Preserve and type the source.
Parse dates and numeric durations, normalise blanks and labels, derive date/hour fields and preserve records requiring review rather than silently deleting them.
Power BI case study · Service operations
A busy contact centre needs a clearer route to action.
I turned 197,278 interaction records into a Power BI monitoring model that connects call routing, abandonment and customer feedback to explicit definitions, quality checks and operational priorities.
Inbound decision mapTotals reconciled
Follow the route before reading the rate.
AI endpoints and known human routes have different operational meanings and different denominators.
Human attention combines 66,075 agent terminal outcomes and 2,290 post-contact abandonments. An AI endpoint does not establish successful resolution.
Interactions analysed197,278
Inbound, outbound and chat records preserved.
Initial abandonment31.61%
Of known inbound human routes, before human attention.
NPS · valid scores 0–1034.28
7,708 responses; zero policy shown below.
Satisfaction · scale 1–109.46
7.50% coverage of human attention.
Operational evidence
Volume, access and waiting answer different questions.
Compare recorded volume, initial abandonment and the 90th percentile of waiting time on known inbound human routes. Period labels are generic in this public version; the final period is partial.
Use the metric buttons, then select a period. Volume counts reflect the length of the captured period. Rates are descriptive observations. Waiting time covers known inbound human routes, including zero waits and excluding AI endpoints; seconds are provisional pending source confirmation. P90 is a distribution measure rather than a service-level target.
View all performance values as a table
| Period | All interactions | Inbound calls | Initial abandonment | Wait P90 | NPS |
|---|---|---|---|---|---|
| Period 1 | 26,153 | 17,965 | 37.94% | 146.1 s | 42.75 |
| Period 2 | 30,865 | 20,234 | 35.16% | 152.0 s | 42.62 |
| Period 3 | 44,436 | 32,827 | 46.09% | 164.3 s | 16.91 |
| Period 4 | 41,990 | 29,426 | 24.92% | 164.8 s | 26.18 |
| Period 5 | 30,250 | 19,066 | 11.58% | 140.6 s | 42.79 |
| Period 6 · Partial period | 23,584 | 14,945 | 17.83% | 156.2 s | 42.93 |
| All periods | 197,278 | 134,463 | 31.61% | 157.9 s | 34.28 |
Inspect the inbound outcome reconciliation
| Recorded outcome | Calls | Share of inbound |
|---|---|---|
| Ends with AI | 34,492 | 25.65% |
| Agent terminal outcome | 66,075 | 49.14% |
| Initial abandonment | 31,602 | 23.50% |
| Abandonment after human contact | 2,290 | 1.70% |
| Unclassified | 4 | 0.00% |
Customer feedback
A strong score still needs a trusted response base.
The recorded zero scores materially change NPS. I retained them in the main 0–10 calculation and built a separate exclusion scenario so this definition decision stays visible.
Inspect the zero-score assumption.
Promoters score 9–10, passives 7–8 and detractors 0–6. NPS equals the promoter share minus the detractor share, multiplied by 100.
- NPS · including zeros
- 34.28
- Valid responses
- 7,708
All periods: 809 zero scores retained. NPS survey coverage is 5.73% of inbound calls.
Same responses. Two definitions.
Both policies remain visible on the full −100 to +100 NPS scale. Changing the highlighted series changes the calculation, not the observed customer experience.
Include zero scoresExclude zero scores
34.28 → 50.02 is sensitivity to the zero policy. It is not a measured improvement. Satisfaction is calculated separately on the documented 1–10 scale; recorded zeros require validation, and the 7.50% response coverage limits representativeness.
View both NPS calculations and their response bases
| Period | NPS with zeros | Response base | NPS without zeros | Response base | Zero scores |
|---|---|---|---|---|---|
| All periods | 34.28 | 7,708 | 50.02 | 6,899 | 809 |
| Period 1 | 42.75 | 973 | 58.56 | 876 | 97 |
| Period 2 | 42.62 | 1,213 | 57.99 | 1,095 | 118 |
| Period 3 | 16.91 | 1,443 | 39.54 | 1,209 | 234 |
| Period 4 | 26.18 | 1,696 | 41.82 | 1,509 | 187 |
| Period 5 | 42.79 | 1,421 | 55.00 | 1,309 | 112 |
| Period 6 · Partial period | 42.93 | 962 | 52.61 | 901 | 61 |
The original field does not establish whether a recorded zero is a submitted rating or a technical default. The main calculation follows the standard numerical scale while keeping this unresolved recording question explicit.
