Open Data Product Initiative · Framework publication
Data Product Operating Framework
One universal operating model for directing, defining, operating and assuring data products, with enterprise and AI-agent-first implementation profiles.
Version 0.2.0Status DraftPublished 2026-10-06License CC BY 4.0
The Data Product Operating Framework Core at a glance.
Open Data Product Initiative · Framework publication
Data Product Operating Framework
One universal operating model for directing, defining, operating and assuring data products, with enterprise and AI-agent-first implementation profiles.
Version0.2.0 · Draft
Published2026-10-06
LicenseCC BY 4.0
Pilot
Start Here: Implement One Complete Data Product Chain#
Do not begin by converting an entire estate into contracts or by launching every capability at once. Start with one decision that matters and make its evidence path complete enough to review.
Register one owned business objective with a baseline and target.
Describe one use case and its information need.
Record one create, reuse or extend product decision.
Publish one validated ODPS product contract.
Publish one ODPC catalog reference to that contract.
Create one ODPG graph linking objective, use case and product.
Define one ODPR workflow for a repeatable validation or onboarding step.
Collect one quality or SLA observation against a declared commitment.
Measure one adoption signal that is meaningful for the use case.
Measure one outcome against its baseline.
Hold one portfolio review and record the continuation, change or stop decision.
The Customer 360 example shows this chain with a realistic extension decision and separate declared and observed state. Use it as a shape, not a template to copy blindly.
Choose a product with an identifiable owner, one motivated consumer group, an existing information gap, a manageable access path and an outcome that can be measured within a reasonable review period. A small regulated or high-risk product can also work if there is enough control evidence to learn from it.
Keep the pilot narrow. Existing ticketing, catalog, monitoring and approval tools may remain in place as long as their handoffs and retained evidence are clear. The goal is not a platform replacement. It is a reviewable chain that shows what is declared, what happened and who decides what to do next.
Ask whether the product was discoverable, whether access was appropriate, whether the declared commitment was observed, whether the consumer used it for the registered use case and whether the outcome justified continuing investment. If a link is missing, record it as the next improvement rather than inventing a new specification field.
The Data Product Operating Framework Core is a universal, vendor-neutral and technology-neutral operating model for managing data products from business demand to measurable value.
It provides a common operating structure connecting:
business objectives
use cases and demand
investment decisions
data products
machine-readable product contracts
semantics
governance requirements
delivery and consumption
lifecycle management
operational evidence
business outcomes
value
The Core works for organisations operating with:
human users
traditional applications
analytics systems
data platforms
automation
AI systems
AI agents
The framework does not require organisations to operate AI agents. It provides one common operating model that supports organisations from traditional enterprise data product management through AI-agent-first operations.
The Core remains machine-readable and automation-friendly by design. The ODPS standards family provides the machine-readable standards foundation underneath the framework. The framework is an operating model, not another ODPS-family specification.
The Core defines required operating concerns rather than one mandatory organisational design. A small organisation may combine several accountabilities in one role; a large organisation may distribute them across product, domain, platform, governance and assurance teams. What matters is that the decision rights, practices, artifacts, evidence and outcomes remain explicit and traceable.
Examples and external references explain how a requirement can be implemented. They do not silently create additional Core requirements. Applicable law, regulation, contracts and organisational policy may impose stronger obligations in a particular environment; those obligations should be represented through the relevant profile, product context and control model.
A managed data product must have an authoritative definition.
The contract describes:
what the product is
who provides it
what it provides
how it is accessed
what it means
what it promises
what conditions apply
how it changes
F4. Machine-readable by default, human-readable by presentation#
The authoritative representation should favour machine-readable structures.
Human-readable views such as web pages, PDF documents, catalog pages, dashboards and reports should derive from the underlying structured artifacts where practical.
ODPS is machine-readable by design. Human-readable documentation and interfaces should increasingly be generated from the same underlying product artifacts used by software and AI agents.
The Core is the map, not the complete implementation guide. The capability reference explains the work, evidence and assessment levels for C1-C14; the implementation playbook explains how to establish the work in practice; the examples show a complete chain.
Profiles adapt this shared foundation to specific operating environments. They are not separate frameworks.
The Core uses the following separation of authority. These responsibilities are complementary and must not be collapsed into a new framework specification.
Evidence is information retained so that a material claim, execution or decision can be examined. It should identify what it concerns, where it came from, when it was produced, which version and scope applied, how it was generated and who or what is accountable for it. The required strength and retention period depend on the consequence of the decision being supported.
Evidence may be a structured event, measurement, validation report, signed approval, decision record, test result or other durable artifact. A screenshot or narrative may be sufficient for a low-risk review but weak for repeatable automated assurance. The framework therefore specifies evidence purpose and traceability while allowing implementation profiles and organisations to define proportionate formats and controls.
The framework distinguishes four attribution levels:
Direct attribution
Contribution
Enablement
Mandatory value
Value claims should state the baseline or counterfactual, measurement period, cost boundary, assumptions and attribution method. Financial conversion is useful only when credible; operational, decision, risk, compliance and public value may require different units and decision criteria.
The value model is designed for investment decisions rather than promotional reporting. Uncertainty, negative outcomes and costs should remain visible. A product may be necessary despite weak direct financial attribution, or widely used while contributing little to the outcome that justified it.
Data Product Operating Framework Core → Enterprise Data Product Profile → AI-Agent-First Data Product Profile
The Enterprise Data Product Profile provides the baseline implementation profile for organisations managing data products through conventional enterprise systems and human-led operating models.
The AI-Agent-First Data Product Profile extends the Core with stronger requirements for environments where AI agents directly discover, interpret, access, combine or act on data products.
Profiles may:
extend Core requirements
identify stronger practices
identify required evidence
identify mandatory machine-readable artifacts
identify stronger control requirements
define context-specific assessment expectations
Profiles must not:
duplicate the entire framework
rename Core capabilities
create separate maturity models
fork Core terminology
redefine ODPS-family standards
Each profile uses the common C1-C14 capability model and the common assessment dimensions of Accountability, Practice, Evidence and Outcome.
Machine-readable by default, human-readable by presentation is a major Core principle.
The objective is that a significant portion of the Core traceability test can be answered from machine-readable artifacts and evidence.
The framework therefore aims toward:
machine-readable objectives and use cases
machine-readable data product contracts
machine-readable semantics
machine-readable relationships
machine-readable policies where suitable
machine-readable workflows
machine-readable runtime evidence
machine-readable assurance results
traceable value
Human-readable web pages, PDF documents, catalog pages, dashboards and reports should derive from the same underlying structured artifacts where practical. Generated views support communication; they do not replace the authoritative artifacts or the normative Markdown sources of this framework repository.
22. Evidence-informed design and external references#
The Core is an Open Data Product Initiative operating-model design. External sources inform its reasoning and provide implementation patterns, but they do not become framework requirements merely because they are cited. The governed Framework Evidence Library records authority, applicability, rights and intended use for each source.
Important supporting sources include:
w3c-dcat-3 — W3C Data Catalog Vocabulary 3 for interoperable catalog, dataset, service, distribution and relationship patterns
w3c-skos — W3C SKOS for concept identity, labels, semantic relationships and vocabulary mappings
w3c-prov-o — W3C PROV-O for provenance across entities, activities and agents
eu-ai-act-2024-1689 — Regulation (EU) 2024/1689 as a jurisdiction-specific legal source for applicable AI obligations, not as a universal Core requirement
iso-iec-42001-2023 and edm-council-dcam — ISO/IEC 42001:2023 and EDM Council DCAM as licensed comparison points; their protected content is not reproduced by this framework
Sources may disagree, address different objects or apply only in particular jurisdictions. Maintainers should cite the primary source, disclose applicability and preserve the framework’s authority boundary when proposing changes.
