Data lineage
Data lineage is the record of where data came from, how systems transformed it, and where it was used. In AI value measurement, lineage connects an executive result back to AI usage, cost, workflow, quality, and completion records so another reviewer can understand and reproduce the metric.
Lineage does not prove that AI caused an outcome. It proves how the evidence was assembled.
What should AI measurement lineage capture?
Start with the reported metric and trace backward. Record the source system, dataset or event, field definition, transformation, join rule, filter, time zone, owner, and refresh date. If a model or analyst classifies records, retain the method and version.
For an AI-enabled support workflow, the chain may connect a reported cost-per-case measure to provider billing, application logs, case records, acceptance status, reopen events, and the calculation that joins them. The unit of work definition must remain visible within that chain.
Microsoft Purview represents lineage through relationships between datasets and the processes that move or transform them. Its documentation also positions lineage as a way to investigate the root cause of data-quality issues (Microsoft Purview lineage guide, Microsoft Purview overview).
OpenLineage provides an open model for recording dataset, job, and run metadata. Its job-lineage specification records explicit data-flow relationships rather than inferring that every input feeds every output (OpenLineage overview, Lineage Job Facet).
Why does data lineage matter for AI ROI?
AI value reporting often joins systems owned by different teams. Cost may come from a model provider or finance platform. Usage may come from an application log. Completed work and quality may live in a CRM, ticketing system, code repository, or finance system.
Without lineage, a dashboard can display a precise number while hiding incompatible definitions, missing records, duplicate joins, or outdated transformations. With lineage, reviewers can challenge the assumptions and repair the calculation without rebuilding it from memory.
Lineage also supports an AI audit trail. The two are related but not identical. Data lineage follows data and transformations. An audit trail records actions, decisions, approvals, and changes. A defensible investment review often needs both.
How does lineage support an AI experiment?
An experiment design defines the hypothesis, comparison, metrics, and decision rule. Lineage shows whether the final analysis used those definitions consistently.
It should identify which records belonged to each comparison group, how exclusions were applied, which model or workflow version operated during the pilot, and whether late-arriving data changed the result. The measurement window should appear in the calculation rather than only in a slide title.
This does not require one enterprise-wide lineage program before any pilot can begin. Start with the small set of sources that support the decision. Expand when a metric becomes reusable across teams or material to budget, governance, or external reporting.
What is the minimum viable lineage record?
For each decision metric, keep its business definition, formula, source fields, transformations, owner, observation period, refresh date, known gaps, and link to the underlying query or model. Record manual adjustments separately rather than overwriting source data.
A metric should fail review when its source cannot be located, its denominator changes without documentation, or its result depends on an unexplained spreadsheet step. AI governance can set the control, but the lineage record supplies the evidence.
Common mistakes
Common failures include documenting only the final dashboard, treating a data catalog entry as complete lineage, losing model and workflow versions, and omitting manual overrides. Another mistake is collecting every technical dependency while leaving the business rule undefined.
Useful lineage follows the path needed for a decision. It gives finance, data, and the workflow owner enough evidence to reproduce the result and state its limits.