Your dashboard says qualified leads grew 30%. Did the definition change?

Direct answer: treat any change to the definition of a lead, opportunity, customer, conversion, revenue, or margin metric as a new version. Choose one of three defensible treatments: recalculate history under the new rule, run an overlap period that bridges old and new versions, or close the old series and start a new baseline. Do not splice different meanings into one trend or train AI across the boundary without an explicit marker. Faster reporting cannot repair a false series; it distributes the wrong conclusion faster.

This problem appears when demand generation changes a score threshold, sales redefines an accepted opportunity, the CRM adds a lifecycle stage, finance moves from gross to net revenue, or analytics registers a different key event. Each change may be reasonable. The error is presenting the before and after values as the same unit. Leadership can then celebrate growth created by a denominator, cut spend because of a taxonomy decline, or teach a model from incompatible labels.

Write the Metric Version Contract

• Decision — which budget, target, queue, forecast, campaign, or model consumes the metric?
• Name and meaning — what does the metric claim, and what does it not claim?
• Unit and grain — event, person, lead, account, opportunity, order, product, market, or period?
• Population — who enters, who is excluded, and how are unknown cases represented?
• Formula — numerator, denominator, aggregation, filters, joins, and zero or missing-value treatment.
• Source and evidence — system of record, fields, query, timestamps, currency, time zone, and maturation.
• Version and validity — identifier, effective date, rationale, approver, and review date.
• Historical bridge — recalculation, overlap, mapping table, or explicit series break.
• Consumers — dashboards, targets, automations, prompts, models, and exports affected.
• Owner and correction — who approves, monitors, communicates, reprocesses, and resolves exceptions?

The contract is small, but it must travel with the number. dbt documents a Semantic Layer that centralizes definitions for downstream applications, including AI tools. That can reduce competing formulas. It does not by itself preserve temporal meaning: a single current definition can erase historical context unless versions and effective dates remain explicit.

Recalculate, bridge, or break the series

• Recalculate — use when historical data can support the new rule reliably and the company accepts republishing prior numbers with a restatement record.
• Bridge — run both versions when full recalculation is impossible but an overlap period can explain the difference by segment and decision.
• Break — create a new baseline when the old data cannot support the new definition, the eligible population changed materially, or a mapping would imply false precision.

There is no honest fourth option called “explain it in the slide note.” A footnote does not stop a dashboard, agent, or model from treating a column as continuous. The version has to exist in the data, query, and distributed artifact.

Example: fewer MQLs, better evidence

A U.S. B2B software company defined an MQL as any contact with a score of at least 60. It later required ICP fit plus a demo request or confirmed buying signal. Monthly MQL volume fell from 1,200 to 840, while sales acceptance rose from 22% to 35%. “MQLs fell 30%” is mathematically correct and commercially incomplete. “Quality improved” is also not proof of value until mature pipeline and revenue are observed.

For four weeks, calculate MQL v1 and v2 over the same eligible population. Compare volume, coverage, acceptance, opportunities, value, delay, and unclassified cases. Keep v1 for targets or agreements still in force; run v2 in shadow before changing routing or media feedback. Then decide whether to restate history, maintain a bridge, or restart the baseline. The stage-evidence framework at https://makinai.co/insights/en/ai-opportunity-stage-evidence-crm-pipeline prevents opinion from becoming progress; the missingness contract at https://makinai.co/insights/en/blank-crm-fields-ai-missingness-freshness-contract keeps unknown data from becoming “not qualified.”

Build only the level the decision needs

• Level 1 — versioned catalog: definition, owner, query, validity, and consumers for the 10–20 metrics that change decisions.
• Level 2 — versioned queries and tests: formulas in source control, example data, reconciliation, and change alerts.
• Level 3 — semantic layer: shared definitions across BI, spreadsheets, reports, and agents when several consumers repeat the logic.
• Level 4 — metric service: API, access control, lineage, multiple markets, low latency, and continuous operations when reuse pays for complexity.

A governed registry may be sufficient for twelve metrics and one executive report. A semantic layer becomes defensible when several teams and applications need the same logic. Microsoft describes Power BI semantic models with preparation, relationships, calculations, refresh behavior, and access control. Google publishes exact meanings for Analytics dimensions and metrics. Neither tool selects the right commercial meaning or historical treatment for the company.

Six-week proof

• Week 1 — choose one recurring meeting and no more than twelve metrics; record decisions, formulas, sources, and current disputes.
• Week 2 — reconstruct two recent definition changes and identify every affected dashboard, target, model, and automation.
• Week 3 — implement contracts, version IDs, effective dates, tests, and reconciled examples.
• Week 4 — dual-run old and new versions; explain variance by rule, population, source, and outcome maturity.
• Week 5 — connect the versioned packet to an AI-assisted report in shadow mode; block incompatible comparisons and require the version in output.
• Week 6 — publish the bridge or break, notify consumers, observe one real decision, and choose whether to keep, expand, or stop.

Deliverables include the catalog, contracts, queries, validity table, bridge, restatement rules, tests, lineage, consumer register, runbook, communication, and backlog. A “standard dashboard” is not sufficient if the organization cannot reproduce why the number changed.

Where AI helps — and where it must abstain

AI can find duplicate formulas, compare documentation, summarize conflicts, map consumers, generate test cases, and block a narrative when the version is absent. It must not silently decide which definition is true, fabricate an unrecoverable history, turn correlation into causality, or rewrite approved targets. Google Analytics documents that marking an event as a key event affects reports from creation and does not alter historical data. Configuration has an effective date, and the non-retroactive boundary belongs in the analysis.

Metrics, costs, and owners

• Coverage — decisions and dashboards with an identified metric, version, owner, and source.
• Integrity — reconciliation variance, duplicates, incompatible denominators, unversioned values, and detected breaks.
• Change — time from approval to implementation, bridge, communication, and downstream correction.
• Decision — reopened meetings, rebaselined targets, reversed budgets, and actions blocked by uncertain meaning.
• AI — outputs citing the version, incompatible comparisons blocked, taxonomy errors, and performance evaluated within each version.
• Business — mature pipeline, revenue, or margin under stable definitions, without counting semantic change as lift.

Cost grows with sources, history, markets, currencies, time zones, grain, consumer count, reprocessing, access, and change frequency. Marketing owns the decision; RevOps owns funnel semantics; finance owns value; analytics owns formula and comparability; engineering operates data; source-system owners stand behind fields; agencies document changes; executives approve a rebaseline. NIST recommends documented metrics, benchmarks, uncertainty, and regular reassessment throughout operation.

Start with the reporting workflow in https://makinai.co/insights/en/ai-weekly-marketing-report-data-decisions, preserve campaign identity through https://makinai.co/insights/en/canonical-campaign-identity-media-analytics-crm-ai, and bring MAKINAI at https://makinai.co/services/en/data-content-intelligence-systems one meeting, twelve metrics, and two recent definition changes. We can scope the smallest versioned catalog and bridge before a larger semantic-layer investment.

The sources explain platform capabilities and behavior; they do not prove that a semantic layer improves results by itself. The Contract, ladder, example, and proof are MAKINAI editorial guidance to validate against the actual stack, finance policies, and business decisions.