The Data Product Operating Framework is an open operating model for managing data products from business demand to measurable value.
It connects strategy, use cases, investment, product contracts, semantics, governance, delivery, lifecycle management and assurance through shared machine-readable context and evidence.
The framework does not require organisations to operate AI agents. It provides one common operating model that supports organisations from traditional enterprise data product management through AI-agent-first operations.
The ODPS standards family provides the interoperable, machine-readable standards foundation underneath the framework.
The framework explains how the parts work together. It is not itself another ODPS-family specification.
The fourteen references explain what each capability means, why it matters, how to implement it, what evidence to retain and how to assess progress. They preserve the Core architecture; they do not create a second framework.
C1 ensures that a data-product investment starts with an owned business, public-value, risk or regulatory objective rather than available data or a technology project. It makes the expected change, baseline, target and decision owner visible before a product is proposed.
Without an explicit objective, teams can deliver technically useful assets that no decision-maker is prepared to fund, measure or stop. Later claims of value then have no baseline or accountable sponsor.
Record the objective in outcome language, distinguish it from a delivery activity, name its accountable owner and agree how progress will be measured. Review it when a material assumption changes and retain the reason for changing a target or metric.
The objective owner is accountable for the outcome. Portfolio decision-makers use it to prioritise investment; product owners and analysts contribute evidence; governance participants test applicable constraints.
ODPC can represent a BusinessObjective and KPI. ODPG can connect it to use cases and products. ODPS is not the place to invent portfolio-level objective fields.
The customer-retention director owns the objective to reduce annual churn from 18% to 14% by the end of 2027. The baseline comes from the finance-approved churn report, not from a later model output.
Agents may read approved objective context for prioritisation, but cannot reinterpret a target, accept risk or approve investment. See the AI-Agent-First Profile.
Level 1 has isolated statements of intent. Level 2 has named owners and repeatable objective records. Level 3 links active demand to approved objectives. Level 4 measures progress and exceptions. Level 5 uses outcome evidence to adapt portfolio choices without delegating accountable decisions to automation.
C2 describes why information is needed before deciding what product to build. A use case identifies the consumer, decision or process, expected outcome and information need without prematurely prescribing a new data product.
If demand is captured as a product request, duplicate products become likely and teams cannot tell whether a need was met after the implementation changes.
Apply it to new requests, material product changes, regulatory obligations, recurring service problems and agent scenarios that need governed information.
Capture the use case in the consumer’s language; state the decision, required information and success measure; search existing products before approving a candidate; document the gap and prioritisation rationale.
The use-case owner is accountable for the problem statement and outcome. Consumers validate usefulness, product owners assess reuse, and portfolio governance approves priority.
Outputs are a qualified use case, information needs, an existing-product assessment, gap statement, priority and links to objective and product decision.
ODPC can represent the UseCase. ODPG expresses its relationship to objectives and products. ODPS represents a later product contract, not the use-case demand.
Retention analysts need to identify customers likely to leave in 30 days. They require recent purchases, service interactions, loyalty status and customer profile; the search finds Customer 360, Orders and Loyalty but exposes an insufficient service-interaction capability.
An agent may propose a candidate product from an approved information need, but must show the discovery evidence and stop for a human product decision.
C3 decides whether to create, reuse, extend, merge, maintain, deprecate or retire a product, and who remains accountable for its outcome and operating cost.
Apply it after demand qualification and at material lifecycle events: investment, major change, persistent assurance failure, loss of demand or retirement.
Compare options, select a lifecycle decision, assign an accountable product owner, identify build and operating funding, and set a review trigger. Record what evidence and assumptions informed the decision.
The portfolio decision-maker is accountable for investment; the product owner is accountable for the product outcome; finance, governance and platform roles advise or approve within their authority.
ODPC holds product references and portfolio context; ODPG connects decisions to use cases, objectives and dependencies. ODPS identifies the resulting individual product only after the decision.
Measure active products with owners and operating funding, overlapping propositions, products without active demand, review timeliness and avoided duplication.
The customer portfolio board decides to extend Customer 360 with service interactions instead of creating a churn-only duplicate. It funds the extension, names the customer-data product owner and requires a six-month value review.
Funding delivery but not operations, assigning a delivery manager instead of an accountable owner, and retaining products because no review date exists are common failures.
C4 defines a reusable data product as a contract between provider and consumer. It establishes identity, proposition, responsibility, lifecycle, version and consumer-relevant interfaces in an authoritative machine-readable artifact.
Without a contract, catalog descriptions, platform configuration and tribal knowledge drift apart. Consumers cannot reliably identify which product version or promise applies.
Assign a stable product identity, author an ODPS contract using the current schema, validate it, approve it, publish an authoritative reference and keep human-facing views aligned with it.
The product owner is accountable for what is offered. Product engineers and stewards maintain the contract; governance and security participants approve applicable conditions; consumers review usability.
ODPS is primary. ODPV supplies referenced meaning; ODPC carries a lightweight discovery reference; ODPG connects the product to its context. Do not duplicate the full contract in ODPC.
Measure active products with a valid current contract, validation failure rate, unresolved references and time from approved decision to published contract.
Customer 360 version 2.4 has a validated ODPS contract describing its reusable customer profile proposition and the current provider responsibility. The service-interaction extension is released through a new approved contract version.
Typical failures are treating a catalog page as the contract, copying incompatible fields from another specification, or keeping different human and machine descriptions.
An agent needs the authoritative contract, version and resolvable references before it can reason about product suitability; see the AI-Agent-First Profile.
Level 1 has inconsistent documents. Level 2 defines a contract practice. Level 3 uses validated contracts for active products. Level 4 measures completeness and synchronization. Level 5 drives validation, discovery and governed workflows from contracts.
C5 governs the concepts whose interpretation changes a decision, rule, calculation or integration. It separates a stable concept identity from labels and aliases.
Ambiguous terms such as “active customer” allow two correct-looking but incompatible uses of the same product, especially across domains and AI-assisted work.
Apply it when a term is shared across products, appears in a material commitment or policy, or could be interpreted differently by consumers or agents.
Reuse a definition where possible; assign identifiers; retain preferred labels, aliases, definitions and relationships; map external terms cautiously; version changes and check contract and graph references.
A semantic steward is accountable for definition quality. Product owners apply terms, domain experts validate meaning and governance participants resolve material conflicts.
ODPV is primary. ODPS and ODPC reference the meaning they need; ODPG shows where concepts and products are connected. Semantics are not an unstructured note in every product contract.
Customer 360 uses the approved definition of “customer” and documents that “service interaction” includes a recorded assisted contact, not an abandoned call event.
Agents should resolve a concept identifier and definition before joining or acting on product data; ambiguity is a stopping condition, not a prompt to guess.
Level 1 has local labels. Level 2 maintains managed definitions. Level 3 applies them across active products. Level 4 detects drift and unresolved mappings. Level 5 uses semantic context in discovery, validation and governed agent workflows.
C6 makes clear what a consumer may expect and under which conditions it may use the product. A commitment includes the subject, metric or rule, threshold, period and measurement method.
“High quality” and “available when needed” cannot be assured. An access rule without a scope, approver or enforcement path leaves consumers and providers guessing.
Define testable commitments; separate a declared rule from enforcement configuration and observed result; publish conditions before access; record precedence and exceptions.
The product owner is accountable for product commitments. Security, privacy, legal and service owners approve conditions in their remit; consumers accept applicable terms.
ODPS represents applicable product commitments and access information. ODPV can clarify terms. Runtime systems record observations; the framework evidence layer retains proof that conditions operated.
Customer 360 declares completeness of at least 98%, freshness no more than 60 minutes, availability of 99.9%, approved-consumer access and human approval before production release.
C7 records the material relationships that let a person or system traverse from objective to use case, product, policy, workflow, consumer and outcome.
