Conversion improved by an average of 15 percent. That is the sentence a reader keeps, and it is the least useful line in the material it came from.
The useful part is what is missing from the invoice beside it. No new tracking went onto a website. No panel was recruited, no survey went out, no customer was asked anything new. Nothing changed about what the business observed or how it observed it. The entire input was behaviour the organisation had been recording, storing and paying to keep for years before anyone proposed reading it.
An international health insurance group was deciding whether to stand up a permanent customer insight function. The question it needed answered was not what such a function would produce, but what had to exist before one could produce anything. The assumed answer was a data programme: a platform, a warehouse, fresh capture across every channel, specialists hired against it, and insight arriving somewhere on the far side of two budget cycles. The proof case we brought inverted that sequence.
The challenge
The evidence came from a completed engagement in a retail lending business, where the method had already run from question to result. Its raw material was the screen log of the system the sales force used all day, and what came out of it was a graph: every screen a case had passed through, with the traffic on every route between them counted. The software had been writing that log as a side effect of doing its job, the log had been retained, and no part of it had ever been read as a description of what a customer was trying to get done.
That is half the finding. The other half is the half that keeps it honest, because behavioural data a business already collects is not the same thing as behavioural data a business can use. Four conditions had to hold before the log became evidence, and only one of them cost nothing.
Transaction logging had to be switched on across the channels, which is configuration of systems already in production rather than new observation of people. Time stamps had to be kept and treated as an asset, rather than overwritten by whichever process last touched the record. Online activity and sales-system activity had to be joined on one case identity, so that a click and a meeting three weeks later could be shown to belong to the same person on the same journey. Only then could the result be rendered as a graph of what happened, rather than a table of what was counted.
Three of those four are engineering, and the join is most of it. Nearly every organisation we work with can produce clicks, calls and appointments. Far fewer can prove that a given click and a given appointment sit on the same journey. That proof is the task, and calling it data readiness rather than data collection is not a quibble: it decides whether the exhaust is an asset or a cost line.
The approach
In the lending case the work opened on a hypothesis blunt enough to come back false, about whether the sales force 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.
Benchmarking did the work an imported target only pretends to do. One hundred and forty-one comparable units, all running the same process on the same systems for the same product, ranked on average throughput time. Fifty-two of them sat above the ten-day line the business had already set for itself, and the spread between the slowest and the fastest ran from roughly twenty-six days down to roughly one. The distance between that distribution and its own best performer is a benefit estimate with a source attached rather than an assertion.
The part that produced the 15 percent is not the analysis. The mining was re-run on a cadence, the ranking was published back to the units, and the performance conversation was held against the current number rather than against a recollection of last quarter. The engagement described the result as a sensor: something that counts continuously, in the way a pedometer counts steps, instead of a study that measures once and is filed. The lift is attached to that repetition. Any single mining run would have produced the same map and none of the improvement, because a map changes nobody's Tuesday and a published ranking does.
The economics of the sensor are what made a standing function arguable at all. The first project pays for the join. Everything after it rides on the same joined log at a marginal cost close to nothing, which is why the follow-on list ran from banner work and a user-experience function out into pricing, service and compliance. The capability's own technique inventory ran to eleven items, and the two that mined recorded behaviour rather than modelled it produced the only result anyone could point at.
The outcome
What was delivered to the health insurance group was a proposition and a proof case. No capability was stood up in this phase, no journey inside the group was mined, and no number inside the group moved. The 15 percent belongs to the lending engagement and is evidence for what the method does, not a result the group has banked.
The proposition carried three constraints. A standing function rather than a programme with an end date. Short-cycle projects as the unit of work rather than a platform build. And a payback floor: at least five times the effort invested, per project, agreed before the work starts rather than argued about once the results are in. 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 definition of success after the fact.
Two things would be built differently if the same case opened today, and neither changes the arithmetic above.
The first is where the current wave of work is pointing. Most of what we are asked for this quarter is retrieval: put the policy wording, the product rules and the correspondence into a vector store, sit a copilot beside the person handling the contact, let it draft and cite. That is good work and it lands quickly. It also grounds the assistant in what the organisation has written down, which is documents, and leaves out what the organisation has watched, which is behaviour. The joined case history described above is the second kind of grounding, and it is the one nobody is buying, because it looks like plumbing rather than intelligence. An assistant that knows the policy but not that this customer has been waiting since the first appointment is answering a different question from the one in front of it.
The second is that the cadence is an operating cost, not a project cost. A sensor that stops being refreshed stops being true, and the discipline that keeps a joined log current is the same MLOps discipline that keeps a model's features honest between releases. It is also what leaves a lineage trail behind a decision, which matters more now that European rules on AI have reached political agreement. Our Platform work starts there rather than ending there, and Consult engagements now open by asking what the organisation already records and what it would take to join it, because that answer sets the ceiling on everything else.
The group did not need a new way of seeing its customers. It needed to read what its own systems had been writing down about them the entire time.