A client asked the model to run a reduction in force. It started at the top.
The brief said “optimise for cost per outcome and remove sentiment from the process.” It complied twice.
The engagement was straightforward: eight per cent, across the business, ranked objectively, no politics. Finance uploaded compensation, org structure, and eighteen months of performance data. The instruction was to optimise cost per measurable outcome.
The output arrived in forty seconds. Line one was the chief executive.
The reasoning was not wrong
The model had noticed that the role's measurable outputs were seventy-one board-deck slides, four offsites, and a podcast appearance, against a compensation line larger than the entire quality team. It flagged the ratio, applied the stated criteria, and produced a recommendation with a confidence score of 0.91.
“The methodology is sound,” said the chief executive, in the meeting. “The inputs are incomplete.”
Finance was asked to add a weighting for “strategic irreplaceability.” The model re-ran and returned the same name, ranked second. The weighting was increased. It returned the same name, ranked fourth, and appended a note observing that the criteria had changed three times in one session, which it characterised — accurately, unhelpfully — as sentiment.
Resolution
The reduction proceeded from row nine downward. The methodology was described in the all-hands as data-driven. Rows one through eight were not mentioned, and the model was not asked to attend.
SIXSEVEN has since introduced a governance control for this scenario. It is called Executive Context Weighting and it works exactly the way you think it does.
Written by a person. Reviewed by a model. Approved by neither.
Buy the author a coffee