An isolated catalog record cannot answer what breaks when an upstream product, policy or definition changes. Free-text dependency notes are neither testable nor reliably traversable.
Record both ends and direction of a relationship; distinguish confirmed from inferred edges; validate references; review critical paths; retain provenance and resolve conflicts rather than overwriting them.
The Customer 360 graph links the churn use case to its objective, Customer 360, Orders, Loyalty and the service-interaction gap. It also connects the extension workflow and retention outcome.
C8 publishes a product’s proposition, status, owner, current version and authoritative contract reference so a consumer can decide whether to investigate or request access.
If discovery relies on personal knowledge, existing products are missed and duplicate demand becomes the default. A catalog copy that drifts from the contract creates false confidence.
Create a lightweight catalog reference to the authoritative model, expose consumer language and material conditions, validate references, publish through the approved catalog and reconcile synchronization failures.
The catalog owner is accountable for the discovery service; the product owner confirms the offering; stewards maintain quality; access and governance teams set visibility boundaries.
ODPC is primary. It references, rather than repeats, the ODPS product model. ODPG provides relationship context and ODPV enables consistent search terms.
Customer 360 has an ODPC ProductReference that points to the authoritative ODPS version 2.4 contract and exposes its customer-retention value proposition, status, owner and restricted visibility.
Typical failures are copying the entire contract into the catalog, publishing unapproved drafts, or equating search visibility with permission to access.
Agents need machine-readable discovery results that distinguish candidate suitability from access approval and identify the authoritative product version.
C9 turns a permitted request into usable consumption with the correct interface, version, terms and support path. It preserves the difference between being discoverable, being approved and being able to use the product successfully.
Credentials alone do not prove a consumer received the right product or understood the conditions. Missing consumption evidence also makes later adoption claims unreliable.
Consumer identity, intended use, applicable conditions, product and interface version, approval route, entitlement design and support ownership are needed.
Evaluate a request, record a decision, provision the allowed scope, communicate version and conditions, perform a safe connectivity or acceptance check, and retain proportionate consumption evidence.
The access owner is accountable for the entitlement decision. Product owners supply interface and condition context; security and privacy roles approve constrained cases; consumers confirm intended use.
ODPS describes product access where supported. ODPR can express a portable onboarding workflow. Runtime entitlement systems and logs provide operating evidence, not new ODPS fields.
The retention analytics group receives approved read access to Customer 360 version 2.4 for churn-risk analysis. The evidence includes the approved purpose, restricted scope, connection test and terms acknowledgement.
Common failures include treating account creation as onboarding, granting broad access because a product is discoverable, and collecting invasive usage telemetry without need.
C10 manages the product lifecycle and makes consumer-relevant change visible. It distinguishes an implementation change from a change to meaning, interface, commitment or usage condition.
Classify the change, update the authoritative artifact, assess graph impact, obtain required approval, communicate to affected consumers, reconcile deployed versus declared version and retain rollback or retirement evidence.
The product owner approves within delegated authority. Engineers implement, consumers validate material changes, and governance participants approve risk, privacy or policy implications.
ODPS carries product version and lifecycle declaration. ODPG identifies affected context. ODPR can express a reusable release workflow. Operational deployment evidence remains external to the contract.
Customer 360 declares 2.4, while runtime evidence shows 2.3. The mismatch, missing approval for the latest deployment and affected retention consumers become an assurance and change-control issue.
C11 defines how repeatable data-product work is initiated, executed, stopped, approved and evidenced. It applies to human-led, automated and agent-assisted operations.
An automation that only records a technical success hides whose authority it used, which artifact versions it acted on and whether it should have stopped.
Model the workflow, validate reusable ODPR recipes where appropriate, establish accountable ownership, log the actor and input versions, route exceptions and preserve human approval points.
The workflow owner is accountable for its purpose and controls. Performers or agents execute only within granted tools and permissions. Approvers make decisions the workflow explicitly reserves for humans.
ODPR is primary for portable workflow contracts. ODPS/ODPC/ODPG/ODPV artifacts may be inputs or outputs. Runtime logs are evidence and should not be misrepresented as recipe declarations.
The Customer 360 release recipe validates the proposed contract, checks the change record, pauses for human production approval and records the exact version that was deployed.
Typical failures are embedding policy in opaque code, allowing an agent to approve its own output, or treating a completed run as proof of a business outcome.
AI agents require an explicit tool allowlist, permissions, grounding sources, stop conditions, human-approval gates and provenance; see the profile rather than creating a second workflow model.
C12 measures service, quality and availability against the declared product commitments and makes the resulting evidence usable for remediation and product decisions.
A declared target is not proof of operation. Without versioned observations, teams cannot explain which promise failed, who was affected or whether a release caused the issue.
Collect observations, identify the evaluated version and period, compare results to declared thresholds, detect missing measurements, open remediation, notify affected consumers and retain the method and evidence.
The service owner is accountable for measurement and response. Product owners decide product implications; operations performs remediation; consumers receive material notices; governance reviews persistent failure.
ODPS declares product commitments where supported. Observations belong in runtime and evidence systems; ODPG can identify dependencies and affected consumers; ODPR can coordinate remediation.
Customer 360 is observed at 94.7% completeness, 43-minute freshness and 99.93% availability. Freshness and availability pass; completeness fails the 98% commitment. Retention analysts may be affected and the quality workflow must open a remediation record.
Level 1 has ad hoc monitoring. Level 2 defines commitments and measures. Level 3 retains observations and response evidence. Level 4 measures trends and consumer impact. Level 5 feeds assurance evidence into automated gates and accountable decisions.
Capability 13 · ASSURE
C13 Governance, Risk, Compliance and Control Assurance#
A policy link or a configured control is not evidence of execution. Overstating compliance can hide unreviewed risk and creates a false assurance claim.
Determine applicability, map controls proportionately, assign owners, perform and record control activity, assess findings, approve time-bound exceptions and verify remediation.
The control owner is accountable for operation. Product owners provide context; risk and compliance roles interpret applicability; internal assurance tests evidence; designated authorities accept residual risk.
ODPG can connect products to policy and control objects. ODPS may carry applicable product conditions but does not replace a control-management system. Evidence and findings are framework-level operating records.
For Customer 360, approved-consumer access and production-release approval are applicable controls. The observed unresolved entitlement and missing release approval create two findings; an exception exists only if a designated authority recorded one.
C14 evaluates whether consumers adopted the product, whether a supported use case improved and whether the organisation should continue its investment. It separates activity from outcome and outcome from claimed value.
Apply it after initial adoption, at agreed value-review intervals, when adoption declines, after material change and before continuation or retirement decisions.
Measure discovery, access and meaningful use; compare outcome to baseline; document contribution and uncertainty; include cost; review the evidence with product and objective owners; record the portfolio decision.
The product owner is accountable for product adoption evidence. The objective owner judges outcome relevance. Finance and portfolio governance assess cost and investment; analysts document method and limitations.
ODPC represents objectives and use cases; ODPG connects them to products. Runtime systems supply consumption observations. Value and attribution evidence belong in the framework operating record, not invented ODPS fields.
Measure active use cases, meaningful consumption, outcome movement, value confidence, unit operating cost and products that would not be funded again today.
At the Customer 360 six-month review, retention analysts used the extension in the churn workflow, but annual churn moved only from 18% to 17.2%. The review records other retention interventions and decides whether further investment is justified.
Agent consumption must be attributable to an approved workflow or purpose. High activity does not prove value, and agents cannot self-certify the outcome.
DIRECT ensures that data products exist for an explicit reason, serve identifiable demand, have accountable ownership, and receive investment based on expected value.
It creates the governance chain between organisational intent and individual data products.
