Three Sources, Three Truths

None of the three is wrong; each measures a different quantity, and the disagreement is the finding.

• The survey measures remembered satisfaction at the moment of a 90-minute-later prompt, among a 7% self-selected sample. It is biased toward people with a strong (often positive) recollection and toward visits that ended well. It is legitimate evidence about overall sentiment and almost useless as a queue-time instrument.
• The timestamps measure elapsed time for the population, with no selection bias. 4:12 median is genuinely fine; an 11:40 90th percentile is genuinely bad. These are not contradictory - they describe a distribution with a long right tail, and the tail is where abandonment, complaint, and attrition live.
• The intercepts measure what the tail feels like and why people leave before they are ever surveyed. A customer who abandons the queue typically never receives the post-visit survey, which mechanically removes the worst experiences from the survey population.

So: trust the operational data for magnitude and the qualitative data for mechanism; use the survey for trend and for relationship-level sentiment. The flat 14-month survey trend alongside 40 added tellers is consistent with a program that fixed the median (already fine) and left the tail untouched.

Explaining it up: do not say 'the survey is biased' to a COO - it reads as excuse-making. Say: the survey tells us how people felt about their visit; it cannot see people who left, because they don't get surveyed. That framing is verifiable, non-defensive, and points directly at the operational fix.

Objective: measure the invisible population and locate the tail in time and place - not to re-measure satisfaction.

1. Abandonment measurement (week 1-3). Instrument the queue: ticket issued vs. ticket served. Any ticket closed without service is an abandonment event with a timestamp and branch. This is the single highest-value data in the program, because it converts 'people say they leave' into a rate.

2. Time-of-day and branch segmentation of the existing timestamps (week 1-2). No new collection; just re-cut 14 months of data by hour, day, branch, and transaction type. Hypothesis to test: the 90th percentile is concentrated in lunch hours and in branches with a small number of complex transaction types (business deposits, estates, disputes) that block a teller window for 15+ minutes. If true, the fix is queue architecture, not headcount - which also explains why 40 tellers did nothing.

3. Exit intercepts targeted at the tail (week 2-6). 40-60 short intercepts, sampled only in the 10:30-14:00 window at the eight branches with the worst tail, including people who leave. Quota by walkout vs. served. Twelve interviews is enough to generate hypotheses, not to size anything.

4. One structured observation shift per worst branch (week 3-5). Count arrivals, service times per transaction type, and staff idle time. Cheap, and it exposes whether the constraint is staffing, scheduling, or transaction mix.

What would change my mind: if abandonment is under 2% of tickets and the tail is evenly spread across the day and across branches, then the queue is not the real problem and the intercept quotes are anecdote, not signal - and I would redirect the work to whatever the tail branches have in common instead.

Give a number that is defensible and honest about its object; refuse a number that implies a precision you don't have.

Give: the share of branch visits completed within a defined service standard, e.g., '78% of visits in Q3 were served within 5 minutes; 6% of visits exceeded 10 minutes.' It is computed from operational data, has no response bias, moves with real interventions, and is board-legible. Pair it with the abandonment rate once instrumented - even a rough first reading is more decision-useful than a survey percentage.

Also give: relationship sentiment (the survey) as a separate, clearly labeled line, described as 'how customers rate their visit overall' rather than as a wait-time measure.

Refuse: a single blended 'customer experience score' that merges survey and operational data into an invented index. It will be gamed, it cannot be decomposed when it moves, and it will be the number the board holds the COO to for years.

Refuse, more quietly: any statement that the survey proves waits are fine. The board packet should carry a one-line limitation note - 7% response, post-visit timing, abandonment not captured. That note protects the CX team's credibility later, when the abandonment data arrives and the picture shifts. Presenting the limitation yourself on Friday is far cheaper than being shown it in three months.

Sequencing note: tell the COO on Friday what the 60-day plan will replace this number with. Boards tolerate an imperfect metric that is improving; they do not tolerate a metric that quietly changes definition between quarters.