A meeting that happens is not a meeting that leads anywhere. Every appointment system in insurance distribution records the first one the same way: requested, held, closed. The record is accurate. It is also the reason nobody was working on the largest loss in the customer journey, because a loss that closes cleanly does not look like a loss to anything that reads the calendar.
An international health insurance group asked what a standing customer intelligence function would be worth before it agreed to fund one. That question has no honest answer in the form of a business case. It has one in the form of a process map from somewhere the work has already been done, so we brought one: a mined event log from a completed engagement in a retail lending business, where the customer's online activity and the sales force's own system-of-record clicks had been joined into a single case history and replayed as a graph.
The challenge
The raw material was never the constraint, in the lending business or in the group. The log was already being written and paid for. In the joined estate, twenty-nine distinct screens in the sales system carried enough traffic to be worth mapping, and the busiest one alone had been viewed 1,421,228 times. Nobody instrumented anything new. The event stream was the exhaust of software that had been running for years, and reading it cost less than the three instruments an operational review normally starts from: an assumption, a sample and an interview.
What the replay showed at the appointment stage is the finding this piece exists to carry. Four appointment steps sit in sequence. The first carries 103 cases. The second carries 41. The third carries 12. The fourth carries 2. That collapse is the number people reach for, and it is the weaker of the two available.
The sharper number is the arc leaving the first step. Seventy-one of the 103 first appointments go directly to an end state. Roughly sixty-nine in every hundred. Not rescheduled, not recorded as lost to a competitor, not marked declined with a reason attached. Ended. The step-to-step flows among the few cases that do continue are single digits all the way down.
None of those seventy-one endings was logged as a failure. Each one was a meeting requested, a meeting held and a meeting closed, and on every measure the meeting itself owned, the meeting succeeded. The path failed. No function owned the path, so no function was accountable for it, so it never appeared on anyone's improvement list. A drop-off invisible to the operational reporting is not one anyone is working on, and this one had been invisible for as long as the process existed.
The approach
The work opened on a hypothesis blunt enough to come back false, about whether the sales force actually delivered the cross-channel experience the organisation had promised, and it was written with guard rails on both sides. Throughput time down and customer effort down, while conversion and satisfaction hold. That single conjunction is what separates an efficiency exercise from one that quietly destroys value.
The appointment finding came out of rendering the same process twice. Once by frequency, which produces the counts above. Once by duration, which produces a different graph of the same events. Step durations average 6.2 days at the first appointment, 15.6 at the second, 14 at the third, and 57.9 hours at the fourth. Read the gaps between the steps instead and the picture inverts: 4.1 days, then 22.4, then 37.1, then 43.9. The final step carries 57.9 hours of actual work sitting behind a wait of nearly forty-four days.
That contrast is the part most automation proposals get backwards. They target the work. The work is a rounding error against the queue in front of it. Before anyone buys a system to make a task faster, the measurement that decides whether the money is well spent is how long the case sits between tasks, and that measurement is already in the log.
Benchmarking did the work an external target usually pretends to do. One hundred and forty-one comparable units, all running the identical process on the identical systems for the identical product, ranked on average throughput time. The slowest averaged around twenty-six days. The fastest averaged around one. Fifty-two of the 141 sat above the ten-day line drawn across the chart. Best practice did not have to be imported or invented; it was already being executed inside the same organisation, and the distance between the distribution and its own best performer is a benefit estimate with a source attached rather than a guess.
What converts a study into an instrument is the cadence: re-run the mining, publish the ranking back to the units, and hold the performance conversation against the number rather than a recollection. In that lending programme, conversion improved by an average of 15 percent. That figure belongs to that engagement. It is evidence for what the method does, not a result the health insurance group has banked.
The outcome
What was delivered here was a proposition and a proof case. No capability was stood up in this phase and no number inside the group moved.
The proposition had three constraints in it, and the constraints are the interesting part. A standing function rather than a programme with an end date. Short-cycle projects as the unit of delivery rather than a platform build. And a payback floor: each project expected to return at least five times the effort invested, agreed before the work starts rather than argued about after it. The scorecard was closed to five measures, leads, cost per sale, conversion, retention and net promoter score, so that no project could invent its own success criterion once the results were in. Privacy protection sat inside that portfolio as a project with a KPI, competing for the same resource as the revenue work, rather than outside it as a veto.
Three things would be built differently now, and none of them changes the arithmetic above.
Data readiness is not the log's existence, it is the case key. Every organisation we work with can produce clicks, calls and appointments; far fewer can say that a given click and a given meeting belong to the same person on the same journey. That join is the whole engineering task, and it is also the thing that gives a decision a lineage trail later, which matters more now that European rules on AI are in force and the trail stops being optional.
Second, evaluation sets have to be written against the path rather than the step. A retrieval-augmented assistant that answers an adviser's question correctly, scored on answer quality alone, will happily reproduce the seventy-one. It succeeded locally. So will a carefully scoped autonomous experiment that chases dormant cases, unless the thing it is measured on is end-to-end completion rather than tasks closed. Our Platform work starts by fixing the unit of measurement before anything is deployed against it, because a system optimised on the wrong unit gets very good at the wrong thing.
Third, the loop belongs in production rather than in a slide cycle. The mining, the ranking and the conversation are an operational cadence, not an analysis, and the MLOps discipline that keeps a model honest is the same discipline that keeps a process map honest. Consult engagements now open with that cadence rather than closing with it.
The organisation was measuring the meeting because the meeting is what the calendar knows how to record. Sixty-nine percent of them were the end of the story, and the system had no field for that.