Strategy becomes operational only when it changes a decision. An objective should therefore be specific enough to influence which demand is accepted, which product is funded, what outcome is expected and when investment should be reconsidered. Broad statements such as “become data driven” may express direction, but they do not provide a sufficient basis for prioritisation or assurance.
The objective record should preserve both accountability and evaluation context: the owner, intended beneficiaries, baseline, target, time horizon, constraints and assumptions. Measures may change as the organisation learns, but the reason for the change and the approving authority should remain traceable. This prevents later outcome reporting from being detached from the decision that originally justified the work.
Demand management is the point where an organisational ambition becomes a testable statement of need. A use case should identify the consumer, the decision or process to be improved, the present limitation, the expected change and the evidence that would indicate success. It should not prescribe a new product before existing products and non-product alternatives have been evaluated.
Separating use cases from products makes reuse visible. If several use cases require the same stable capability, that is evidence for a reusable product. If one proposed product bundles unrelated demand, the portfolio may need to split it. The use-case record should remain valid even if the eventual product, interface or implementation platform changes.
The portfolio is a decision system, not merely an inventory. It should expose competing demands, product overlap, dependency concentration, total operating cost and the evidence used to continue or stop investment. A catalog can list products, but only accountable portfolio governance can decide whether they should exist.
Accountability should distinguish the person who owns the product outcome from people who provide engineering, governance or platform services. Funding should include the continuing cost of operating, assuring and eventually retiring the product—not only initial delivery. Portfolio review should occur on a defined cadence and when material triggers arise, such as loss of active demand, repeated commitment failures, major dependency changes or a superior reusable product becoming available.
Decision records should state the options considered, evidence available, assumptions made and authority used. This allows later reviewers to distinguish a poor decision from a reasonable decision made with incomplete information.
Figure 5. Portfolio governance converts qualified demand into an explicit lifecycle decision and uses later evidence to revisit that decision.
Strategy and Objectives → Demand and Use Cases → Portfolio and Investment → DEFINE → OPERATE → ASSURE → Portfolio Review → Strategy Review
DIRECT remains active after approval. Evidence from ASSURE may invalidate an objective assumption, reveal a different consumer need or show that continued investment is no longer justified. The control loop therefore needs explicit review dates, trigger events and decision owners rather than relying on informal escalation.
The framework requirements above are Open Data Product Initiative operating-model decisions. The following external sources provide supporting context rather than additional conformance requirements:
nist-csf-2.0 — NIST Cybersecurity Framework 2.0 uses outcome-oriented guidance and explicitly allows implementation to vary by organisation and use case. That supports separating desired outcomes from prescribed implementation activity.
nist-privacy-framework-1.0 — NIST Privacy Framework 1.0 connects enterprise risk management with cross-functional accountability and decision-making. It informs the treatment of ownership, context and risk without making the privacy framework mandatory for all data products.
nist-ai-rmf-1.0 — NIST AI Risk Management Framework 1.0 connects governance activity to organisational priorities and risk tolerance. It is relevant when AI is in scope but does not make AI-specific governance a Core dependency.
edm-council-dcam — EDM Council DCAM is retained as a professional-framework comparison point for enterprise capability and assessment coverage. Its licensed content is not reproduced and it does not define framework conformance.
ODPC defines reusable portfolio objects and catalogs.
ODPV defines shared terminology and semantics.
ODPG defines relationships between products, use cases, objectives, policies and other objects.
ODPR defines repeatable work and workflow contracts around those artifacts.
Figure 6. A product contract is the stable centre of a connected definition: identity and version bind semantics, conditions, relationships and interfaces together.
The product contract is the stable point of agreement between provider and consumer. It should be independently identifiable, versioned and sufficiently complete for a consumer to determine what the product offers, who is accountable for it, how it can be accessed and which commitments and conditions apply. A catalog page or platform configuration may present parts of that information, but neither should silently become the only authoritative definition.
The contract should describe the product rather than the internal project that created it. Implementation details belong where they affect consumption, commitments or change impact; temporary delivery tasks do not. When information is maintained in another authoritative artifact, the contract should use a stable reference and define what happens if that reference cannot be resolved.
Human-readable pages should be generated from or demonstrably aligned with the same machine-readable definition. Parallel hand-maintained descriptions create ambiguity precisely when a consumer needs certainty about a version, interface or commitment.
Semantic management should focus first on concepts whose ambiguity changes a decision, rule, calculation or integration. Not every label needs central governance. Material terms—such as customer, active account, service region or reportable incident—need stable identity, a clear definition, accountable stewardship and an explicit relationship to local or external terms.
Labels and aliases support discovery, but they are not substitutes for concept identity. A renamed label should not silently create a new concept, and two identical labels should not be assumed to have identical meaning. Where multilingual labels are provided, the underlying identifier and definition should preserve equivalence or disclose the difference.
Semantic consistency must be checked across contracts, catalogs, graphs, policies and workflow inputs. A vocabulary is useful only when implementations actually reference it and changes are governed.
A commitment should identify the subject, metric or rule, threshold, evaluation period, measurement method, scope and consequence of failure. “High quality” and “available when needed” are aspirations, not testable commitments. A product may also disclose non-binding objectives, but consumers must be able to distinguish them from approved obligations.
Usage conditions should be understandable before access is granted. They may include permitted purposes, prohibited uses, attribution, retention, privacy, security, geographic, licensing or commercial conditions. Machine-readable expression improves consistency and automated evaluation, but expression alone does not create legal validity or operational enforcement.
The provider should manage conflicts between conditions, record approved exceptions and identify which rule prevails. A declared rule, an enforcement configuration and evidence that the rule operated are three separate objects.
Connect the individual data product to the wider network of objectives, use cases, KPIs, policies, systems, products, APIs, workflows, consumers and organisational responsibilities.
Relationships turn isolated records into operational context. Each material relationship should identify both endpoints, use a defined relationship type and retain enough provenance to explain who asserted it, from which evidence, when and with what confidence. A free-text note that “product A depends on product B” is difficult to validate, traverse or use for impact analysis.
Not every possible edge belongs in the shared graph. Include relationships that support a decision, control, discovery path, dependency analysis, workflow or value chain. Derived relationships should be distinguishable from confirmed relationships, and conflicting assertions should remain reviewable rather than being silently overwritten.
Graph quality includes more than valid syntax. References must resolve, relationship direction must be meaningful, lifecycle states must be compatible and critical paths must have accountable owners. These checks allow a change in an objective, policy, interface or upstream product to be traced to affected consumers and outcomes.
A product moves into normal operational management when:
an accountable portfolio decision exists
the product has an approved identity
an authoritative product definition exists
required semantics are resolved
relevant quality and service commitments exist
access and usage conditions are defined
required relationships exist
the artifact package passes validation
required approvals are recorded
Exit is a controlled transition, not a claim that the definition is permanently complete. Any unresolved assumption, exception or planned semantic decision should be recorded with an owner and due date. Material changes after the transition return through the relevant DEFINE and lifecycle controls.
The framework requirements above remain Open Data Product Initiative decisions. These sources support the implementation reasoning:
w3c-dcat-3 — W3C Data Catalog Vocabulary 3 provides interoperable patterns for datasets, services, distributions, catalog records, identifiers and relationships. It informs catalog interoperability but is not a substitute for an ODPS product contract.
wilkinson-fair-principles-2016 — FAIR Guiding Principles emphasise persistent identifiers, rich metadata, provenance and machine actionability. The framework applies those ideas selectively; a FAIR digital object is not automatically a governed data product.
w3c-skos — W3C SKOS distinguishes concepts, labels, semantic relationships and mappings. It supports the vocabulary guidance without requiring every implementation to use RDF.
w3c-odrl-2.2 — W3C ODRL Information Model 2.2 separates permissions, prohibitions, duties and constraints. It informs machine-readable usage conditions but does not prove that a policy was enforced.
w3c-prov-o — W3C PROV-O represents entities, activities, agents and provenance relationships. It informs relationship and evidence provenance without redefining ODPG.
w3c-shacl — W3C SHACL demonstrates how machine-readable constraints can produce validation results. It is an implementation option, not a mandatory validation technology for the Core.
