Skip to main content

Measuring the return

Project sponsorsLeadershipBoard members
In one minute

The single most important number is the cost of answering the next question. If it is falling, the platform is working. If it is flat, something is wrong with the model, not with the technology.

Everything else on this page supports that one measure.

Take the baseline first

It takes an afternoon, it is the most commonly skipped step, and without it a successful project cannot prove it succeeded.

Before anything is installed, record honestly:

Baseline metricHow to capture it
Time to answer the chosen questionAsk the person who would answer it today. Include waiting time, not just working time
People involvedHow many distinct people must be available
Systems touchedHow many separate systems someone must access
Reporting cycle effortPerson-days per cycle for the relevant report
ReproducibilityWould two competent people get the same number? Yes / roughly / no
TimelinessTime from a condition occurring to the person who can act knowing about it. Push delivery is measured against exactly this number
Questions declinedHow often something goes unasked because it is too expensive. Ask the team; they will know
Mean time to diagnosisFor incidents in the area you are covering

The last two are the ones people are surprised by. "Questions declined" is often the largest hidden cost in the organisation, and it is only ever visible by asking.

At 90 days

The first project should produce evidence on four axes.

Time
To answer the question
Compare directly against the baseline. Days to minutes is common for investigation-type questions
People
Needed to answer it
Often the more striking number, from four specialists to one engineer
Repeat
Can it be re-run?
A yes here is worth more than a faster first answer
Reuse
Model reused for question two
The leading indicator of compounding value

That last one deserves emphasis. At 90 days you cannot yet show a large financial return, be honest about that. What you can show is whether the second question was cheaper than the first. If it was, the compounding thesis holds and the investment case for continuing is sound. If it was not, find out why before scaling: usually the model was built too narrowly around question one, or it was built by people who do not know the operation.

At one year

By a year in, the measures become financial and operational.

Recurring time saved

  • Reporting cycle preparation, in person-days per cycle × cycles per year
  • Investigation time, in hours per incident × incidents in the covered area
  • Ad-hoc question turnaround, across everyone who asks them

This is the most defensible number because it is directly observed and it recurs.

Operational outcomes

Only claim these where the platform is genuinely in the causal path. The same observed-not-asserted discipline is how agent work gets measured too, a warning acted on against one dismissed, a draft accepted against one rewritten. The signals that steer agents →

  • Downtime avoided, where a monitored condition produced a warning that was acted on. Count the specific instances; do not extrapolate.
  • Output recovered, where faster diagnosis shortened a stoppage.
  • Maintenance effort redirected, work no longer done on a calendar because condition data said it was not needed.

Risk reduction

Harder to quantify, appropriate to state qualitatively for a board:

  • Reported figures now reproducible: two people get the same number. Full traceability with data-quality flags arrives with lineage, which is planned
  • Reduced dependence on individuals for critical knowledge
  • Audit responses answerable from evidence rather than reconstruction

The compounding measure

Plot the effort for each question answered, in order. The shape of that line is the whole argument.

Falling
Healthy
Each question reuses more model than it adds
Flat
Investigate
Model too narrow, or built around one department's view
Rising
Stop and review
Usually means source integration, not modelling, is the bottleneck

A worked example

Numbers invented for shape, not for accuracy. The point is which lines are defensible.

An engineering team baselines one recurring question, "which assets contributed to last month's availability shortfall", before starting.

LineBeforeAfterBasis
Time to answer3 weeks elapsed, ~5 person-daysUnder a day, ~0.5 person-daysObserved, both times
People needed4, across 3 departments1Observed
Times asked per year1212Unchanged
Recurring saving~54 person-days a year4.5 days saved × 12
Questions previously declined~6 a year, by the team's own estimateNow attemptedStated qualitatively, not totalled
Incidents in scope20 a year, ~6 hours to diagnose~2 hoursObserved on 5 incidents so far, extrapolated to 20 with that caveat stated

What makes this credible is not the size of the number. It is that every line says where it came from, one line is deliberately excluded from the total, and one is marked as extrapolated from a small sample. A board that sees that will believe the rest.

A single unmarked extrapolation discredits an entire analysis. Mark them.

Leading and lagging indicators

Most of what is easy to measure arrives too late to steer by. Watch both.

Leading, watch monthlyLagging, review at 6 and 12 months
Effort to answer question N versus N-1Person-days saved per reporting cycle
How many people can answer without a specialistMean time to diagnosis
Number of trials or investigations attemptedDowntime avoided, counted by instance
Model reuse: what the last question added versus reusedMaintenance work not done because condition data said so

If a leading indicator goes flat, you have roughly a quarter to act before it shows up in the lagging ones. That is the whole reason to track them separately.

What not to claim

Credibility is the scarce resource in a data programme. Three claims that destroy it:

  • Attributing broad improvements to the platform. If site availability rose 2% and DataHub was one of six things that changed, say so. Overclaiming once makes every subsequent number suspect.
  • Extrapolating from one avoided failure. One caught bearing failure is worth reporting as one caught bearing failure. Multiplying it by the fleet is how a real result becomes an unbelievable one.
  • Counting theoretical savings from questions nobody asked. Real, and unprovable. Mention it qualitatively; do not put it in the total.

Costs to track against

For a complete picture, track:

  • Infrastructure, the servers and storage the platform runs on. Predictable, and it scales with data volume, which above all means with sampling rate, the one dial worth setting deliberately. Choosing a sample rate → · Data lifecycle →
  • People, the domain expert and steward time modelling, and the administrator time running it. This is the dominant cost.
  • Integration, per source system, and highly variable. The single biggest driver of first-project duration.
  • Licence, none. The platform is open source under the AGPL-3.0.

Note what dominates: this is a people cost, not a software cost. That is worth being explicit about with a board, because it sets the right expectation about where attention should go.

Go deeper