Method and data quality
Build the definitions before the dashboard.
I kept the source grain intact, separated channels and terminal outcomes, and designed a reusable semantic model so filtering changes the context without changing the meaning of a KPI.
One fact table, two dimensions.
CalendarOne row per date
InteractionsOne row per source record
DestinationsUnique endpoint mapping
Calendar and destination dimensions filter the fact table through one-to-many relationships. Initial routing team and terminal destination remain distinct.
The full implementation was opened, refreshed, saved and reopened in Power BI Desktop. The public page reproduces safe aggregates and definitions; it does not expose the contact-level model.
Separate routing from attribution.
Use unique destination mappings and a calendar dimension. Keep inbound operational rates separate from outbound and chat activity.
Protect each denominator.
Calculate human-route rates, inbound survey coverage and scale-valid feedback measures in the current filter context. Return meaningful zeros for empty counts.
Reconcile results in the native report.
Compare totals and outcome splits against independent calculations, inspect filter behaviour and check all four report pages in the native exported PDF.
Metric logicInspect the initial-abandonment definition
Initial abandonment rate =
Inbound calls abandoned before human attention
÷ Known inbound calls routed to humans
All-period calculation:
31,602 ÷ 99,967 = 31.61%
Known human routes =
66,075 agent terminal outcomes
+ 31,602 initial abandonments
+ 2,290 post-contact abandonments
AI terminal outcomes and 4 unclassified inbound calls
are outside this denominator.Quality controls flagged 242 inconsistencies between terminal classification and registered human time, four unmapped inbound destinations and one duration record for review. These checks identify investigation cases; they do not prove that the underlying outcomes are wrong.
Decision logic
Prioritise the service queue, then inspect the handoff.
The findings support a targeted investigation plan. They identify where to look; capacity, routing and process evidence are still needed to explain the causes.
Concentration
Support dominates initial abandonment.
Action
Support contributes about 95% of initial abandonments. Start with its arrival pattern, queue coverage and waiting-time distribution; test staffing and callback changes against a stable baseline.
Handoff signal
Sales needs a transfer-path review.
Action
Sales records a 12.2% post-contact abandonment rate among calls that had human attention. Inspect transfer destinations, ownership and handoff steps before attributing the pattern to agent performance.
Measurement
Resolve the feedback recording rules.
Action
Confirm whether zero scores are submitted answers or defaults, retain the sensitivity comparison and track survey coverage alongside every score. The satisfaction mean alone cannot describe all callers.
Compare broad teams with the denominator visible
| Team | Known human routes | Initial abandonments | Initial abandonment rate | Post-contact abandonment rate |
|---|---|---|---|---|
| Customer support | 86,374 | 30,017 | 34.75% | 2.15% |
| Sales | 9,628 | 764 | 7.94% | 12.20% |
| Network operations | 3,742 | 692 | 18.49% | 0.00% |
| Other | 223 | 129 | 57.85% | 0.00% |
Monitoring sequenceSignal → investigation → measured change
BaselineDefinitions, volume and quality
InvestigationSupport access and Sales transfers
EvaluationControlled changes and comparable periods
No intervention was implemented in this case study. The recommendations are proposed next steps, with impact to be measured after implementation.
Scope and interpretation
Clear limits make the findings usable.
The report establishes a reproducible operational baseline. It keeps recording uncertainty, partial coverage and the distinction between endpoint and resolution inside the analysis.
A terminal classification describes where the interaction ended. Some AI terminal records also contain human time, so they are not labelled as independently resolved.
NPS coverage is 5.73% of inbound calls. Satisfaction coverage is 7.50% of human-attention outcomes; these are different response bases.
Raw period volumes are not equivalent monthly totals. Generic public labels retain sequence and partial-period status while omitting source dates.
This page and its downloads omit contact-level records, staff identifiers, original exports and internal case documents. Broad team and period aggregates preserve the decision logic.
Project evidence
Inspect the numbers and the decision memo.
The public package separates auditable aggregate outputs from the concise business narrative. Definitions and response bases travel with the results.
CSV · Public aggregate package
Download the chart data as JSON ↘Reconciled totals and operational comparisons.
Channel and outcome totals, generic period summaries, broad team comparisons and a metric dictionary.
PDF · Executive case summaryBusiness question, findings and proposed next steps.
Analytical method, operational priorities, feedback sensitivity and limitations in a concise document.
Contact
Need a clearer answer from complex data?
I combine analytical discipline with multilingual SEO and editorial judgement to turn evidence into decisions people can act on.