
• Timeliness: Is the information current enough for the task?
• Validity: Do formats and values follow the agreed rules?
• Accuracy: Do records match verifiable reality?
• Also record provenance, the validity period and ownership.

A specific goal is more useful than “We want to use AI”. One possible pilot: Service staff find the appropriate approved instruction more quickly and check the drafted answer before sending it.

You do not need every company record immediately. List the necessary sources, locations, formats, owners, update frequency and access rules. Decide which system is authoritative for each type of information.

• Define the use case and permitted use.
• Set required fields, terminology, versions and access rules.
• Measure a baseline, such as missing required fields, detected duplicates and outdated documents.
• Collect real-world test questions, including incomplete and conflicting cases.
• Define acceptance and correction: Who reports errors and who decides which version is correct?

A one-off cleanup is insufficient if data entry continues to create duplicates. Examine the collection process, shared identifiers and ownership. A correction must not unintentionally destroy an original source or necessary historical data.

Define review intervals and a feedback channel. Reassess data and process quality after changes. This foundation helps distinguish whether a poor answer comes from missing data, unsuitable retrieval, an unclear task or the model.