If "customer" doesn't have the same definition everywhere, AI can't work
The executive committee looks at the monthly dashboard. Marketing reports 12,000 active customers. Accounting counts 9,500. Customer service sees 10,800. Three figures, three different definitions, zero confidence in the data. And when the company launches an AI churn prediction project, the model is fed by this inconsistent data and produces unusable results. The business glossary solves this problem at the root: a unique, shared and documented definition for each critical business term.
The problem
In a 300-employee mid-market company, the term "revenue" has at least 4 interpretations: billed revenue (accounting), collected revenue (treasury), ordered revenue (sales) and delivered revenue (logistics). When the CEO asks "what's our revenue this month?", they get 4 different answers. And no one is wrong — each answers according to their own definition.
This problem is found everywhere:
- Contradictory "customer" definitions — The CRM counts the records created (including prospects never converted). The ERP counts the accounts with at least one invoice. Marketing counts the newsletter subscribers. The result: a 20 to 40% gap between the counts, and monthly reports that never line up.
- KPIs calculated differently — The conversion rate is calculated from unique visitors for marketing and from quotes sent for sales. Both departments report a conversion rate but it doesn't measure the same thing. The executive committee compares incomparable figures.
- AI projects polluted by ambiguities — A sales prediction model is trained on "ordered revenue". The reporting compares it to "billed revenue". The natural gap between the two (cancelled orders, credit notes, billing delays) is interpreted as a model error. The AI project is judged unreliable and abandoned.
The root cause is always the same: absence of a shared glossary. Each department has developed its own definitions over time, in its own Excel files, without coordination. The business glossary breaks down these silos. See our data quality workshop for a guided approach.
The AI solution
The business glossary centralizes the business definitions and connects them to the technical data. AI speeds up its construction and its upkeep.
Centralized business glossary
A single dictionary accessible to everyone that defines each business term: name, precise definition, calculation formula, examples, Data Owner, source systems and quality rules. AI helps write the definitions by analyzing the real usage of the terms in existing SQL queries, dashboards and Excel files. It detects inconsistencies between departments.
Mapping business terms / technical columns
Each glossary term is linked to the physical columns of the databases. "Active customer" points to the exact SQL query that calculates it. AI assists the mapping by analyzing the column names, SQL comments and existing formulas to suggest the correspondences. No more asking "which table is that?" — the glossary says it.
Automatic conflict detection
AI compares the definitions between departments by analyzing the SQL queries used by each team to calculate the same metrics. It identifies the gaps: "marketing calculates revenue with the orders table (status = completed), accounting with the invoices table (type = invoice)". It quantifies the impact of the gap and generates a reconciliation report.
Implementation
Building the business glossary happens in three phases over 6 to 10 weeks.
Inventory of critical terms (weeks 1-2)
Identify the 30 to 50 most used business terms in your company: "customer", "revenue", "margin", "active product", "validated order", etc. For each term, ask 2 to 3 departments: "how do you define this term?". Document the current definitions and the gaps. Prioritize the terms that cause the most confusion.
Definition workshops (weeks 3-6)
Organize a 1h30 workshop per domain (customers, products, finance, HR). Bring the relevant Data Owners together. For each term, present the divergent definitions and their quantified impacts. Converge toward a single definition. Document: name, definition, formula, Owner, source systems. Validate with the executive sponsor for strategic terms.
Publication and adoption (weeks 7-10)
Publish the glossary in a tool accessible to everyone (Notion, Confluence, DataHub). Create an update process: any new definition or change goes through the Data Owner. Integrate the glossary into the AI projects: each variable of a model must point to a glossary entry. Train the teams and establish a quarterly review.
Results
Here are the results observed at our clients after setting up the business glossary.
Frequently asked questions
What is a business glossary?
A business glossary is a shared dictionary that defines the company's business terms in a unique and unambiguous way: "active customer", "net revenue", "product in stock", "qualified lead". For each term, it specifies the definition, the calculation formula, the responsible Data Owner, the source systems and the associated quality rules. It's the cornerstone of data governance.
Why do you need a glossary before launching an AI project?
Because an AI model trained on data whose definitions vary between sources will produce inconsistent results. If "active customer" means "purchase within 12 months" for the CRM and "non-closed account" for the ERP, the model mixes two different populations. The predictions are biased and the decisions based on those predictions are wrong. The glossary aligns everyone before AI steps in.
How do you handle definition conflicts between departments?
Conflicts are normal and frequent. The method: bring the relevant Data Owners together (e.g. sales director + finance director for the definition of "customer"), present each department's current definitions, identify the gaps and their impacts (e.g. a 2,500 "customer" difference between the two definitions). The lead Data Owner decides. If the conflict persists, the executive sponsor (CEO or CFO) arbitrates. Document the decision and the history.
Do you need a dedicated tool to manage the business glossary?
To start, a collaborative spreadsheet (Google Sheets, Notion, Confluence) is enough to document the first 30 to 50 terms. A dedicated tool (Atlan, Alation, DataHub with glossary) becomes necessary when the glossary exceeds 100 terms, when several teams contribute simultaneously, or when you want to link the business terms to the technical columns (data catalog). The rule: start simple, add tooling when it hurts.
For tech profiles
Technical implementation of the glossary
Metrics-as-code
dbt lets you define business metrics directly in the SQL code: semantic models, dimensions and measures. Each metric has a unique definition, versioned in Git, with its exact SQL formula. BI tools (Tableau, Looker, Power BI via the Semantic Layer API) consume these definitions instead of recalculating locally. No more divergences.
Glossary connected to the catalog
DataHub (open source) offers a Business Glossary module integrated into the data catalog. Each term is linked to the corresponding datasets, columns and dashboards. Full-text search, tags, term hierarchy and validation workflow (draft → review → approved). Deploys via Docker in 2 hours.
Recommended tools
Comparison of approaches
| Criterion | dbt Semantic Layer + DataHub | Shared spreadsheet | Atlan / Alation |
|---|---|---|---|
| Cost | Free (open source) | Near zero | 2,000-10,000 €/month |
| Link to technical data | Native (columns + SQL) | Manual | Native |
| Versioning | Native Git (dbt) | File history | Built-in |
| Non-tech adoption | Average (DataHub interface) | Excellent | Excellent (premium UX) |