Translate enterprise outcomes into data theses

A data thesis is not a technology project list. It is a falsifiable claim explaining how specified data changes an enterprise decision to produce a defined outcome. Start with the recurring decision under uncertainty: who makes it, which realistic alternatives exist, and what information would reverse the choice? Only then identify data and capability. A strong thesis names outcome, denominator, period, assumption and falsifying fact, while separating an early signal from the later business result.

A data thesis is a testable explanation connecting an enterprise outcome, a recurring decision and a defined data capability. It is neither a project list nor a promise that more data improves performance. Its validity depends on three links: information reaches the decision owner in time, changes the choice or its speed, and explains part of the outcome while guardrails remain within bounds. Governance approves purpose and risk boundary; management designs and runs the trial.

Build the thesis on a seven-field sheet: intended outcome, current decision, decision owner, alternatives, required data, hypothesis and falsifying fact. Put denominator and period beside every measure. Then trace the chain from source event to the information seen by the decision owner and mark where causality may break. Reject “improve data” unless the changed behavior is explicit.

Compare repair at the decision point, a bounded capability trial and a broader platform. If an incident lead cannot see service dependencies, a two-service map tests the cause quickly. If dependencies are known but ignored, workflow and training are the intervention. A broader platform becomes justified only when the trial shows a recurring problem, a transferable pattern and scaling cost lower than expected impact.

Median recovery falls from 95 to 74 minutes, an improvement of 21 minutes or about 22.1%. Reopenings rise from 9% to 14%, breaching the 10% limit. Do not combine the measures into a score that hides harm. The speed link is supported; closure quality is not. Revise verification and rerun the trial. The thesis is not accepted until both conditions hold.

Track three separate signals: input quality, decision change and business result. In the maintenance trial, accurate cause code is an input signal, dispatching a second crew is a decision and reopening is the outcome. If the first improves without the second, workflow is blocking use; if the decision changes without outcome improvement, the causal hypothesis may be wrong. Never collapse these signals into one score.

When moving the trial to a third site, do not copy its threshold alone. Compare asset mix, supervisor skill and emergency-order share because lower reopening may reflect easier work. The thesis record retains denominator, measurement period and screen version, and scale pauses when the channel changes before a new baseline exists.

Build a data thesis for a service with recurring outages. The enterprise outcome is lower customer impact, not production of a dashboard. The decision to improve is the incident lead’s choice of which component to isolate or restore. Inputs are affected services, dependencies, the latest change and component state; the capability makes those relationships visible with known freshness. State a falsifiable hypothesis: if dependencies for critical services appear within two minutes of incident opening, median time to identify the causal component will fall by at least 25%, without pushing reopenings above 10%. Select incidents with comparable severity and disclose exclusions; a quiet week is not a valid baseline for a month containing a major release. In the trial, expose the map for two services and retain the former workflow. Before each incident, the lead records an initial hypothesis; afterward, the lead records whether the map changed the choice. Identification falls from 38 to 25 minutes and recovery from 95 to 74, yet reopenings rise to 14%. Review shows that relationship direction was unclear, so responders isolated a dependent component instead of the source. The decision is neither declare success nor abandon the thesis. Correct relationship direction, add verification before closure and continue for ten bounded cases. Extend to two more services only if recovery remains below 75 minutes and reopening returns to 10% or less. Data creates value here by changing an observable professional choice within a guardrail, a competing explanation and a rollback option.

Before approving the thesis, ask someone who did not write it to identify the decision, alternative and falsifying fact from the sheet alone. Then test an excluded case, such as a safety failure requiring immediate dispatch regardless of cause code. A clear boundary proves the thesis guides a specific choice rather than becoming a general slogan for funding data. Implementation review: retain an unchanged comparison channel and adjust for asset type and failure severity. Recalculate after excluding safety orders outside the trial decision. If the effect disappears, revise the claim instead of scaling. If it persists, move the trial to one new site before enterprise adoption. The record also preserves every exclusion reason for replay.