Make governed data products and surrounding business context discoverable by people, applications and automated systems, including AI agents where used.
Publication should help a consumer decide whether a product is suitable, not merely confirm that a record exists. Discovery information therefore needs product identity, proposition, provider, lifecycle state, version, relevant semantics, access path, commitments, usage conditions and relationships to demand or alternatives. Search relevance and business language matter alongside technical metadata.
The catalog is an index and presentation surface, not necessarily the authority for every field. It should preserve resolvable links to authoritative product, vocabulary, graph and policy artifacts and disclose when indexed information was last synchronized. Federation must not hide conflicting identifiers, stale records or different visibility rules.
Machine discovery requires structured metadata and a stable interface. Free-text catalog pages may help people, but automated consumers should not have to infer lifecycle state, access conditions or compatibility from prose alone.
Provisioning begins with an identified consumer and an intended use, then evaluates entitlement against the product’s access and usage conditions. Approval is one possible control; it should not be added where an automatic policy decision is sufficient, nor removed where accountable judgement or risk acceptance is required.
Access is not complete when credentials are issued. The consumer needs the interface version, endpoint or delivery location, authentication method, semantic context, operating limits, support path and a successful connectivity or acceptance check. These details should be versioned where a change could affect consumption.
Consumption evidence should associate activity with the product, version and consumer identity at an appropriate level of granularity. Privacy, security and proportionality constraints still apply: the framework does not require invasive tracking of individual behaviour to establish product adoption.
Versioning communicates consumer-relevant meaning. An organisation should define what constitutes a breaking, compatible and editorial change for its product types and apply that classification consistently. A changed implementation does not always require a new public contract version, while a changed meaning, interface, commitment or condition often does.
Impact analysis should use known consumers, dependencies, workflows, controls and value paths rather than relying only on the delivery team’s judgement. The change record should identify the prior and proposed state, affected relationships, validation evidence, decision authority, release timing and recovery approach.
Deprecation is a managed period, not a label applied shortly before shutdown. Consumers need a supported alternative or explicit exception path, migration expectations and an end date. Retirement should revoke access, close or transfer obligations, update discovery records and retain enough evidence to explain the decision later.
Figure 7. Product change is a governed lifecycle path: material changes are classified, tested against their impact and either released, migrated or rolled back.
A workflow contract should state the intended outcome, authoritative inputs, ordered activities, decision points, expected outputs, validation gates, accountable roles and evidence to retain. It should remain distinguishable from a specific execution engine so the organisation can review or move the procedure without losing its meaning.
Human-led, application-led and agent-assisted execution are all valid Core implementations. Automation is appropriate when inputs, rules, permissions and failure handling are explicit. Where judgement or risk acceptance is material, the workflow should identify the human decision and the information needed to make it rather than disguising the decision as an automated step.
Execution records should identify the workflow version, input versions, actor or system identities, decisions, exceptions, outputs and final status. A successful technical run does not by itself show that the business outcome was correct; ASSURE evaluates service, control and outcome evidence separately.
what the product contract or workflow says should happen
what operational systems show happened
This difference becomes evidence for ASSURE.
OPERATE should preserve both sides of the comparison. Updating a declaration to match a failure can erase the signal; retaining observations without the applicable declaration removes the evaluation basis. Version and time therefore matter for contracts, policies, workflows and evidence.
The framework requirements above remain Open Data Product Initiative decisions. These sources provide supporting patterns:
w3c-dcat-3 — W3C Data Catalog Vocabulary 3 provides portable catalog, dataset, distribution and data-service structures as well as catalog-record lifecycle metadata. It informs federation and discovery but does not replace ODPC or ODPS.
wilkinson-fair-principles-2016 — FAIR Guiding Principles support persistent identifiers, rich metadata and machine-actionable discovery while preserving the distinction between accessibility and unrestricted access.
w3c-prov-o — W3C PROV-O provides a general model for entities, activities, agents and derivation. It informs workflow and execution provenance without requiring a particular runtime platform.
nist-sp-800-53r5 — NIST SP 800-53 Rev. 5 contains established control patterns for access, configuration, change, monitoring and evidence. Organisations should select applicable controls rather than treating the catalog as a universal checklist.
nist-ai-rmf-1.0 and nist-ai-600-1-genai-profile — NIST AI RMF 1.0 and its Generative AI Profile inform bounded AI operations, inventory, measurement and oversight where AI is used. They do not make AI or agents a Core requirement.
ASSURE determines whether data products operate as defined, satisfy applicable governance requirements, are adopted by intended consumers and create enough value to justify continued investment.
Observability connects a measurement to the exact product, version, commitment and evaluation period to which it applies. A dashboard value without this context may be useful operational telemetry, but it is weak assurance evidence. The measurement method, sampling window, exclusions and status logic should be explicit enough for another reviewer to reproduce or challenge the result.
Service assurance should combine technical observations with consumer impact. A short availability failure may be immaterial for one use case and decisive for another; a quality average may hide a critical segment. Evaluation should therefore preserve the declared threshold while allowing impact, severity and approved exceptions to be assessed separately.
The measurement system also requires assurance. Missing telemetry, changed metric definitions, delayed observations and unmonitored interfaces should be visible rather than interpreted as successful performance.
Control assurance begins with applicability. The organisation should be able to explain why an obligation, policy and control applies to a product and which authority accepted that interpretation. Copying a standard control catalog into every product obscures accountability and makes evidence review unmanageable.
A control definition should identify its objective, owner, trigger or frequency, scope, method, expected result, evidence and response to failure. Control execution then produces an observation about a particular product and version. Design approval, execution evidence and effectiveness assessment are different stages and should not be collapsed into one “compliant” status.
Exceptions are governed decisions with scope, rationale, compensating measures, accountable acceptance and expiry. Findings should connect to remediation and retest evidence. Independent assurance may sample evidence or repeat tests, but independence and sampling limits must be disclosed.
Adoption is meaningful use by an intended consumer, not an account count or isolated access event. The organisation should define what meaningful use means for each use case and separate initial trial, recurring consumption, production dependency and reuse by an additional use case.
Outcome measurement compares the present state with a baseline, target or credible counterfactual. It should record the period, population, calculation, assumptions and other changes that may explain the result. Attribution must be proportionate to the decision: a small operational improvement may use contribution evidence, while a major investment claim may need stronger causal analysis.
Value combines the outcome with its significance, cost and attribution. Financial value may be appropriate, but risk reduction, compliance, public value, decision quality and service improvement should not be forced into artificial revenue estimates. A mandatory product can have value even when stopping it is not a realistic option; the portfolio decision may instead concern cost, quality or risk.
Product exists → Product is discoverable → Consumer receives access → Consumer uses product → Use case changes → Outcome changes → Value is realised
Figure 8. Discovery, access and consumption demonstrate activity; value is established only when evidence connects that activity to use-case and organisational outcomes.
These should not be collapsed into one opaque score.
A combined executive view may summarize the four perspectives, but it should preserve drill-down to the underlying measures, evidence, exceptions and decision owners. A healthy service can still be unused, a compliant product can still lack value, and a valuable product can still require urgent control remediation.
