Digital Tools: Friendship Data, Filed Between a Pay Rise and Better Hiring

Turnover is expensive in a way that is easy to underrate. One study cited in the chapter costed it precisely: a call centre with 31 agents and 4 supervisors, losing agents at roughly 60% a year, at a total annual cost of about $400,853. Rates vary enormously by sector, and the two figures the chapter reaches for are not measured the same way: voluntary turnover in educational services runs about 14%, while in many food service firms turnover including involuntary departures reaches around 86% a year — replacing almost the entire workforce, annually.

So the pressure to predict who is about to leave is real, and it is financial rather than sinister. What is worth looking at closely is what the field now reaches for in response, and how the reaching gets described.

The paragraph

Under the heading DIGITAL TOOLS, the chapter documents three programmes in about sixty words:

SAS’s employee-retention program sifts through employee data on traits like skills, tenure, performance, education, and friendships. It can predict which high-value employees are more likely to quit soon, and to try to prevent that. Alliant Techsystems created a statistical “flight risk model,” to calculate the probability an employee would leave and to take corrective action. Google’s “Googlegeist” survey contains five questions for identifying Googlers who are more likely to leave; if a team’s responses fall below 70% favorable, Google takes corrective action.

Three named companies, three approaches to the same problem, presented in one continuous list. Read them as designs rather than as examples and they diverge immediately.

Observation and participation are different inputs

The first two programmes work on records that already exist. An employee generated them by being employed — appraisal scores, a start date, a qualifications field, and in SAS’s case something the chapter calls friendships.

The chapter does not say how friendships are determined, and that gap is worth sitting with rather than filling in. Four of the five named inputs correspond to a field in an HR system. The fifth names a human relationship with no stated method of measurement attached. Whatever operationalises it is doing a great deal of the model’s work and is not on the page.

The Google programme works on answers. People sat down and responded to five questions, knowing they were being surveyed.

It would be too generous to call that contestable. Googlegeist is anonymous and aggregated — there is no personal score to look at, and so none to dispute. If a team lands under 70%, the intervention arrives whether or not any particular member gave low marks.

The real difference is not agency, then. It is that the Google design never produces a per-person artifact at all. The other two do: a row about a named individual, held somewhere that individual has no route to. There is no screen showing it and no process expecting an objection — not out of malice, but because nobody designing a retention dashboard anticipated the subject of the row asking to see it.

The same record, pointed two ways

Earlier in the same chapter there is a second use for all this data, described just as briefly: employers accumulate a great deal of information on their people — jobs held, appraisals, training, skills — and many now use AI and analytics to predict how current employees would fit the requirements of particular roles, and to identify possible career paths for them.

That is the same file, read by the same kind of model, pointed in the opposite direction. One question is where could this person go next. The other is is this person about to leave. The first produces an offer the employee finds out about, because making the offer is the entire point. The second produces a score the employee does not find out about, because telling them is at best pointless and at worst accelerates the thing it predicts.

So the posture toward the employee is what separates them, not the technology and not the data. Both are described in the chapter in a couple of sentences, several pages apart, without either being connected to the other.

”Corrective action,” twice

The phrase appears twice in sixty words, attached to two very different objects.

The distinction matters most when the prediction is wrong. A team score below 70% that turns out to be a bad quarter costs a manager a conversation. An individual flight-risk score that is wrong costs the individual — quietly, through decisions they never see, such as not being staffed onto the eighteen-month project, which is a reasonable thing for a manager to do with someone the system says is leaving.

None of which makes the survey benign. A published threshold is a target, and targets get managed. A team that learns 70% is the line has an obvious reason to keep its answers above it, and the pressure to do that falls on the same people the survey exists to protect. Passive observation and direct questioning fail in different directions. Neither fails gently.

The heading is doing work the text does not

Nothing in the chapter argues that modelling employees’ friendships is appropriate. It does not defend the practice, because it never raises the question. What it does is place the practice second in a list, in the same typographic register as RAISE PAY and HIRE SMART.

It would be a stretch to call that a position. A textbook in its seventeenth edition accumulates vignettes, and new examples get filed under existing headings because that is how an edition ships on schedule. A subheading called DIGITAL TOOLS is a filing decision more than an argument.

The effect does not depend on the intent. A list of tactics carries an implicit claim — that its items are alternatives to one another, comparable, selectable according to fit and budget. A reader arrives at that equivalence whether or not anybody placed it there, and no one has to have decided anything for the framing to transfer.

This is worth noticing because the summary is where the framing survives. Compress that paragraph into a sentence for a slide or a report and the natural compression is leading firms use analytics to predict and prevent attrition — which keeps every claim and drops the only distinction in the passage. The equivalence gets inherited rather than examined, because the layout never flagged that there was anything to examine.

Two questions that separate the three

Neither of these is a general ethics test. Both come from the differences visible in this one paragraph, which is what makes them cheap enough to actually use.

Accuracy is the question these systems are usually defended on, and it is the wrong first question. A flight-risk model that is right most of the time still produces a set of people who are treated differently on the strength of a forecast, and the ones it gets wrong absorb that treatment without ever learning why.

The cheapest intervention in the section needs no model

A few paragraphs earlier, the same chapter records what Walmart found: it significantly reduced voluntary turnover by giving candidates very realistic previews of the job’s demands and hours.

No data mining, no probability, no dashboard. People left less often when they had been told accurately what the work was before they took it. It is the least technical item in the whole retention section, it is sitting in the same few pages as the friendship model, and on the evidence presented it worked.

Thanks to a chapter that put all of them in one place, which is the only reason the comparison is available at all.

If you have seen a retention model in use: was it aimed at people, or at the conditions people were leaving? I would be interested in how often it is the second.