Ticket counts and pipeline uptime are the metrics data leaders show executives because they are easy to graph. Neither one tells you whether the data function is producing decisions, trust, or compounding leverage. Plumbing-level KPIs answer plumbing-level questions; they do not answer “is this team worth what it costs?”
Below are five KPIs that do. Each one points at a failure mode that quietly erodes the function’s standing — and each one is measurable today, with infrastructure most teams already have.
1. Decision latency
The wall-clock time from “stakeholder asks a decision-relevant question” to “decision is made with data in hand.” Not delivery time. Decision time. A dashboard shipped in three hours that sits unread for two weeks has a decision latency of two weeks.
How to measure it. Tag inbound requests in your intake tool with a decision_by field. When the request closes, capture the date the decision was actually made (a one-line confirmation from the requester is enough). Track the median and the 90th percentile separately — the long tail is where political damage accumulates.
What good looks like. Median under five business days for tactical asks; under three weeks for strategic ones. If your median is four hours but your P90 is forty days, you have a triage problem masquerading as a velocity story.
2. Dashboard-trust score
A monthly survey-derived metric: for each top-20 dashboard by usage, ask the primary audience “would you make a six-figure decision based on this dashboard alone?” on a 1-5 scale. Average the scores, weighted by dashboard usage.
How to measure it. Pull dashboard view counts from your BI tool. Send a one-question Slack survey to the top three viewers of each top-20 dashboard, monthly. Five minutes of stakeholder time per month, in aggregate. Score below 3.5 means the dashboard is decoration; below 2.5 means it is actively eroding trust in the function.
What good looks like. Weighted average above 4.0, with no top-10 dashboard below 3.5. The point is not the absolute number — it is the trend. A trust score sliding from 4.2 to 3.6 over two quarters is a louder signal than any incident report.
3. Model time-to-production
For analytics-engineering and ML work alike: median elapsed days from “first commit on the model” to “model serving its first production query or prediction.” Includes review, QA, deploy, and the oft-forgotten interval where the model is built but nobody has wired it into a downstream consumer.
How to measure it. Git timestamps for the start; deployment-tool or warehouse-query logs for the end. Most warehouses log first non-developer query against a table — that is your end timestamp. Exclude models that never reach production from the median, but track the rate separately as a “shelf rate.”
What good looks like. Under 10 business days for analytics models; under 30 for ML models. Shelf rate under 15%. Higher shelf rates indicate the team is building speculatively rather than against confirmed demand — usually a symptom of weak intake discipline, not a technical problem.
4. Data-incident MTTR
Mean time to resolution for incidents that materially affect a downstream consumer — a broken executive dashboard, a stalled reverse-ETL sync, a model serving stale predictions. Distinct from your warehouse-uptime SLA, which captures none of the cases that actually hurt.
How to measure it. Define “material” up front: any incident a consumer noticed before you did, plus any incident affecting a top-20 dashboard or a revenue-touching pipeline. Track from detection (whoever noticed first) to consumer-confirmed resolution. Do not stop the clock at “the fix is deployed” — stop it when the affected user says the data is correct again.
What good looks like. P50 under four hours, P90 under one business day. The P90 is what gets your function fired. Pair this metric with detection source: every incident a stakeholder caught before you did is a gap in observability you owe yourself an explanation for.
5. Data-asset usage skew
The Gini coefficient of usage across your active data assets — tables, models, dashboards. A perfectly even distribution scores 0; a distribution where one asset accounts for everything scores 1. Most mature data platforms cluster between 0.6 and 0.8.
How to measure it. Pull 30-day query counts per table from your warehouse logs and 30-day view counts per dashboard from your BI tool. Compute the Gini in a sheet — twenty rows of arithmetic, no library required. Recompute monthly.
What good looks like. Stable or slowly rising over time, in the 0.65-0.80 band. A skew below 0.5 usually means you have unmaintained sprawl pulling the average down — a long tail of dashboards and tables nobody uses but everyone is afraid to deprecate. A skew above 0.85 usually means concentration risk: one or two assets are doing all the work, and an outage in either is a function-wide outage.
None of these five replace your existing operational metrics — they sit on top of them. Pipeline uptime keeps the warehouse running. These five tell you whether the warehouse is earning its keep. Pick the two that hurt most to look at, instrument them this quarter, and report them to your executive sponsor monthly. The conversation about your team’s value gets concrete the moment the numbers do.