Data & Analytics: Charter

Every function in this framework claims to be evidence-led. This is the function that decides what counts as evidence. Its authority comes from one rule: one authoritative number per concept. The moment two dashboards disagree about revenue, both stop being used and decisions revert to instinct.

In an AI-operated company the stakes rise, because agents consume these numbers at machine speed and without the human instinct that says that looks wrong.

What this function owns#

The metric dictionary. Each concept defined once, what it counts, what it excludes, which source it comes from, who owns it. A metric without a written exclusion rule is not defined.

Lineage. For any published figure, the path back to the raw record. If the path cannot be walked, the figure is a rumour.

Data quality gates. Freshness, completeness, range and referential checks that run before anything is published, not after someone complains.

Access and privacy boundaries. Who and what may read which fields, jointly with Security and Legal.

The measurement of outcomes. Not activity. Counting what was done is easy and mostly useless; measuring what changed is the job.

What is NOT delegated to an agent#

• Defining a metric. Definitions encode judgement about what matters.
• Overriding a failed quality gate. A human explains why publishing stale data is acceptable today, in writing.
• Granting access to sensitive fields.
• Retiring a metric. Numbers people depend on do not disappear quietly.

KPIs#

AI agents in this function#

Quality-gate agent. Runs freshness, completeness and range checks on every dataset before publication and blocks the publish on failure. It checks values, not file age: a file rewritten on schedule with a stale payload inside passes an age check and fails a value check, and that distinction has caught real incidents.

Lineage agent. Traces a figure back to source on request and reports where the trail breaks.

Anomaly agent. Flags distribution shifts rather than threshold breaches, because the interesting failures rarely cross a round number.

Query assistant. Drafts queries against the modelled layer for other functions, with the generated SQL always shown. Read-only access, and never against raw production tables.

The pattern holds: agents block, trace and flag. A person decides what a number means and whether it is fit to publish.

SOPs#

• Definition before dashboard. No metric is visualised before it is defined in the dictionary.
• Publish gate. Quality checks run first; a failed gate blocks publication and raises an alert rather than degrading silently.
• Incident protocol for a wrong published number: withdraw, correct, state what was affected, record the cause. Correction is fast; the record is what stops repetition.
• Quarterly metric review. What is unused, what is disputed, what should be retired.
• Access review on the same cadence, with Security.

Templates#

SOP, Security Checklist, metric definition record, data quality report, access request.

Workflows#

In: raw sources from every system, question requests from all functions, privacy constraints from Legal, threat model from Security.

Out: the metric dictionary, governed datasets, scheduled reporting, answers to specific questions, with their lineage attached.

Handoffs: Finance for anything financial (their definitions win on financial concepts), AI Strategy & Transformation for baselines and benefit measurement, AI Engineering & Product for features that consume data at runtime.

The loop that protects everything downstream: define → gate → publish → review. Skip the gate once and the next wrong number is discovered by a customer.

FAQ#

Why insist on one number per concept?#

Because reconciliation is a tax paid forever. Two definitions of "active customer" means every meeting begins by arguing about which is right, and the argument is never resolved, it is simply re-fought.

Is freshness enough to trust a dataset?#

No, and assuming so is a specific and common failure. A pipeline can run on time, exit cleanly and write a file whose contents are weeks old. Check the values, and check the maximum date inside the payload, not the timestamp on the file.

How much data should be kept?#

As little as answers the questions actually asked. Every retained field is storage cost, privacy exposure and breach surface. The unused-dataset metric exists to make that cost visible.

What changes when agents consume these numbers?#

Error propagation speed. A human reading a wrong figure often pauses; an agent acts on it and produces ten downstream artefacts before anyone notices. That is why the gate blocks publication rather than warning after it.

What else is coming for Data & Analytics