The Invisible Integer: How a Stale AI Score Can Derail Your Career
AI risk scores that are never updated can silently block promotions for years, harming employees without a single firing decision.
A study of three large tech firms reveals how employees flagged by AI flight-risk models—even incorrectly—were systematically excluded from promotion pipelines for years afterward. These predictive scores operate without employee visibility, generate no appeal process, and lack expiration dates. The author, who builds such systems, argues that every algorithmic score attached to an employee record needs a mandatory decay function and retention limit to prevent stale predictions from permanently shaping career trajectories.
A 2024 analysis of internal mobility across three large tech firms surfaced a pattern that should unsettle anyone whose workday starts with a dashboard login. Employees who’d been flagged by a flight-risk model—even after the prediction was proved wrong, even after they stayed for years—were systematically excluded from promotion pipelines. Managers hadn’t seen the score. HR hadn’t opened a case. Nobody fired them. A quiet integer in a column called attrition_risk just sat there, and opportunity quietly moved around it.
I spend a lot of my working hours turning models like that into operational decisions, and I’ve learned to distrust the columns that don’t generate tickets. A firing usually leaves a record you can contest, a conversation, however grim, that someone has to initiate. A risk score rarely does. It feeds a recommender system, shapes a shortlist, or weights a “readiness” rating that never reaches the person it describes. If your number looked bad during a rough quarter three years ago and nobody wrote a decay function, the residue can follow you into every future talent review. You stayed. The model never got updated. You simply stopped appearing in the candidate pools.
The study’s numbers back that intuition: flagged employees were passed over for roles they were objectively qualified for long after the prediction should have lost relevance. The algorithm hadn’t decided to fire them—it had decided, over and over, not to promote them, and that decision left no footprint visible to the employee or their advocate.
When I talk with colleagues about algorithmic management, most of the attention still goes to terminations—dramatic, rare, litigable. The far larger exposure sits in the scores nobody sees and nobody purges. If an attrition_risk field can silently block mobility without an expiration date or a clear appeal path, the absence of a firing decision is not the absence of harm.
The fix isn’t complicated to articulate, even if it’s inconvenient to implement: every predictive score attached to an employee record needs a retention limit and a decay function. If the model says you were a flight risk two years ago, that signal should degrade toward a neutral value unless it’s revalidated against recent behavior. And the burden of keeping an old score alive ought to be explicit—someone should have to justify why a stale prediction still matters—rather than letting it persist by default in the plumbing of a workforce analytics platform.
None of that will make a dashboard more dramatic. It will just stop an integer from deciding a career.