The least interesting object in most digital strategy packs is the screenshot of the company's own website. It goes in small, near the bottom, as proof that the channel exists. Turn it round and it is the only instrument on the page, because it is the one place in the journey where a structured record of a customer comes into being before anybody at the company has spoken to them.
An international health insurance group was weighing whether to run customer insight as a standing function rather than as a sequence of one-off studies. The case put to it did not open with a platform diagram. It opened with a screen from a consumer lending journey, on which a couple who have started looking at houses answer three questions, what they can borrow, what they already have, and how much they need, and then fill in seven areas of their own finances: income, debts and credit, the home they live in now, savings, the loan they already carry, the home they want, and the work it will need. At the bottom of that screen sits one action, which books an appointment and shares the completed file with an adviser. That action, not the calculator above it, is the subject of this piece.
The challenge
The question the group wanted answered was not whether it had digital channels, but whether the people who actually meet customers deliver the experience the channel strategy assumes, and whether the two halves are joined at all. A review built on assumptions, samples and interviews returns an answer with the fidelity of its inputs: the assumptions are the analyst's, the sample is small, and an interview records what somebody remembers doing rather than what they did.
The objective in the worked example was written as a constrained one, which is rarer than it should be. Throughput time down, and the effort the customer has to spend down, while conversion and satisfaction hold. Two numbers to push, two numbers that must not move. Most efficiency programmes write only the first half of that sentence.
There is a structural reason the answer is hard to obtain. The journey in the worked example, as it was drawn before anything was mined, runs across eight phases, beginning at a life event long before the company enters the customer's mind and ending when that customer recommends it to someone else. Eleven named steps sit on that spine. Three functions own stretches of it, marketing at the front, sales in the middle, and the team that services the relationship once the sale is done, and three channels carry it: online, the call centre, and the office where the appointment takes place. The value being chased rotates as the journey moves, volume early, satisfaction in the middle, margin late, and it rolls up to four business outcomes rather than one. No single owner sees the whole line, so no single owner's reporting can be the answer.
The approach
The intake screen is what makes the line observable. The customer does the data entry, in their own time, and the structure of that form becomes the schema of everything downstream. Then the share action supplies the join key. On one side of it sits a web event stream. On the other sits the record system the adviser works in. Neither needed a new system. Transaction logging had to be switched on and the two sides linked to one case identity, but both were already writing time stamps against every interaction. The time stamp is the asset, and it was already being paid for.
Mined together, the record-system side of that worked example produced a process map of twenty-nine screens with observed frequencies attached. The busiest single screen had been viewed 1,421,228 times, and 848,617 of those visits arrived directly from the start of a session, which makes it a front door rather than a step. The heaviest internal move between two screens carried 544,241 cases. Nothing here was built for the study, only switched on and read by somebody willing to treat the log as evidence.
Zoom in on the appointment process the shared file feeds, and the numbers turn small and uncomfortable. There were 103 first appointments, 41 second, 12 third, 2 fourth. The collapse from 103 to 2 is not the finding. The finding is that 71 of the 103 ended immediately after the first meeting, and every one of those was recorded as a meeting held. The step succeeded. The journey stopped. Anything measured at the level of the step would have reported a working process.
The same process rendered in time rather than in frequency is a different picture, which is why it is rendered twice. The first step averages 6.2 days, the second 15.6, the third 14, the fourth 57.9 hours. The waits between steps run 4.1 days, then 22.4, then 37.1, then 43.9. Roughly two and a half days of work sit behind six weeks of waiting on that last hop. Automation aimed at the work would be aimed at the rounding error, while the file the customer filled in sits in a queue.
The outcome
Once the same event log is aggregated by organisational unit, the benefit case stops being an estimate. In the worked example, 141 comparable operating units were ranked on three mined measures: steps per product, throughput time per product, and conversion. The best unit averaged about one day of throughput and the worst about twenty-six, on the same process and the same systems. Fifty-two of the 141 sat above the ten-day line drawn on the chart. The size of the prize is the distance between that distribution and a performer already inside the building, which is more defensible than any external benchmark.
Repeating the exercise on a cadence is what turns a study into an instrument. The source records an average conversion improvement of fifteen percent from running that loop, quoted with no baseline attached, so it belongs here as a reported figure rather than a verified one.
What the insurance group received in this phase was a case, a method and a worked example from another industry, plus a recommendation to run insight as a standing capability with a payback test on every project it takes on. Nothing described above was built at the group. The measurements come from the separate lending engagement, and the money on the intake screenshot is indicative calculator output, a demonstration of the widget rather than a result, which is why it is not reproduced here.
The pack reads differently now because of what has arrived since. An adviser copilot that can retrieve the customer's own file, the history of the case and what was said at the last appointment is a different product from one that retrieves a policy document, and the difference is not the model. It is whether a structured file exists at all, and whether the two event streams were joined when it was handed over. Retrieval-augmented answers inherit the quality of what an intake form collected years earlier, which makes this a data readiness and lineage question settled long before anyone shortlists a vendor.
The mined log also supplies what a pilot usually lacks, a set of labelled failures. The 71 cases that stopped at the first meeting are the beginning of an evaluation set, because the outcome is known and nobody had to be interviewed about it. With European rules for AI now agreed, the trail showing which record produced which recommendation will eventually be something an organisation is asked to produce. And early, carefully scoped autonomous experiments belong on the waits, the 22.4 days and the 43.9 days, where the case sits still and nobody owns the clock, rather than on the 57.9 hours where a person is doing work they understand.
Our Consult engagements start where this pack did, at the screen where the customer does the typing, because the context an assistant works with is not assembled at deployment. It is assembled by the intake somebody designed years earlier, and by whether anyone joined it to what happened next.