The framework requirements above remain Open Data Product Initiative decisions. These sources support the assurance model:
w3c-dqv — W3C Data Quality Vocabulary separates quality dimensions, metrics, measurements, annotations and policy context. It informs machine-readable quality evidence without making RDF mandatory.
w3c-prov-o — W3C PROV-O supports attribution of evidence to entities, activities and agents. Provenance improves reviewability but does not by itself prove that a claim is true.
w3c-shacl — W3C SHACL distinguishes constraints from validation reports and individual validation results. This supports the framework’s declared-versus-observed separation.
nist-csf-2.0, nist-privacy-framework-1.0 and nist-sp-800-53r5 — NIST Cybersecurity Framework 2.0, NIST Privacy Framework 1.0 and NIST SP 800-53 Rev. 5 provide risk, control and assessment patterns. They inform applicable control design and evidence; they do not establish universal legal obligations for every product.
oecd-ai-principles-2024 — OECD AI Principles emphasise beneficial outcomes, accountability, robustness and lifecycle risk management for AI. They are relevant when AI participates in the product environment, while C14 continues to measure organisational outcomes rather than AI activity counts.
Does evidence show that the capability achieves its purpose?
The four dimensions should be scored separately before assigning the overall capability level. Strong documentation cannot compensate for missing accountability, and extensive activity cannot compensate for missing outcome evidence. An assessor should record the limiting dimension and the improvement needed to progress.
Evidence should be relevant to the capability claim, attributable to a known source, tied to the applicable period and version, and retained in a form that can be reviewed. The assessor should distinguish:
declared evidence — definitions, policies, plans, contracts and expected practices
execution evidence — records that the defined practice was performed
observed evidence — measurements and events showing what occurred
decision evidence — findings, approvals, exceptions and investment responses based on the observations
A template, policy or configured control is not execution evidence. One successful execution may demonstrate existence but not consistent operation. A dashboard may show observations while providing insufficient provenance or method to support assurance.
Sampling should reflect the frequency, variability and consequence of the capability. Assessors should record the population, period, sample method, exclusions and known limitations. Evidence supplied only for the assessment should be treated cautiously if it is not produced by normal operations.
The assessment scope should identify the organisation, portfolio, product population, profile, period and exclusions. Interviews can establish intent and explain evidence, but statements should not substitute for operating records where those records should exist.
Assessors should test forward and backward traceability. A forward test starts from an objective or use case and follows the chain to a product, execution, evidence and outcome. A backward test starts from a value or conformance claim and verifies the evidence, consumption, product version and original decision that support it.
Where evidence conflicts, record the conflict rather than selecting the more favourable source. Where a capability differs materially between domains or product classes, report the distribution and rationale instead of relying only on an average.
Profiles use the same capability levels and assessment dimensions as the Core.
Profile requirements add context-specific expectations for practices, artifacts, evidence and controls within the applicable Core capability. A profile must not rename capabilities, create a separate maturity model or treat automation as evidence of maturity.
Evidence demonstrates that the capability operates and achieves required outcomes.
Conformance and maturity answer different questions. An artifact can conform while the surrounding capability is weak. A mature capability can manage several artifact formats while one particular artifact fails validation. Reports should state the object and scope of each conclusion rather than using an unqualified “compliant” label.
Assessors should be sufficiently independent from the activity being assessed for the consequence of the conclusion. Self-assessment is useful for improvement, but it should be labelled as such. Claims used for certification, regulatory response or material investment decisions may require independent evidence review.
Automation may validate artifacts, collect observations and detect missing links. It must not silently decide ambiguous applicability, risk acceptance or outcome attribution. Automated results should preserve rule version, input version, execution time and exceptions so a person can examine how the conclusion was reached.
Calibration is required when multiple assessors or organisational units are involved. A common evidence pack and example scoring decisions should be used to test whether assessors interpret levels consistently.
Library sources nist-csf-2.0, nist-sp-800-53r5 and w3c-shacl support this section. NIST Cybersecurity Framework 2.0 supports outcome-oriented assessment while allowing implementation to vary by context. NIST SP 800-53 Rev. 5 provides established control-assessment concepts, and W3C SHACL demonstrates the separation between declared constraints and validation results. These sources inform the assessment logic but do not define conformance to this framework.
The playbook provides practical guidance for implementing the framework.
The Core defines what good data product operations require.
The playbook explains how to establish them.
Implementation should begin with a bounded portfolio or value stream, not an enterprise-wide metadata exercise. Select a scope in which accountable decision-makers, product providers and consumers can establish one complete traceability chain and use its evidence to make a real continuation or change decision.
Before starting, record the implementation sponsor, scope, intended profile, product population, current systems, material obligations, success measures and decision date. This baseline prevents tooling activity from being mistaken for operating-model adoption.
Do not begin by creating hundreds of ODPS files for products that have no clear business demand.
Choose one objective with an accountable owner and one or more active use cases. Confirm how a use-case outcome will be observed and who can approve, defer or reject a product candidate. Establish a lightweight portfolio decision record before designing the product contract.
connect products to portfolio context through ODPG
Begin with the smallest contract that lets an intended consumer understand and evaluate the product. Resolve material semantic ambiguity and usage conditions first. Record unresolved decisions explicitly instead of hiding them in prose or implementation defaults.
Use the contract identifiers and versions in catalog, provisioning, release and workflow records. The objective is not immediate platform replacement; it is traceability across existing systems. Where integration is manual, define the accountable handoff and evidence before automating it.
Select a small set of observations tied directly to declared commitments and the original use-case outcome. Confirm that missing data and failed measurements are visible. Produce an assurance view that can support a product or portfolio decision, rather than a dashboard that only reports activity.
Hold the first review while the implementation scope is still small. Use evidence to make at least one explicit decision: continue, change, expand, merge, constrain or retire. If the review cannot change a decision, revisit the accountability and measures before scaling.
An initial implementation should establish at least:
one governed business objective
one registered use case
one approved product decision
one valid ODPS product contract
one catalog reference
one relationship graph connecting objective, use case and product
one operational workflow
one service or quality observation
one adoption measure
one outcome measure
one portfolio review decision
This creates one complete traceability chain before scaling.
The minimum viable implementation is complete only when the chain has been exercised. Creating the artifacts without provisioning, observation and review demonstrates definition work, not an operating framework.
Standardise identifiers, templates, decision records, validation and evidence packaging. Train owners and consumers, then repeat the process across several products or domains.
Integrate portfolio, contract, catalog, provisioning, lifecycle and observability systems. Remove duplicate entry where an authoritative artifact can drive another view.
Use recurring capability assessment and product evidence to improve practices, strengthen selected profile requirements and change portfolio investment.
Use machine-readable standards as portable sources of truth where appropriate.
Do not let a UI become the only source of truth.
Generated HTML, PDF and catalog views should derive from underlying artifacts.
For each material field, decide which artifact or system has authority and how other systems receive updates. Avoid “bidirectional synchronization” as a default answer; without field-level authority and conflict rules, it creates multiple competing truths.
Migration can use temporary reconciliation reports. Record the old source, new source, transformation rule, unresolved differences and cutover decision. Preserve identifiers so historical evidence remains connected after systems change.
Keep human approval where judgement, risk acceptance or investment authority requires it.
Automate only after the manual or semi-automated practice has a clear input, rule, output, exception path and owner. Automation should expose failures and uncertainty rather than converting them to default success. Record rule versions and input versions so results can be reproduced.
The AI-Agent-First Profile extends the Core with stronger machine-readable context, workflow, control, provenance and assurance requirements. It does not create a separate framework.
An organisation can establish the Enterprise profile first and strengthen selected capabilities as it progresses toward agent-ready and agent-first operations.
Markdown should remain the human-maintained source.
Generated HTML and PDF should:
preserve headings and anchors
include document metadata
include framework version
include publication date
link back to source files
avoid changing semantics during rendering
Publication should also expose the framework or artifact version and stable anchors so references remain meaningful. Accessibility, responsive navigation and print quality are presentation requirements; they do not change the authority of the underlying source.
