An adoption number is half of a fraction. The other half normally lives somewhere else, in a different report, owned by somebody who was not in the room when the first half was celebrated. On one lending landing page both halves were printed on the same screen, and nobody had put them in the same sentence.
The page belonged to a large European cooperative banking group. Down the middle it sold a self-service lending file: work through the modules and within thirty minutes you would know what your new loan and your monthly costs looked like. Under that promise sat a badge, written in the friendly register of a queue you were welcome to join. Sixty thousand people had already been through it. Over in the right rail, beneath a heading explaining why to bank here at all, was the other number. More than a million customers already held a loan with the group.
Sixty thousand against more than a million is roughly six percent. That division is ours, not the bank's, and the two populations are not the same population, which is precisely why the figure is worth saying out loud rather than badging one side of it. The badge counts people who opened a self-service file, most of them orienting rather than borrowing. The rail counts an installed book assembled over decades. Nothing obliges those two to be comparable. What the ratio does establish is an order of magnitude, and the order of magnitude is the finding. A channel a few percent of your base has ever touched is a pilot that shipped, not a channel.
The challenge
Read the rest of the page and the six percent stops being surprising. Three separate controls on it invited the customer to book an appointment: one in the right rail, one at the foot of the quick calculator, one directly beside the button that started a calculation. Three other entry points invited a calculation instead, at three different depths of commitment. Along the right, a servicing rail carried six links for people who already had a loan and wanted to view it, repay it, raise it, or get help with payments they were struggling to make.
So the page was doing three jobs at once. It was acquiring, it was servicing, and it was offering to book a human at every point where either of the first two got difficult. None of those is wrong on its own. Stacked, they mean self-service and assisted service were competing for the same click rather than sequencing into each other, and the customer had no way of knowing which door the organisation wanted them to use, because the organisation had not decided.
That stack also poisons the measurement. An acquisition funnel and a servicing rail sharing one clickstream produce a single traffic figure that answers no question anyone actually has. The review used this page as evidence for a finding about ordering, and the confound sat on the page the whole time: some of that traffic was existing borrowers doing maintenance, not prospects orienting. Nothing in the material separates them, and this piece does not pretend otherwise.
The approach
The instrument underneath all of this was ordinary and the discipline was not. Mobile records, online records and customer-relationship records were stitched into one case history per customer, and a process model was discovered from that log rather than drawn in a workshop. Three sources, one model, one identity across all of them.
That stitch is what produced the review's second named finding, and it is the one that reframes the ratio. Customers contacted the bank first and went online afterwards. Not orientation followed by an appointment, which was the assumption every digital roadmap of the period was written on, but contact followed by digital self-service.
No channel-level report could have produced that sentence. A channel silo always believes it is the front door, because every session it can see begins inside it. Only a log with one case identity spanning the phone, the branch and the web can order the touches, and ordering was the whole finding. The drill-down into a single member unit shows the same shape mined from the relationship records rather than asserted: a lead, an inbound call, an outbound call, an appointment, a quote, a purchase, with two distinct routes into the appointment, one running through a phone step and one going direct.
Put the ordering next to the ratio and the self-service file stops looking like an underperforming acquisition channel. It was a servicing surface for people already in the funnel, being measured as though it were the top of it, and being optimised for a cold visitor who mostly arrived by telephone. Six percent is not a marketing failure. It is what you get when the front door is somewhere the analytics cannot see.
The outcome
What the engagement delivered was findings, a discovered model, a mined per-unit funnel and a set of recommendations. The recommendation that mattered was structural rather than clever: stop commissioning studies and embed process indicators permanently, inside the reporting that already wrapped the straight-through processing chain, inside the comparison programme already running across the member units, and inside the team that ships the product. One precondition was printed in small type at the edge of a diagram, and it is the most honest line in the pack. None of the comparing works unless every unit records the same way in the same system. Data readiness before instrument, as ever. No before-and-after conversion figure exists in the material, and none is claimed here.
Three things would be done differently with what is available now.
Publish the fraction, not the badge. Sixty thousand is a number a launch review can applaud. Sixty thousand out of a defined eligible population, refreshed monthly, is a number that forces a decision, and the definition of eligible is the hard part rather than the counting.
Then make the ordering measurable rather than discoverable. The finding arrived as a one-off study because no durable case identity spanned the channels. Carrying a case key across the telephone system, the branch appointment book and the web session is unglamorous integration work with an MLOps and lineage discipline attached, and it converts the sharpest insight in the review into a standing measurement. That is the Platform work, and it is the part nobody demonstrates.
And the early language-model copilots now appearing in customer-facing lending do not move the ratio on their own. A copilot grounded on retrieval over the group's own product documentation, evaluated against a proper set of held-out questions before it is allowed near a customer, genuinely improves the thirty minutes for the people already inside those thirty minutes. It changes nothing about a page with seven doors, and it must sit inside a decision trail an auditor can reconstruct later, which the agreed European rules on AI will hold the group to when they bite. Our Consult work starts by measuring the fraction rather than the experience, because the fraction is what tells you whether an interesting pilot has a channel underneath it.
The badge was accurate. Sixty thousand people really had been through it, and every one of them was a real customer doing a real thing. The number simply had no denominator attached, and the denominator was printed on the same page.