From person-dependent to system-supported: a framework for governing data definitions

Most institutional research teams already know how to write a good definition. Ask any seasoned analyst what "full-time" means in an IPEDS submission versus a state report, and they can tell you precisely. The knowledge exists. What is usually missing is the structure that keeps that knowledge findable, current, and survivable after the person who holds it leaves.

That is the pattern across our research with 20 IR and IE professionals in 13 states. Creating definitions is achievable. Maintenance is where efforts collapse. A 300-definition document built by one IT staffer, then shelved when that person left after about a year, is not a documentation success; it is a maintenance failure waiting to be discovered. As one finding from the study puts it plainly: the institutional intelligence gap is not a gap in knowledge, it is a gap in infrastructure.

If you have read the first two posts in this series, you have the diagnosis. The first explained why one common term carries several valid definitions and why that fragments reporting. The second framed key-person dependency as the institutional intelligence gap and gave you a self-assessment. This post is about the build: how to move from person-dependent to system-supported without buying a single tool first.

A definition you can actually govern carries two dimensions at once: the functional (what a non-technical consumer needs to understand the number) and the technical (the logic a system needs to reproduce it). The first post lists the seven parts in full. The governance point is narrower: most documented definitions in the study captured the business meaning and stopped there, which is exactly the half that does not survive a staff change or a system migration. The governance metadata, who owns it, when it was last reviewed, what changed and why, is the part that decays first and is hardest to reconstruct after the fact.

Of the 20 respondents, exactly one maintained a dictionary that covered both the functional and technical dimensions with active governance, versioning, and campus-wide accessibility. That office documented 495 definitions and sat in the lowest-risk tier. The lesson is not the number. It is that completeness plus governance, not volume, is what separated the one survivable dictionary from the rest.

The framework is deliberately tool-agnostic. It pairs institutional process with selective AI rather than handing the whole problem to software. In the study, only one office ran anything close to a continuous-improvement model, updating definitions through a monthly KPI review by a governance council. That was the outlier, not because the others lacked tools, but because they lacked the process the framework formalizes.

Here is the part that makes this urgent rather than tidy. Adopting AI analytics, natural-language querying, or predictive models on top of an undocumented definition layer does not give you faster answers. It gives you faster wrong answers. A tool querying a warehouse where "enrolled student" has several undocumented meanings will return whichever one the query happens to hit, with no flag and no audit trail. The number looks authoritative because a machine produced it.

AI accelerates the work but does not replace the judgment. In one test from the research, roughly 30% of AI-generated definitions were usable on a first pass; the rest needed editing by someone with deep institutional knowledge. Scale that to a 500-definition dictionary and about 150 are ready immediately while 350 still need human context. That ratio is the whole argument: AI is a powerful drafting and auditing partner, and a poor substitute for the person who knows why a column named "master" is less authoritative than an unlabeled computed flag sitting next to it.

So closing the intelligence gap is not a prerequisite for using AI. It is a prerequisite for trusting what AI produces. The gap was always there; AI does not create the risk, it magnifies it. The institutions that document and govern their definitions first are the ones that will be able to believe their own dashboards once the querying gets automated.

Whichever tier you are in, the returns compound. Each definition you formalize lowers the cost of the next one, because the process, the template, and the precedent already exist. The first definition is expensive; the fiftieth is nearly free. That is the quiet argument for starting now rather than waiting for a migration or an accreditation year to force the issue. For more on turning that documentation into something leadership trusts, see our work on data-informed advocacy and best practices for IR and IE request workflows.

Clema is an AI data-intelligence platform built for IR and IE teams, and it maps onto the framework rather than replacing it. It connects to nine federal data sources, including IPEDS, College Scorecard, EADA, Pell Grants, DAPIP, and PSEO, and lets you query institutional and federal data in plain language. It flags discrepancies across sources, which is the audit step at scale. It serves both a technical dictionary and a consumer-facing glossary from the same underlying definitions, which is the democratization step. And it builds institutional memory that persists beyond any single person's tenure, which is the part the research kept identifying as the missing infrastructure. Thirty-five institutions use it today.

The honest framing matches the AI-readiness argument above: the tool accelerates lineage, audit, drafting, and maintenance, and your team still supplies the judgment about which definition is right for which context. Clema makes the framework cheaper to run. It does not run it for you.