Deck line: Metadata work is funded in four places, owned in none, and measured nowhere. This is what it looks like run as a single operated practice, with a scoring model attached.
Contents
SectionSubtitleExecutive summaryI · Why metadata work stallsFour budgets, four owners, one estateII · The practiceHarvest, Model, Version, Compile, VerifyIII · The measurementFive dimensions, one boundary, four operating numbersIV · The maturity modelDescribed, Partial, Traceable, OperatedV · OwnershipWhere the practice sits and who runs itVI · Evaluating a platform against the practiceTen questions, no vendors
The thesis: the metadata estate is an operated system with a lifecycle, a measurable state, and an owner. Most organizations run it as four disconnected projects, which is why the outputs keep disagreeing.
Executive summary
Most organizations fund metadata work four times.
A cataloging project buys discovery. A lineage project buys traceability. A governance program buys policy and stewardship. An AI initiative buys whatever context its application needs this quarter. Each is reasonable. Each has an owner, a budget line, and a roadmap.
None of them owns whether the estate as a whole is described well enough for a report, an agent, and an auditor to reach the same answer. That question has no owner, no measurement, and no line item, which is why it goes unanswered for years while all four projects report progress.
This paper describes the alternative: a single practice, run with the operating discipline that every other layer of the data stack acquired over the last fifteen years. Application code got version control and a deploy pipeline. Infrastructure got declarative state and a way back. Transformation logic got tested and promoted through environments. Business definitions, which govern financial reporting and increasingly govern what AI systems tell people, are still edited directly in production interfaces by whoever holds permissions.
Practitioners have started calling the practice MetadataOps. The operational half of the name is the substantive part: a build, a test, a deploy, a diff, and a rollback, applied to meaning.
What follows is the lifecycle, a five-dimension scoring model, a four-level maturity ladder, and the questions to hold any platform to. No vendor is named anywhere in this document, including the one that wrote it.
Part I · Why metadata work stalls
The pattern is consistent enough to be diagnostic.
An organization completes a substantial cataloging effort. Eleven thousand assets harvested from every warehouse, every BI tool, and most of the transformation code. Nine months of work, delivered on schedule.
The revenue figure in the executive dashboard still disagrees with the one in the finance model.
This is the signature of a program where one stage succeeded and nothing downstream of it did. The organization now holds a complete inventory of things it still cannot reason about, because an inventory answers what exists and the question on the table was what a number means.
Four budgets produce four partial answers
The four workstreams have genuinely different objectives, and each one stops at a boundary that made sense when it was scoped.
Discovery projects optimize for breadth of coverage and adoption. Success is measured in assets cataloged and searches performed, so depth of description and ownership of definitions fall outside the brief.
Lineage projects optimize for traceability. Success is a graph that renders, and the resolution of that graph is usually set by whatever technique was cheapest to implement.
Governance programs optimize for policy and accountability. They produce definitions, owners, and approval workflows, and they typically have no write access to the code that determines what the systems actually do.
AI initiatives optimize for a working application. They need context now, for one use case, and whatever they assemble is scoped to that application and dies with it.
Each program succeeds against its own definition of success. The estate stays undescribed in exactly the places that fell outside all four sets of success criteria.
The reconciliation tax that goes uncounted
The visible cost is easy to name and rarely tallied. Two teams present different numbers for the same metric, someone spends a week finding the discrepancy, and the incident is absorbed as a normal cost of doing business.
Ask an analytics team how many hours a quarter go to reconciling definitions across tools. The number is usually available and usually worse than expected. It is also the single most persuasive figure available when arguing for a consolidated practice, because it is already being paid.
Part II · The practice
Five stages. Each produces an artifact, each depends on the one before it, and each one skipped produces a failure that surfaces two quarters later and gets attributed to something else.
Harvest
Coverage of the estate: metadata collected from databases, transformation code, BI tools, and cloud platforms, across the systems that predate the current stack as well as the ones that replaced them.
The artifact is a record of what exists, with a known denominator. A harvest that covers the modern slice and stops is a harvest with an unstated denominator, which reports well and answers nothing.
Skipped: a well-adopted catalog people search and then leave to go ask a person.
Model
Business concepts defined once, each with a named owner, each bound to the specific physical assets that implement it.
The binding is the part that gets dropped, because it is the slow part. A definition written in prose and attached to nothing is a glossary entry, and glossary entries do not survive contact with a disagreement. The test of a modeled concept is whether it resolves to a table, a column, or a metric.
Skipped: four plausible definitions of revenue, all documented, none authoritative.
Version
Change history for the definitions themselves: attribution to a person, a retrievable prior state, promotion through environments, and a way back.
This is the stage that makes the practice operational. A change record that captures who and when, and leaves what unanswered, records activity. An audit trail records the thing being asked about.
