Decision provenance is not a replacement for data lineage. It is the missing business-semantic join that connects lineage, runtime evidence, authority, action, and outcomes at the level of an individual decision.

A dataset can be perfectly governed, a model can be approved, and a prompt can be tested—yet the resulting decision can still be indefensible because the evidence chain was never joined.

AI makes traditional lineage necessary—but no longer sufficient

Classic lineage was designed to answer questions such as: Where did this field originate? Which transformation changed it? Which report consumes it?

Those questions still matter. But an AI-enabled decision may also depend on a prompt version, a retrieved document set, an agent’s selected tool, a policy evaluation, a delegated identity, a human override, and the business event that followed.

None of these is reliably represented by dataset-to-dataset lineage alone. The governing object is no longer just the data pipeline. It is the decision path.

Traditional lineage

  • Sources, datasets, transformations
  • Jobs, tables, schemas, reports
  • Upstream and downstream dependency
  • Best question: “Where did this data come from?”

Decision provenance

  • Context, model, prompt, retrieval and tools
  • Identity, delegation, policy and approvals
  • Actions, reversals, incidents and outcomes
  • Harder question: “Why was this action produced and allowed?”

The evidence stack enterprises actually need

“Provenance” is often used as a catch-all. Separating the layers makes the control gap visible.

Data provenanceSource systems, datasets, features, transformationsWhere did the input originate?
Model provenanceModel version, training lineage, evaluation, approvalWhich model was authorized?
Execution provenanceSpans, agent steps, retrievals, tools, messagesWhat happened at runtime?
Decision provenanceContext + execution + authority + actionWhy did decision X occur?
Outcome provenanceOverrides, complaints, incidents, value, harmWhat happened because of it?

The strategic gap is usually between execution provenance and outcome provenance. Decision provenance closes that gap.

From context to consequence: the decision evidence chain

Decision Provenance Record

Entities Activities Agents Policies Outcomes
User / system inputrequest, subject, purpose
Retrieved contextdocuments, versions, scores
Model & promptmodel, prompt, parameters
Tools & agentscalls, arguments, delegation
Identityhuman, service, agent authority
Policy decisionrule, version, evaluated facts
Human oversightapproval, override, rationale
Integrity evidencehash, signature, retention
Context assembly
DECISION ID
Evidence-backed action
Business workflow
Outcome
Human reversal
Incident / complaint
Measured value or harm

Do not begin by capturing everything

A defensible program can mature in four stages. Each stage should be judged by the questions it can answer—not by the volume of telemetry it stores.

Level 1

Inventory & accountability

Identify the application, model, prompt, environment, owner and timestamp.

“What system produced this?”
Level 2

Technical reconstruction

Add trace IDs, inputs, dataset references, retrieved documents and tool calls.

“What technical path led here?”
Level 3

Authority & governance

Add identity, delegation, policy version, approvals, risk tier and exceptions.

“Why was this allowed?”
Level 4

Outcome-linked assurance

Connect decisions to later outcomes, reversals, incidents, value and harm.

“What happened because of it?”

Decision provenance is assembled—not purchased whole

Catalogs, model registries, observability tools, policy engines, workflow platforms and evidence stores each cover part of the problem. A durable design uses two complementary planes.

01

Event plane

Immutable runtime events preserve source truth: model calls, retrievals, tool invocations, policy decisions, approvals and outcome events. CloudEvents-style envelopes and OpenTelemetry-compatible telemetry are useful building blocks.

02

Provenance plane

A typed graph—or graph-indexed relational model—materializes entities, activities, agents and causal relationships. W3C PROV offers a durable conceptual foundation.

Stable identifiers + point-in-time versions + integrity evidence = reconstructable decisions

Why the outcome link changes governance

Scenario: AI-supported supplier risk review

An AI agent summarizes supplier evidence, checks policy, calls a sanctions service, and recommends whether to approve onboarding. A human reviewer accepts the recommendation.

Three months later, the supplier is suspended after a compliance incident. Traditional logs can show system events. Decision provenance lets the enterprise reconstruct the exact evidence, policy, authority and reasoning path associated with the original approval.

1Which supplier documents and versions were retrieved?
2Which sanctions result and tool arguments were used?
3Which policy version allowed approval?
4What evidence was shown to the human reviewer?
5Did similar decisions under the same policy later fail?

Six implementation moves that create real evidence

01

Define the decision unit

Start with a stable decision_id. Without it, telemetry remains disconnected.

02

Tier capture by risk

Use full fidelity for high-impact decisions and selective capture for lower-risk flows.

03

Version the context

Persist document versions, hashes, retriever configuration and model/prompt versions.

04

Make authority queryable

Link identity, delegation, policy input, approval and exception identifiers.

05

Close the outcome loop

Connect reversals, incidents, complaints and realized business results back to decisions.

06

Measure reconstruction fidelity

Test whether sampled decisions can be reconstructed within an agreed SLA.

So—is decision provenance truly the next evolution of lineage?

Yes, with an important qualification. It is not “better lineage” and it should not erase the disciplines of data or model lineage. It is an evidence architecture built above them. Its purpose is to make decisions reconstructable across technical and organizational boundaries—and to connect them to consequences.

LINEAGE
+ AUTHORITY
+ OUTCOME

Standards and guidance informing this perspective

  1. W3C PROV-DM — conceptual model for entities, activities, agents and provenance relations.
  2. NIST AI Risk Management Framework and the Generative AI Profile.
  3. OpenTelemetry Generative AI semantic conventions for runtime AI telemetry.
  4. OpenLineage object model for jobs, runs, datasets and extensible facets.
  5. Open Policy Agent decision logs for policy evaluation evidence.
  6. EU Artificial Intelligence Act, including logging, documentation, human oversight and post-market monitoring obligations for high-risk systems.

This article presents an analytical synthesis. “Decision provenance” is an emerging enterprise architecture pattern rather than a single universally standardized product category.

About the author

I help organizations turn governance from a policy layer into an operating model—connecting data quality, metadata, stewardship, platform architecture, and trusted consumption across modern cloud ecosystems.

My work has consistently focused on the point where business trust breaks down: not only in bad source data, but in weak transformation controls, disconnected metadata, and ungoverned decision pipelines.