The attraction of a workforce prediction is understandable. A communications business has many teams, varied working conditions and managers who cannot examine every pattern unaided. If analysis can draw attention to a developing problem, ignoring it may seem careless. I share the interest in using evidence well. My concern begins when a useful question about a population becomes an apparently settled statement about a named employee. The journey between those two uses is a management decision, even when the underlying model does not change.
I encountered that distinction in a turnover-model pilot covering 686 employees in ten service teams. The proposed output ranked individuals by estimated likelihood of leaving. The intended response was supportive: managers would know whom to approach before a departure became inevitable. Yet the estimate did not tell us what a person intended, what they wanted to discuss or how a manager would interpret the label. A tool built to guide attention could alter the relationship between manager and employee before either had explained their own account.
The output changes when its audience changes
Within an analytical team, a probability may be understood alongside assumptions, uncertainty and the conditions of the dataset. In a busy operating meeting, the same value can become a shorthand: this person is likely to leave. Once that shorthand is attached to a name, it may influence decisions the original project did not intend to address. A manager might offer additional attention, but might also hesitate over a longer assignment or interpret an ordinary disagreement as confirmation of the score. The possibility of a benevolent use does not remove the need to examine other plausible uses.
I therefore ask a question that is sometimes treated as secondary: who will see the result, and in what form? A named list invites a different action from an aggregate pattern. A label in a personnel system may persist beyond the review that produced it. An employee may never know that an estimate affected the conversation. These are not objections that a confidence interval alone can resolve. They concern authority, explanation and the conditions under which a judgement about a person can be corrected.
The strongest objection to my caution is that managers already make imperfect judgements. An analytical estimate might improve on intuition or reveal a pattern that personal relationships miss. I take that objection seriously. The choice is not between a flawed model and flawless human understanding. It is between possible decision arrangements, each with its own weaknesses. A responsible comparison should ask what the proposed tool adds, what new errors it introduces and whether the people using it can recognise its limits in the actual setting where it will operate.
In the pilot, I worked with data director Daniel Corvess and the relevant advisers to separate the analytical question from the intended employment use. I did not ask the people function to certify technical validity or settle specialist privacy and employment questions alone. I asked management to own the use it was proposing. A technically interesting result could still be unsuitable for circulation as a named ranking. Conversely, a narrower question about team conditions might be useful even if we chose not to release the individual output.
A narrower use can be a more responsible one
I stopped the named-list rollout and supported an aggregate review of working conditions. That decision made the output less direct for managers who wanted to know whom to contact. It also changed the proposition we could explain. The work could help identify conditions worth examining across teams without claiming knowledge of an individual’s intention. Managers continued to hold ordinary conversations, respond to expressed concerns and make decisions through the appropriate processes. Analysis became one source of questions rather than an undisclosed reason to treat a particular employee differently.
There was a financial consequence. The original pilot had cost USD 390,000, and redesign required another USD 160,000. I did not present the earlier expenditure as recovered simply because some analytical work remained useful. The executive team needed the full account of what had been spent and what the changed use would require. A project’s sunk cost can create pressure to distribute something, especially when colleagues have worked hard. That pressure is real, but it is not evidence that a particular use is justified.
A narrower use also requires discipline. Aggregation is not a magic protection against misunderstanding or disclosure. Small groups, unusual circumstances and repeated reports can create their own concerns. The intended audience, population and limitations still need review by the appropriate people. I would not describe the revised pilot as a universal answer or suggest that specialist participation made every future use acceptable. The decision had a defined scope and needed continued examination if its audience or purpose changed.
There is also a risk that management uses caution as an excuse not to investigate difficult working conditions. Withholding a named prediction should not mean abandoning the underlying question. If a pattern suggests that employees are encountering a recurring problem, the company still needs a responsible way to understand it. That may involve listening, examining operating arrangements or testing a different hypothesis. The alternative to a questionable tool should be a better decision process, not an empty statement that human judgement will somehow suffice.
Keep the decision owner visible
I want the proposal for any workforce tool to name the decision it will support before discussing the sophistication of the method. The intended user should be able to explain the population, the limits and what would cause the output to be reconsidered. An employee-facing consequence needs an appropriate route for questions and correction. Where the organisation cannot explain those conditions, the honest status may be that the work remains exploratory. That is a useful boundary, even when it delays a visible launch.
I also want later review to examine use, not merely whether the model still runs. Has a score begun to influence a decision outside the original purpose? Do managers understand it as a prompt or as an instruction? Is the organisation collecting more information simply because the tool permits it? These questions belong to management alongside technical monitoring. They help keep a project from acquiring authority by habit while its original assumptions disappear from view.
The pilot did not establish that our change reduced turnover. We did not isolate such an effect, and I would not claim it from the decision alone. What we changed was the organisation’s use of uncertain evidence about people. That is a substantive result, but a bounded one. My broader conclusion is that predictive work should make management more answerable for its reasoning. When an output allows the decision-maker to disappear behind a score, the organisation may have gained a new calculation while losing sight of the judgement it still needs to own.