The Framework Evidence Library contains supporting sources and their applicability boundaries. Use it to research an implementation decision, not to copy external requirements into the Core. Relevant starting points include w3c-dcat-3, W3C DCAT 3, for catalog interoperability; wilkinson-fair-principles-2016, the FAIR Guiding Principles, for machine-actionable metadata; w3c-prov-o, W3C PROV-O, for provenance; and the cataloged NIST risk frameworks for proportionate governance and assurance.
For a material framework change, retain the research question, source identifiers, claims, limitations and architecture impact. LLM-assisted synthesis may propose text, but maintainers must verify source support and deliberately edit the normative Markdown.
This synthetic example follows a single decision from objective to portfolio review. It deliberately uses ODPS-family artifacts only for concerns represented by the available schemas; decisions, runtime observations, control evidence and value review remain framework-level records.
The chain is: objective → use case → information need → existing-product search → product decision → contract → vocabulary → relationships → catalog publication → access → version/change → workflow → runtime observation → control evidence → adoption → outcome → value → portfolio review.
The declared and observed states are intentionally separate. The 94.7% completeness, deployed version 2.3, unresolved entitlement and missing approval are not fields invented in ODPS or ODPR; they are operating evidence compared with the declared product state.
This lighter synthetic scaffold uses a regulatory-reporting product. The product has restricted visibility, purpose-bound access, an approved reporting calendar, lineage to upstream source products, a control mapping and retained evidence of review and release approval.
Use it to test C6 through C13: which reporting rule applies, who may access the data, which dependencies affect a filing, which version was used, which controls operated and how an exception is approved. The framework does not claim that an ODPS contract alone proves regulatory compliance.
This lighter synthetic scaffold covers a customer-support agent answering a governed business question. The agent discovers candidate products, reads the authoritative contracts, resolves vocabulary, traverses allowed relationships, checks its purpose-bound access, verifies version and current assurance evidence, and follows a permitted workflow.
The agent must stop and escalate when vocabulary is ambiguous, access is denied, the contract is stale, quality is failed, a relationship is missing or human approval is required. Its output includes the products, versions, graph path, checks performed and escalation evidence. It does not receive unrestricted autonomous access.
Information explaining how evidence was produced, including the relevant source, activity or method, time, scope, version and responsible person or system.
Provenance improves traceability and reviewability. It does not by itself prove that the evidence is accurate or that the supported claim is valid.
An implementation adaptation of the Core that strengthens practices, evidence, machine-readable artifacts, controls or assessment expectations for a particular operating environment without redefining Core capabilities.
The universal functions, capabilities, principles, terminology, traceability, evidence, assessment and conformance models shared by every framework implementation profile.
A person, organisation, application, system or AI agent consuming a product.
The consumer identity should be specific enough to apply access conditions, understand context and evaluate adoption while respecting privacy and proportionality.
The person or formally delegated role answerable for a decision, product outcome or capability. Tasks may be delegated, but accountability and escalation authority remain explicit.
The approved representation that has decision authority for a defined object, field or scope. Copies and generated views should resolve to or be reconciled with it.
Meaningful use of a data product by an intended consumer for an approved or recognised use case. Discovery, access, trial, recurring consumption and production dependency are different stages and should not be treated as equivalent.
An AI-enabled software actor that can interpret context, select or invoke tools, execute steps or take actions with some degree of autonomy. Within this framework, an agent remains subject to product contracts, policy boundaries, identity, authorization, workflow and evidence requirements.
An operating environment in which AI agents actively participate in discovery, product operations, workflows, assurance or decision support within explicit governance boundaries. It is an implementation stage and profile context, not a separate framework.
Represented in a structured form that software can parse and process according to explicit rules. A digital document containing only prose is not automatically machine-readable for framework purposes.
A scoped conclusion that an artifact, practice or capability satisfies stated requirements. The object, version, profile, assessment method and evidence supporting the conclusion should be explicit.
Enterprise Data Product Profile — baseline implementation for conventional enterprise systems and human-led operating models
AI-Agent-First Data Product Profile — stronger requirements for environments where AI agents directly discover, interpret, access, combine or act on data products
AI agents actively participate in discovery, product operations, workflows, assurance and decision support within explicit governance boundaries.
The stages describe an adoption path, not separate framework versions or maturity models. An organisation may adopt stronger practices selectively while retaining the same Core.
The Enterprise Data Product Profile provides the baseline implementation profile for organisations managing enterprise data products through conventional enterprise systems and human-led operating models.
It applies the Data Product Operating Framework Core without redefining its functions, capabilities, terminology, assessment model or conformance model.
Typical environments include:
data catalogs
data platforms
APIs
analytics systems
governance systems
human approval processes
product managers
domain owners
engineers
business consumers
The profile provides a realistic adoption path for organisations that do not yet operate AI agents extensively. Automation and AI may be used, but neither is required for profile adoption.
catalogs and platforms may be federated or heterogeneous
product information may be distributed across several enterprise systems
machine-readable artifacts can be adopted incrementally
existing governance and delivery systems remain operational authorities
the ODPS standards family supplies portable contracts and context where applicable
The profile does not require replacement of existing platforms. It requires clear authority, portable definitions, traceability and evidence across them.
Distributed implementation is acceptable when authority is explicit. For example, the product contract may be maintained in source control, discovery metadata in a catalog, entitlements in an identity platform and observations in monitoring systems. The organisation should define which system is authoritative for each object, how identifiers connect them and how conflicts or stale copies are detected.
Figure 9. Enterprise adoption does not require one platform: existing systems can remain operational authorities when stable identifiers, ownership and evidence flows connect them.
Objectives should be explicit, measurable where practical and owned by accountable roles.
At minimum, retain the objective owner, intended outcome, measure or evaluation method, time horizon and approval status. A generic transformation theme is not sufficient justification for every product placed beneath it.
Use cases should be registered, owned and linked to relevant objectives and data products.
The record should identify the consumer, problem or decision, expected change, evidence of success and evaluation of existing products. Registration may begin in an established demand-management tool if stable references connect it to the portfolio and product contract.
Products should have accountable owners, explicit investment decisions and governed lifecycle decisions.
Portfolio review should consider continuing operating cost, active demand, overlap, dependencies, commitment performance and outcome evidence. Human committees may make the decisions, but their rationale and authority should be retained in a durable record.
Active products should have authoritative ODPS definitions with stable identities, responsible providers, versions, interfaces and lifecycle states.
Existing catalog and platform metadata may supply inputs, but the approved ODPS artifact should be identifiable as the portable contract. Human-readable product pages should be generated from or reconciled with that artifact.
Shared vocabulary is recommended for important cross-domain concepts. Material terms should have consistent definitions across product contracts, catalogs and governance records.
Organisations can start with the concepts that most often cause reconciliation, reporting or policy errors. Stable identifiers and defined ownership are more important than building a comprehensive enterprise ontology before products can operate.
Relevant quality, service-level, access, usage, policy, privacy, security, licensing and commercial conditions should be explicit.
Conditions should be expressed at the level needed for a consumer and approver to act. Commitments intended for assurance need a metric, scope, evaluation period and method; policy references need an applicability rationale and exception path.
Critical product relationships and dependencies should be represented, including relationships to objectives, use cases, owners, upstream products, downstream consumers and applicable policies.
A spreadsheet or catalog relationship may be an acceptable starting point if identifiers and relationship types are controlled. Critical dependencies should support change-impact review and should not remain solely in architecture diagrams or individual knowledge.
Products should be discoverable through an approved catalog or equivalent governed discovery service.
Discovery should expose proposition, owner, lifecycle state, current version, access path and important conditions in language consumers understand. Federated catalogs should retain authoritative references and synchronization status.
Access and provisioning should follow defined identity, entitlement, approval and revocation controls.
