The Algorithm That Sees Everything Except the Building It's Standing In

A technology company has spent seven figures on a machine learning platform designed to predict when customers are about to leave, while its senior engineering team quietly packs its collective desk drawers and heads for the exit. The algorithm is sophisticated, the dashboard is beautiful, and the people who understand why either of them works are currently updating their LinkedIn profiles.

A major supermarket chain once rolled out a state-of-the-art self-scanning app, promising to eliminate checkout friction forever. Sleek, fast, revolutionary. What they hadn't accounted for was that the app required a member of staff to override it roughly every four minutes — for reduced yellow-sticker items it couldn't recognise, for age verification it couldn't complete, for the inevitable moment it decided your entire trolley needed a random audit. The intervention rate was so high that the queues for the self-checkout help desk became longer than the original queues the app was designed to dissolve. The technology was perfectly engineered for a world in which humans behaved consistently and edge cases didn't exist. It was, in other words, engineered for a world that has never existed.

This is a precise and accurate description of what is happening inside a statistically uncomfortable number of technology organisations right now.

The Dashboard Sees Customers. It Does Not See People.

The business logic behind customer churn prediction is not foolish. Retaining a customer costs less than acquiring a new one. Silent departures are worse than noisy ones, because at least the noisy ones leave a forensic trail you can learn from. If you can read the behavioural signals — the slowing logins, the drifting engagement, the support tickets that take slightly too long — you can intervene before the customer makes a decision they can't be talked out of.

The problem is not the theory. The problem is what the theory requires to function.

Building a churn model that works in production — not in the demo, not in the slide deck, but in the real, messy, CRM-integrated world where the data has three different customer ID formats depending on which acquisition channel the customer came through — requires people with deep institutional knowledge. It requires the engineer who knows why certain fields are null for customers who joined before the 2022 platform migration. It requires the architect who understands that the mobile engagement metrics and the desktop engagement metrics are measuring subtly different behaviours and cannot simply be added together like you're tallying a shopping list.

That knowledge does not live in the documentation. The documentation is six months out of date. That knowledge lives in people. Specifically, it lives in the senior engineers who are currently leaving.

The Climate That Produces the Algorithm Also Produces the Attrition

The organisations most aggressively investing in AI-driven churn prediction are, with some frequency, the same organisations haemorrhaging the engineering talent those investments depend on. This is not coincidence. The same leadership culture that treats engineering talent as a cost centre to be optimised will, over time, optimise away the very people who make the technology viable.

Headcount freezes. Reorganisations without consultation. Engineering leadership quietly replaced by programme managers who have read the right books but never shipped anything. Technical debt perpetually deferred in favour of feature velocity. Performance cycles that reward visibility over craft. These are not isolated policy decisions. They are a climate. And senior engineers, who have options and read climates accurately, leave.

They do not leave loudly. They don't write the scorched-earth LinkedIn post on the way out. They work their notice with quiet professionalism, and then the working knowledge of why the data pipeline breaks every third Tuesday walks out of the building with them.

What remains is a team doing their absolute best with documentation that no longer reflects reality, maintaining a model whose design decisions they didn't make and cannot fully explain. Four months after launch, someone in customer success mentions, with careful diplomacy, that the model keeps flagging customers who then renew without intervention, while missing the ones who actually churn. The data science lead, who joined eight weeks ago, explains she is still getting up to speed. The VP of Product says something about iteration.

The algorithm continues producing confident predictions into the void.

The Signal Is Not in the Customer Data

The customers themselves are experiencing something the churn model cannot see. Their product is getting incrementally, almost imperceptibly worse. Support tickets take fractionally longer. Old bugs linger across multiple releases. An integration that worked perfectly eighteen months ago now has an intermittent issue nobody can reproduce. No single thing is catastrophic. It is the slow erosion of quality that happens when the people who cared most deeply about the product have been replaced — not by people who care less, but by people who simply don't yet know enough to catch what's quietly going wrong.

This signal does not appear on the customer experience dashboard. It lives in the Glassdoor reviews that reach the C-suite only after heavy curation. It lives in the exit interview feedback that gets filed and not read. It lives in the average engineering tenure, which dropped from four years to fourteen months over the past two years — a fact that appears nowhere near the churn risk tiers.

You can build a remarkably sophisticated radar for detecting customer dissatisfaction. You can train it on years of behavioural data, tune it with genuine care, and surface its outputs in a beautifully colour-coded interface.

It will not tell you that the engineer who designed the feature weighting left for a competitor because her skip-level manager took credit for her work in a board presentation and nobody said anything. That information does not fit in a cell.

The Technology Is Not the Intervention

There is a version of this story that ends better. It requires a leadership team willing to accept one uncomfortable truth: the measurement of customer health and the measurement of organisational health are not separate disciplines. They are the same discipline, viewed from different angles.

An algorithm designed to retain customers is only as good as the team maintaining it. That team's effectiveness is a direct function of how they are led, how they are developed, and whether the people carrying the most institutional knowledge feel enough of a stake in the organisation's future to stay.

The technology, in other words, is the artefact of an intervention that already happened — or didn't. The supermarket didn't need a better scanning app. It needed enough people on the floor to make the shop actually work.

#TechLeadership #EngineeringCulture #CustomerRetention