The evidence trail — defensibility was always the job
Before AI, you already knew this discipline. An adverse action — a rejection, a termination, a demotion — gets documented: what was decided, by whom, for what reason, with notice to the person. Not because documentation is fun, but because eighteen months later, when a claim arrives, the file is the difference between "here is exactly what happened and why" and a shrug with counsel present.
AI adds one new column to that file, and it's the column the next decade of employment litigation will read first: what the tool recommended, and what the human did with it.
The studio's decision record:
ai_recommended— what the tool said: advance, reject, whatever it produced. Logged verbatim. If you can't log it, the tool shouldn't be in the loop — a recommendation that vanishes after influencing a decision is the worst of both worlds: it shaped the outcome and left you nothing to defend with.decided_by— a human name. Notauto. A record where the decider is the tool is a record that says the thing Mobley v. Workday is suing about.reason— the human's reason, in words, tied to the req's requirements. "Missing the CPP certification the req requires" defends itself. An empty reason field defends nothing.notice_date— when the person was told. Notice is the through-line of every AI-hiring statute in lesson two — NYC's ten business days, Illinois's consent, Colorado's adverse-action process. If your records can't show notice, it didn't happen.- Retention — Colorado's law lands on three-year retention for records of automated decisions that materially influence major employment decisions, effective January 2027. Adopt the three years everywhere and stop thinking about it:
decided year + 3is a computed field, not a policy debate.
Here's the part worth saying plainly: the human disagreeing with the tool is not a problem for this record — it's the record working. Human advances a candidate the tool rejected, with a written reason? That's your best evidence that review is real, not ceremonial. Colorado's structure — notice, human review of adverse outcomes, retained records — assumes exactly this. The record you should fear is the quiet one: tool recommended reject, candidate rejected, no name, no reason, no notice. That record exists in thousands of ATSes right now. In yours, after this lesson, it structurally can't.
The last two steps build it: a completeness checker that names each gap in a record, then the checkpoint — a full docket audit that tells you, record by record, what would survive.