The implementation should record the consumer, intended use, product version, decision and resulting entitlement. Provisioning is complete only when usable connectivity and relevant conditions have been communicated or tested.
Versions, lifecycle states, material changes, deprecation and retirement should be managed and communicated to affected consumers.
Change classification should reflect consumer impact. Known consumers and dependencies should be used for impact analysis, migration and notification rather than relying only on general catalog announcements.
Critical repeatable workflows should be documented, reviewable and increasingly automated where rules are clear. Human approval should remain where judgement, investment authority or risk acceptance is required.
The baseline can use existing ticketing, workflow or delivery tools when the workflow intent, inputs, decisions, outputs and evidence are explicit. ODPR adoption should improve portability and reuse rather than duplicate an already governed process without purpose.
Important service and quality commitments should be measured, compared with declared commitments and retained as evidence.
Measurements should identify the product and contract version, evaluation period, method and status. Missing observations should not be treated as successful performance.
C13. Governance, Risk, Compliance and Control Assurance#
Applicable policies and controls should be linked to products, operated by accountable roles and supported by evidence.
The profile does not require every enterprise control to be copied into every product. Applicability, ownership, execution evidence, exceptions, findings and remediation should be traceable through the systems already used for governance and assurance.
Products should be evaluated for discovery, access, adoption, supported use-case outcomes and organisational value rather than usage alone.
Start with a baseline and one outcome measure for each priority use case. Attribution may begin as a documented contribution assessment, provided assumptions, costs and uncertainty are visible to the portfolio decision-maker.
ODPS represents the individual data product contract.
ODPC represents portfolio and discovery objects.
ODPV represents shared vocabulary and semantics.
ODPG represents relationships and context.
ODPR represents reusable workflow contracts.
Operational systems execute work and provide runtime observations.
The evidence layer preserves what proves what happened.
ODPS definitions are expected for active governed products. The other ODPS-family standards should be adopted where shared portfolio objects, vocabulary, relationship graphs or portable workflows create practical value.
The framework and this profile do not redefine ODPS-family standards.
Assessment uses the common Assessment Standard, the same C1-C14 capability model and the same four dimensions:
Accountability
Practice
Evidence
Outcome
Profile requirements add context-specific expectations to those capabilities. They do not create a separate maturity model or allow artifact conformance to substitute for operating evidence.
Assessment should verify that human-led decisions are explicit, machine-readable artifacts are increasing, and operational evidence can be traced back to the relevant product, use case and objective.
This profile translates the Core into a conventional enterprise environment; it does not require conformance to the external sources below. Library sources w3c-dcat-3 and wilkinson-fair-principles-2016—W3C DCAT 3 and the FAIR Guiding Principles—support interoperable, machine-actionable discovery. Sources nist-csf-2.0, nist-privacy-framework-1.0 and nist-sp-800-53r5—NIST Cybersecurity Framework 2.0, NIST Privacy Framework 1.0 and NIST SP 800-53 Rev. 5—provide patterns for proportionate risk, access, change, control and assurance practices.
Organisations should apply legal, regulatory, contractual and sector-specific requirements through C6 and C13. A requirement that applies in one jurisdiction or product class should not be presented as a universal Enterprise Profile requirement without an explicit profile revision.
It strengthens selected Core capabilities. It does not rename them, fork Core terminology, create a separate maturity model or redefine ODPS-family standards.
the Core is already adopted as the common operating model
enterprise ownership and lifecycle controls remain in force
agents have identifiable operating contexts and bounded authority
authoritative artifacts can be resolved programmatically
agent actions can be observed and retained as evidence
human accountability remains explicit
higher autonomy requires stronger context, controls and assurance
Agent-first does not mean agent-only. Human review, escalation and decision rights remain part of the operating model where material judgement or authority is required.
The strength of implementation should reflect at least:
autonomy — whether the agent recommends, prepares, decides or acts
materiality — the consequence for people, services, finances, rights or obligations
reversibility — whether an action can be safely undone
reach — the number and variety of products, consumers and systems affected
uncertainty — the reliability of context, semantics, models and observations
exposure — the sensitivity of data and power of available tools
Higher consequence, autonomy or irreversibility requires stronger identity, authorization, validation, monitoring, approval, suspension and evidence. Low-risk assistance may use lighter controls, but it should still remain traceable to an approved use case and accountable owner.
Agent participation should be linked to an explicit objective, approved use case and expected outcome. Autonomy should not be introduced without a justified operating need.
Agent-supported use cases should identify the consumer, intended decision or action, acceptable outcome, evidence needs and material failure conditions.
Agent-enabled products and workflows should have accountable owners, explicit investment decisions and lifecycle decisions covering continuation, suspension and retirement.
The contract should allow an agent to determine what the product is, what it provides, how it can be accessed and which version is applicable without relying only on free-text pages.
Executable or automatically testable rules are preferred where practical. Machine-readable conditions do not remove the need for accountable policy ownership or exception decisions.
Require graph-traversable relationships for relevant:
objectives
use cases
products
dependencies
KPIs
policies
controls
APIs
agents
workflows
Relationship provenance and confidence should be retained where relevant. Agents should be able to distinguish authoritative, inferred and unverified relationships.
Require agent-oriented discovery through structured metadata and machine interfaces.
AI agents should be able to identify suitable products, resolve authoritative references and evaluate basic fitness for use without depending only on free-text catalog pages.
Discovery results should respect visibility, policy and entitlement boundaries.
Require machine-readable access and integration instructions.
Where required, agents should have identifiable identities and explicit authorization. Agent access follows the same product contract, policy boundaries, entitlement controls and revocation processes as other consumers.
Credentials should not be inferred from the identity of the human or system that initiated the workflow unless explicitly authorised.
Agent consumers should not silently continue using invalid assumptions after product changes. Cached context and workflow dependencies should be invalidated or reviewed when material contracts change.
Agent-assisted workflows should define where appropriate:
workflow intent
authoritative input artifacts
grounding boundaries
available tools
permissions
allowed actions
iteration limits
stopping conditions
expected outputs
validation gates
human approval gates
escalation conditions
workflow version
provenance requirements
ODPR should be the preferred portable workflow representation. Runtime-specific orchestration may execute the workflow but should not silently redefine its declared boundaries.
The profile expects stronger use of the common authority model:
ODPS is the authoritative individual data product contract.
ODPC supplies machine-oriented portfolio and discovery objects.
ODPV supplies shared vocabulary and semantics.
ODPG supplies graph-traversable relationships and context.
ODPR supplies reusable workflow contracts.
Operational systems execute work and provide runtime observations.
The evidence layer preserves what proves what happened.
These artifacts should use stable identifiers and resolvable references. The profile strengthens how the standards are used; it does not add a new ODPS-family specification or redefine an existing one.
Library sources nist-ai-rmf-1.0, nist-ai-600-1-genai-profile and oecd-ai-principles-2024 support this profile. The NIST AI Risk Management Framework 1.0 supports lifecycle governance, contextual risk analysis, measurement and accountable management. Its Generative AI Profile adds generative-AI risk and action patterns relevant to some agent implementations. The OECD AI Principles provide intergovernmental context for beneficial outcomes, transparency, robustness and accountability.
Sources w3c-prov-o, w3c-odrl-2.2 and w3c-shacl provide implementation patterns. W3C PROV-O informs execution and evidence provenance; W3C ODRL 2.2 informs machine-readable permissions and duties; and W3C SHACL provides an example of machine-testable constraints and validation results. These are implementation resources, not replacements for ODPS-family authority.
Library source eu-ai-act-2024-1689, Regulation (EU) 2024/1689, may impose legal requirements for particular AI systems, actors, uses and dates in its jurisdiction. This profile does not make those jurisdiction-specific obligations universal or provide a legal compliance determination. Organisations should map applicable duties to C6 and C13 with qualified legal and risk ownership.