Skipped: every retrospective question becomes a reconstruction project, staffed by the team whose work is under review.
Compile
Governed definitions written into the native formats of the systems that consume them, so each tool inherits the definition it uses.
Without this stage, definitions are modeled, owned, and versioned in one place while the BI tools, the transformation code, and the AI applications each re-implement them independently, because that is what those systems require to run. The governed definition stays correct and the shipped definitions drift, each on its own schedule.
Skipped: strong governance artifacts alongside persistent number disagreements. The two look contradictory and are the expected result.
Verify
Drift detection running continuously, impact analysis before schema changes ship, and an evidence trail available when someone asks where a number came from.
Every stage above produces state, and state goes stale. Schemas change, pipelines get rewritten, source systems get replaced.
Skipped: confidence with no expiry date on it, which is worse than doubt, because it stops anyone checking.
The order is a dependency chain
Harvest determines what can be modeled. Modeling determines what is worth versioning. Versioning determines whether compilation is safe to automate. Compilation determines whether verification has anything to check.
Teams that attempt compilation before modeling ship four definitions of revenue into six tools very efficiently.
Most metadata programs fail by working on stage four while stage two is incomplete. The work looks productive the whole time, the artifacts accumulate, and the results never arrive, because each stage can only be as good as the one feeding it.
Part III · The measurement
A practice that cannot be scored gets assessed on activity, and activity assessments are how metadata programs lose their budget in year three.
Five dimensions describe the quality of the metadata itself. Each is a measurable fact about the estate, established by tracing a metric by hand.
Coverage. What share of the estate has harvested metadata at all, stated with its denominator. The systems that predate the current stack belong in the denominator, because they are usually the ones feeding regulated reports.
Resolution. Column level or table level. Knowing that a report touches a system is a different thing from knowing which field produced the number, through which transformation, from which source. Table-level records answer the questions asked in a demo and fall silent on the question that matters, which is which of four similar columns the finance close actually uses.
Derivation. Parsed from the code that defines the transformation, or inferred from execution history. Inferred lineage reports what ran during the observation window, so jobs that did not fire leave gaps in a map that still looks complete. The test is direct: run a trace on a job that has not executed this quarter.
Freshness. Whether the record matches the estate as it stands today. A description maintained by hand describes the estate as of the last day someone had time to update it.
Governance state. Whether each definition has a named owner who would answer if asked today, a retrievable version history, a policy applied at the point of use, and a binding to a specific physical asset.
Scoring
Score each dimension from 0 to 4 across the three most-reported metrics in the business, and sum the averages for an index out of 20.
Choose metrics that appear in at least two consuming systems and that someone senior has asked about recently. Trace each one backward from the report to the source, by hand, one hop at a time. Budget an hour per metric and expect three. The tracing is where the findings come from, and the list of hops where the trace went dark is more useful than the number it produces.
The boundary that makes the score credible
These five dimensions describe the metadata. Bad values, null rates, data recency, and pipeline reliability belong to data quality tooling and data engineering.
A strong score says the estate is well described. It says nothing about whether last night's load was correct.
Stating that boundary is what makes the rest defensible. A measurement that claims to cover everything gets discounted by the first engineer who tests it.
The operating numbers
The index measures state. It is traced by hand, it takes three hours, and it moves on a timescale of quarters. That makes it the right instrument for a roadmap and the wrong one for a monthly report.
Every operations discipline that took hold did so on a second kind of number, measuring how fast, how often, and how reliably it runs. Software delivery got deployment frequency and lead time. The practice described here needs the equivalent, and four numbers cover it.
Definition-change lead time. Time from a proposed change to a business definition until every consuming system carries it. Measured per change, reported as a median. A lead time of weeks says the estate is documented; a lead time of days says it is deployed.
Definition-change failure rate. The share of definition changes that were rolled back, or that broke a report, a model, or an application after promotion. This is the number that tells you whether the Version stage is a control or a log.
The reconciliation tax. Incidents per quarter in which two consuming systems disagreed on the same number, and the hours spent resolving them. Most organizations already pay this and have never counted it. Start counting in week one; the log is the strongest argument you will have by week six.
Time to evidence. Minutes from a provenance question (where did this figure come from, what did this definition say last year) to an answer someone other than the author can reproduce. Assertion takes seconds and proves nothing. Reconstruction takes weeks. Evidence takes minutes, and the whole practice exists to move this number from the second to the third.
Report the four alongside the index. The index says where the estate stands. The four say whether the practice is running, and they are the numbers a data leader can put on a monthly dashboard without a briefing. None of the four requires a particular platform to measure.
Part IV · The maturity model
Four levels, corresponding to the index bands above. Each describes what an organization can and cannot answer, which is more useful than describing what tooling it has installed.
