Ask a large insurer where the customer journey begins and the answer arrives shaped like a funnel. It begins at a quote request, an inbound call, a click from a comparison site. Everything upstream of that click belongs to brand marketing, everything downstream of signature belongs to service, and the map on the wall covers the part in between. That part is not the journey. It is the section of it the organisation happens to be able to see.
An international health insurance group was working out whether to stand up a permanent customer intelligence capability rather than commission one more round of analysis. What went into that room was not a proposal document. It was a journey map drawn during a completed engagement in retail lending, and what made it worth showing had nothing to do with lending. It was where the map started.
The challenge
The first step on that map is a marriage. Two people get married. Then they start looking at houses. Then they find the one they want. Only at the fourth step does anything appear that the institution could possibly have logged: somebody opens the online borrowing calculator. Seven further steps run out behind that one, through the call back and the invitation to an information session, the completing of the file, the signature, the keys, the renewal, and finally the customer recommending the bank to a friend.
Eleven customer steps in all, laid across eight phases, with three delivery lanes running underneath every one of them: online, call centre, branch. Three of the eleven happen before the institution exists in the customer's mind at all. The last one happens long after the money has been lent. The funnel most organisations actually map covers something like the middle third of that, and it is the middle third precisely because that is where the systems already write records.
This matters more than it sounds, because a map is not a picture. It is a permission structure for what counts as a finding. A map that begins at the quote request cannot produce a finding about the weeks before it, and the analysis built on that map will faithfully report that the important moments are the ones inside the funnel. Start at the life event and different moments become eligible. The map used here also refused to carry a single measure of goodness across its span: volume mattered early, satisfaction in the middle, margin at the end, with those three drivers tied back to four business outcomes rather than rolled into one. A step is not judged by a number belonging to a different part of the journey.
Three functions owned pieces of that one journey and three channels carried it, so no single owner could see the whole thing.
The approach
A drawn map is still an assertion. The next move was to measure it, and the useful part is that nothing new had to be instrumented. The transaction and CRM systems were already writing the log. Four things were required: transaction logging switched on, time stamps treated as an asset rather than as metadata nobody reads, online and CRM events joined into one case identity, and the result rendered as a graph.
The scale of what was already sitting there is worth stating. Twenty-nine distinct CRM screens made up the mined process. The busiest single screen carried 1,421,228 observed views, of which 848,617 were sessions arriving straight into it at the start of a case, and 544,241 transitions ran into the screen holding application requirements. That is the substrate a journey argument stands on, and the organisation was already paying to store it.
Then the map was narrowed to one thing it had claimed mattered: the appointment. Four steps, mined twice. Read as frequency, the surviving path runs 103 cases, then 41, then 12, then 2. Read as duration, the same four steps take 6.2 days, 15.6 days, 14 days and 57.9 hours, with waits of 4.1, 22.4, 37.1 and 43.9 days sitting between them. The busy graph and the slow graph are not the same graph, which is why reading only one of them sends improvement effort to the wrong place. The last step holds 57.9 hours of work behind a 43.9-day wait, roughly eighteen times more queueing than working on that single hop.
The number that changed the conversation was a different one. Of the 103 first appointments, 71 went straight to the end of the case. Every one of those 71 was recorded as a meeting held. The step completed. The journey stopped. Nothing in the funnel-shaped view of the world had any way to represent that, because the appointment was a task and the task was completed.
The third move was comparison. The same process, normalised per product, was ranked across 141 comparable organisational units on throughput time, steps per product and conversion. The slowest averaged around 26 days, the fastest around 1 day, and 52 of the 141 sat above a 10-day line. Same institution, same systems, same customers, better than an order of magnitude between the ends of the distribution. Nobody had to estimate the size of the prize, because the best observed performer was already inside the building with its actual path in the log.
The outcome
At the bank, the loop was closed rather than left as a study: the mining was repeated on a cadence, the benchmark was published back to the units, and the conversation between manager and team was held against the number. The reported result of that programme was an average improvement in conversion ratio of 15 percent, stated as an average with no baseline attached to it on the record we hold, which is how it should be read.
For the international health insurance group, this was an evidence case and a recommendation, not a delivered outcome. Nothing was mined for them in this phase. What they were given was the map, the worked example from another sector, and an operating rule: a standing capability running short-cycle projects, each opened with a stated hypothesis and a written business case, each required to return at least five times the effort invested, and each closed with the benefit tracked explicitly rather than assumed. Whether they took it is a separate story.
The reason to set this out now is that agentic systems inherit exactly the blindness the funnel map had. An agent acts on the journey it can see, and the harness we scope for it is a map with the same property as the one on the wall. Point an agent loop at the product funnel and it will improve the product funnel: it will close cases at the first meeting, and its completion rate will rise while the business does not move. The place this bites hardest is the evaluation set. If every example in it starts at a quote request, graded autonomy is being graded against the wrong boundary, and a copilot that never sees the steps before the enquiry will confidently tell you nothing happened there.
Retrieval has the same shape. Grounding an assistant in policy documents and product pages gives it the institution's view of the journey. Grounding it in the event log gives it the customer's. Our Platform work starts with the second, because the log is the only artefact in the building that records what people did rather than what the process says they should have done. With the EU AI Act in force, that recording discipline pays twice over for a health insurer: the joined event trail that grounds a journey map is also what shows a supervisor, later, what actually happened, and it is what model risk management has been asking for on the pricing side for years already.
Start the map where the customer starts and the moments that turn out to matter are rarely the moments the funnel was watching.