The standoff, as it actually happens
The organisation invested in the decision-support capability precisely so that choices like this would be evidence-led. Now, in its first consequential test, the output points to Option A (exit the underperforming line, reallocate the spend, price differently) and the executive team, having read it, still prefers Option B. The data science lead is quietly furious: if the model only counts when it agrees with the room, why was it built? The COO is equally firm: the model has never met a customer. Someone proposes "taking elements of both", which would dodge the question rather than answer it.
Why neither side should simply win
Deferring to the model risks laundering a bad decision through mathematical authority: the training data may encode a market that no longer exists, and the objective function was written by people who made assumptions nobody has revisited. Deferring to leadership risks the older failure: instinct defended as experience, which on inspection is sometimes just seniority plus anecdote. The deeper problem is asymmetric accountability. If Option B fails, an executive misjudged. If Option A fails, "the organisation trusted a machine" becomes the headline, and every person in the room is silently pricing it, which biases the vote against the model regardless of its merits.
The risks that surface too late in each direction
- Following the model, it later emerges a data pipeline change shifted the inputs, and nobody in the meeting knew the recommendation rested on it.
- Overriding the model, the override becomes precedent: within a year the system is expensive decoration, consulted only for confirmation.
- The "blend both options" compromise dilutes each path below its threshold of effectiveness, achieving neither.
- The dissent record is vague, so when the outcome is known, memory politely rewrites who argued what.
- The real disagreement, about the objective rather than the answer, was never surfaced, so it recurs on the next decision.
Questions that reframe the disagreement
- What exactly was the model asked to optimise, and does the executive team actually agree with that objective?
- What does leadership believe that the model cannot see, and can that belief be stated as a testable claim rather than a feeling?
- On what past decisions has this model been right and wrong, and does anyone keep that record?
- If an analyst, rather than an algorithm, had produced the identical recommendation, would the room take it more seriously or less?
- What evidence, available within weeks, would change either side's position?
What to pressure-test before choosing either path
Interrogate the model's inputs and objective before its output. Most disagreements of this kind dissolve into a definitional gap once someone reads what the system was actually asked. Convert leadership's preference into explicit assumptions and test which are load-bearing. Where the options permit it, stage the commitment: run Option A in a bounded market or segment while preserving Option B's core, and let the disagreement be settled by evidence rather than by rank. And document the reasoning either way, because the value of this episode is partly what it teaches the organisation about when to trust the system it bought.
Who can arbitrate credibly
Not the data team, who built the model, and not the executive sponsors of Option B, who own the instinct. Internal arbitration fails here because everyone has a stake in one answer. The useful outside voice is doubly specific: operators who have made this class of commercial decision at scale, and who have also lived with algorithmic decision support long enough to know its characteristic failure modes. Selected senior operators from the Global Board can look at the model's recommendation and the leadership case side by side and say which parts of each survive contact with their experience, an input neither camp inside the building can provide.