Described (0 to 6). Metadata functions as documentation. It supports discovery conversations and human judgment, and the humans supply the missing context from experience. An agent grounded on this estate produces confident answers with no way to check them, and an audit request becomes a reconstruction project.
The move to Partial is coverage. Extend harvesting to the systems currently outside it, starting with the ones feeding regulated reports.
Partial (7 to 12). The modern slice of the estate is in reasonable shape and the edges are dark. Agents are right about the covered path and wrong in ways that are difficult to predict, because the failure correlates with system age rather than question difficulty. Audit answers exist for some systems and get assembled by hand for others.
The move to Traceable is resolution and derivation. Column-level records, parsed from the code that defines the transformation.
Traceable (13 to 16). The organization can answer where a number came from across most of the estate, with named exceptions. This is the first level at which grounding claims survive a motivated technical challenge, and the first at which an audit response is a query for most of the questions asked.
The move to Operated is version and compile. Change history with environments and rollback, and definitions written into the systems that consume them.
Operated (17 to 20). The estate runs as a system with coverage, depth, provenance, freshness, and ownership. An agent grounds on a governed definition, and an auditor receives a query result on demand.
Read the sub-scores alongside the total. A 14 built on strong coverage and weak derivation is a different organization from a 14 built on the reverse, and the two are fixed in different orders.
Part V · Ownership
A practice with five stages and a measurement implies an owner. In most organizations that ownership is currently distributed across four teams, which produces the same outcome as no owner and takes longer to notice.
The placement decision turns on capability
Placement is usually argued from the organization chart. The narrower question is more useful: which team has write access to the transformation code, credentials on the legacy systems, and a deploy pipeline it already runs?
Four of the five stages are engineering work. Harvest requires connectivity. Version requires environments and promotion. Compile requires writing into other systems' native formats. Verify requires automated checks. A team that has to file a ticket for any of those will run the practice at the speed of the ticket queue.
Model is the exception, turning on business judgment. That split produces a workable default: the data platform or data engineering function owns execution, and governance owns the Model stage and supplies definitional requirements.
Two conditions override it. Where a BI consolidation is the forcing function, analytics engineering should lead, because the work is concentrated where they already sit. Where regulated reporting is the driver, governance should lead and embed an engineer with real write access, because the evidence requirements are the hard part.
Both overrides follow the same rule as the default: place the practice where the binding constraint already is.
The anti-pattern
The most common resolution is a virtual team, with representatives from each group, a monthly meeting, and shared accountability.
This reproduces the four-partial-owners condition with a calendar invite attached. The five stages generate artifacts, and artifacts need a single owner who can be asked why one of them is missing. A committee can set the standard. It cannot run the deploy.
Where a virtual team is the only politically available option, give one named person the artifacts and let the committee advise them.
Reporting is a separate decision
Where the work executes and who hears about it are different questions, and conflating them stalls the conversation. The practice can execute inside a platform team and report its score to a data leader monthly. That arrangement gives the engineer the access the work requires and gives the executive the visibility the investment requires.
The number to ask for is the completeness score with its denominator, plus the list of systems still dark.
Part VI · Evaluating a platform against the practice
Bring one of your own definitions to any evaluation. Ten questions, applicable to every vendor in the market.
01. What percentage of our estate can you harvest, and what falls outside it? Ask for the exclusion list. A connector count answers a different question.
02. Column level or table level, across which systems? Expect the answer to vary by source technology, and ask where it degrades.
03. Is lineage parsed from code or inferred from execution? Then test it on a job that has not run this quarter.
04. Show me what this definition said twelve months ago, in full. A platform storing current state alongside a change log will answer partially and confidently. The gap is visible in about a minute.
05. Who changed it, and when? A person, distinct from the service account that wrote the record.
06. Which reports consumed it at that time? Reconstructing consumption after the fact is harder than reconstructing the definition, because the consuming systems have also changed.
07. Show me a definition change moving from a lower environment into production, with the review step visible. This is what separates a controlled change from an edit that happened to be logged.
08. Can a downstream system inherit this definition in its own native format? Or does each consumer re-implement it.
09. What survives if we stop paying you? Turn the platform off and describe what remains in the systems.
10. What does a single answer cost? Where meaning resolves at query time, the resolution is billable. Ask for the figure on a representative question.
The first three sort on the supply side, the next four on the operating discipline, and the last three on what the organization is left holding. A platform that answers all ten by demonstration is rare enough that the exercise is worth running before a shortlist is drawn.
Where to start
Score three metrics. Trace each one by hand, record the five dimensions at every hop, and write down where the record goes dark.
The gap list is the beginning of the roadmap, and it is also the strongest available argument for resourcing the role that owns it.
Written by the team at MetaKarta and released under CC BY 4.0: free to use, adapt, and republish with attribution.
