Profile the extract before looking for patterns

Understand the method

Exploration starts by finding out what the extract can support. A chart can be technically correct yet misleading if the source contains repeated tickets, mixed time zones or a denominator that silently drops unresolved cases. Begin with a small profile that describes the unit, scope, completeness and validity of the records. This is not paperwork to complete after the interesting analysis; it is how you decide whether the apparent patterns deserve attention.

Carry the plan into the evidence

This module reuses Harbor Helpdesk's 24-ticket source from Module 1 without alteration. One row is one ticket, not a customer or a message. The creation cohorts are [1 June 2026 00:00, 8 June 00:00) and [8 June 00:00, 15 June 00:00), UTC. The fixed observation cutoff is 22 June 2026 00:00 UTC. All tickets have sufficient follow-up for a 72-hour outcome. The small extract is deliberately constructed teaching data, not a representative business sample.

Read the dictionary before importing. Ticket identifiers and categories are text. Event fields are UTC timestamps, and ratings are integers from one to five when observed. Some spreadsheet tools import ISO timestamps as text; verify parsing and timezone interpretation before subtracting cells. Do not assume a displayed date means the underlying value is a valid date. A calculator and the worked duration list are sufficient for this lesson; the optional Python source reproduces the calculations locally.

Write a short profile

Count 24 records and 24 unique nonblank ticket IDs. Reconcile 12 tickets to each cohort. There are 12 email tickets and 12 chat tickets overall, but those balanced totals conceal different weekly compositions. Check each response timestamp against creation and the cutoff. Check resolution only where recorded. Count missing values by field and cohort rather than relying on one overall completeness percentage.

The source has all 24 first responses, six blank first-resolution timestamps and eight blank satisfaction ratings. Under this teaching dictionary, a blank resolution means no first resolution recorded by the snapshot; mature unresolved tickets remain in the resolution denominator. A blank rating is unobserved satisfaction, not zero satisfaction. Neither condition authorizes deleting the whole ticket from all measures. Zero elapsed hours, if present in another extract, would be a numeric observation requiring a check for automated responses, not an empty field.

Pair every check with an action

Test unique keys, documented categories, parseable UTC timestamps, nonnegative elapsed times, events no later than the snapshot and sufficient outcome maturity. If a creation timestamp cannot be parsed, quarantine the row from cohort calculations and seek correction while retaining the original. If a key repeats, investigate whether records are copies or conflicting events; do not automatically keep the first one. The supplied source is prepared and passes these checks. Do not invent defects or a cleaning story to make the exercise look more substantial.

Record the file/version, check, observed result, analytical implication and next action in an exploration log. Your profile should end with a readiness decision: proceed with the bounded descriptive comparisons while retaining the documented blanks and limitations. It should not end with “the data is perfect.” Missing routing exposure, staffing and a longer baseline still limit interpretation, even when every supplied timestamp is internally consistent.