Glossary
The words behind the grades.
Plain-language definitions of the terms we use in the Data Readiness Verdict, its read-outs and our articles, each with why it matters when the data has to feed an AI model. What earns each grade is in the grading rubric.
The five grades
- Data quality
Whether the values in a system are correct, complete and consistent, and whether anyone measures that on a schedule.
Why: A model learns whatever errors the data holds, at scale. Unmeasured quality means nobody knows how wrong the model's inputs are.
- Documentation
Whether the meaning of each field is written down, kept current and agreed across the teams that use it.
Why: If nobody can say what a field means, nobody can map it to what a model expects, and training labels end up disagreeing with each other.
- Data lineage
The record of where each value came from, what transformed it on the way, and what depends on it downstream.
Why: Regulators and auditors ask where every model input came from. Without lineage, you can't answer, and you can't rebuild what the data looked like on a past date.
- Access and security
Whether the right people can reach the data through a defined, reasonably fast path, with sensitive fields classified and protected.
Why: AI teams that wait months for data never leave the pilot. Teams that get it without controls on personal or medical fields create a different problem.
- AI consumability
Whether a model can read the data it needs, in a structured form, at the moment it has to make a decision.
Why: Accurate, documented data is still useless to a model if it sits in free text, scanned PDFs or a weekly batch when the model scores in real time.
Data foundations
- Data profiling
Measuring real data, field by field: how complete it is, which values are invalid, where duplicates are, and how records link across systems.
Why: It turns opinions from interviews into evidence. Every grade in a Data Readiness Verdict rests on it.
- Data dictionary
A technical reference listing each table and field in a system, with its type, allowed values and meaning.
Why: It's the starting point for mapping source data to model inputs. Where it's missing or stale, documentation grades fall.
- Business glossary
The agreed business definitions of shared terms, such as what counts as an open claim or an active member, owned by named people.
Why: When claims, finance and actuarial define the same term three ways, a model's training labels disagree before training starts.
- Data steward
The person accountable for the definitions and quality of a set of data, and the first call when it's wrong.
Why: Gaps without an owner don't get fixed. Naming stewards is usually one of the cheapest, earliest roadmap items.
- Golden record
The single, trusted version of an entity, such as a customer, policyholder or provider, assembled from the records different systems hold about it.
Why: Models that compare patterns across records need one identity per entity. Without it, one provider or claimant looks like several.
- Entity resolution
The process of deciding which records, across or within systems, refer to the same real-world person or organization.
Why: It's how a golden record gets built, and it's often the fix that unblocks fraud and network models.
- Narrow path
The handful of systems, fields and joins a specific AI use case depends on, as opposed to the whole data estate.
Why: Fixing the narrow path gets the first use case into production in months; fixing the whole estate first rarely finishes.
- Data Readiness Verdict
A six-week, fixed-fee review of the data behind one AI use case: real extracts from every system it depends on are profiled, each system is graded A–F on five dimensions, and the result is a written verdict: go, fix then go, or no-go.
Why: It answers the funding question before the budget is spent. Profiled, not surveyed: interviews tell you what people believe about the data, and profiling shows what it actually is.
Testing AI
- Decision point
The moment in a process when a model is asked to make its prediction or recommendation, such as first notice of loss or claim submission.
Why: A model can only use what exists at that moment. Everything about readiness is measured against it.
- Point-in-time data
Data as it actually existed at a given moment, before later corrections, coding or adjustments were applied.
Why: Testing a model on point-in-time data is the only way to get an accuracy number that will hold in production.
- Leakage
Information from after the decision point creeping into the data a model is trained or tested on.
Why: It makes demos look far more accurate than production will ever be, and it's the most common reason a funded model disappoints.
- Written accuracy target
An agreed, documented threshold for how well a model must perform, measured on decision-time data, before it's built.
Why: It turns a demo into a commitment, and makes it clear when a model is ready to go live.
Insurance and health plans
- FNOL (first notice of loss)
The first report of a claim to an insurer, by phone, web, app or agent, and the data captured at that moment.
Why: It's where claims triage models decide, usually on the least structured data in the whole claim lifecycle.
- FWA (fraud, waste and abuse)
Improper healthcare billing, from deliberate fraud to wasteful or abusive patterns, and the programs that detect it.
Why: FWA models depend on resolving provider identities and keeping claim lines as originally billed, two common data gaps.
- NPI and TIN
The National Provider Identifier for an individual or organization, and the Taxpayer Identification Number used for billing. One clinician can appear under several of each.
Why: If they aren't resolved to one identity, the pattern a fraud model is looking for is split across what look like unrelated providers.
- Retroactive eligibility
A change to a member's coverage that applies to past dates, often overwriting what the eligibility record said at the time.
Why: Models need eligibility as of the date of service. Systems that overwrite history lose it, which is a lineage gap.
- PHI masking
Replacing or hiding protected health information in a dataset so it can be analyzed without exposing individuals.
Why: It lets profiling measure sensitive fields without anyone seeing patient details.