Most operating problems are not analytical. They are archival. Before you buy another forecasting tool, find out whether your team can explain the last twenty calls it made.
The problem with undocumented calls
Every operating team has a shadow archive: the reasoning behind decisions that were made in a room and never written down. Six months later the decision survives, but the argument for it has evaporated — so nobody knows whether the conditions that justified it still hold.
We started measuring this in 2023. Across fourteen engagements, the median team could reconstruct the rationale for fewer than one in five of its material decisions from the previous two quarters. The decisions themselves were fine. The institutional memory was not.
“The decisions themselves were fine. The institutional memory was not.”
What a decision record actually contains
A useful record is short and unglamorous: the call, the date, who owned it, the two or three alternatives considered, and — the part everyone skips — the conditions under which you would revisit it. That last line is what turns a note into an instrument.
Teams that write the revisit condition down catch stale decisions roughly a quarter earlier than teams that don't. That is not a modelling advantage; it is a filing advantage.
0
% of calls left unrecorded
0
% earlier stale-decision catch
0
engagements measured
Where AI genuinely helps
Language models are unusually good at the boring half of this: drafting the record from a meeting transcript, tagging it against the decisions it supersedes, and surfacing the three prior calls most likely to conflict. What they are not good at is deciding. Keep the judgement human and let the machine keep the ledger.
The failure mode we see most often is inverted: teams automate the judgement and leave the filing manual. That produces confident decisions nobody can audit.
Starting this quarter
Pick one decision forum — the weekly ops review is usually the right one — and record every material call for eight weeks. Do not roll it out further until someone has actually used the archive to reopen a decision. If nobody reaches for it, the format is wrong, and scaling it would only multiply the waste.
Eight weeks is deliberate: long enough for a decision to age into needing a second look, short enough that abandoning the experiment costs nothing.
Priya Raghunathan
Priya leads decision science at Northvane. She spent nine years in operations before deciding that most analytics problems were really filing problems.
Work with Priya →