Automate Everything That Doesn't Change Who Advances

Where is the line?

At the point where a candidate's stage can change without a person reading the application. Everything before that point is logistics: moving information, booking time, formatting notes. Everything at or past it is selection, and federal selection law does not care what the software is called. A procedure used as a basis for an employment decision is a selection procedure ¹.

The Uniform Guidelines define the term as any measure, combination of measures, or procedure used as a basis for any employment decision, and say it covers the full range of assessment techniques from traditional paper and pencil tests through informal or casual interviews and unscored application forms ¹. That is a 1978 regulation and it names no software. What it settles is the category question: a chat widget that dispositions an applicant and a five-minute phone call that does the same are the same kind of thing.

It cuts in the unexpected direction too. Replacing a scored screen with an unstructured conversation does not move the step outside the Guidelines, it only removes the record. The validation burden attaches where adverse impact appears, so the reading to carry is that both automation and judgment are selection, and only one of them leaves you something to show afterwards.

The public draws the line in roughly the same place. In a Pew Research Center survey of 11,004 US adults fielded in December 2022, 71% opposed AI making a final hiring decision against 7% in favour, while opinion on AI merely reviewing job applications was far softer: 41% opposed, 28% in favour, 30% not sure ². Those were attitudes toward AI in hiring as respondents imagined it, before mainstream chatbot use, so the honest reading of the 71% is a judgment about who holds the decision rather than about whether AI appears in a process at all.

What is safe to automate this quarter?

Anything whose output is the same regardless of who the candidate is. Scheduling and rescheduling, reminders, status updates, answering process questions, assembling the interview kit from the requisition, formatting scorecard notes, and chasing feedback out of interviewers. None of those change who advances, and most of them are currently being done late by someone with no time to do them.

• Scheduling and rescheduling. The single largest source of dead time in a loop, and the one candidates notice most.
• Status updates on a clock. A candidate who has heard nothing for three weeks has already started somewhere else. An automated update is not a courtesy, it is retention of people you already screened in.
• Process questions. Timeline, format, what to bring, whether the exercise is paid. Every one of these has a fixed answer that does not depend on the applicant.
• Interview kit assembly. Pull the questions, the rubric and the role context into the calendar invite so the interviewer reads them.
• Note formatting and feedback chasing. Turning a panel's scrappy notes into the scorecard structure changes no outcome and takes real time off a coordinator's week.

The rejection message is where this category earns its reputation back. In a four-scenario vignette experiment with 921 working-age Austrians, a rejection from an AI with no explanation scored lowest on all four outcomes measured, from 1.49 for recommendation intention to 1.86 for outcome fairness on a five-point scale, while an AI rejection carrying an explanation drew the same ratings as a human rejection without one ³. Those were hypothetical rejections judged by an online panel and every condition sat below the scale midpoint, so read it as a ranking among unhappy outcomes rather than evidence anyone felt well treated. Automate the send. Do not let the explanation go with it, and settle in advance what you owe a candidate who asks why.

What changes who advances?

Pre-qualification chat that dispositions on an answer, ordered shortlists that decide what a recruiter opens, auto-rejection on a threshold, and assessment features that return a number attached to a person. Each can end a candidacy with nobody reading the file, which is the whole test, and each brings duties that the scheduling features never carry.

New York City has required audit-plus-notice since 2023. Local Law 144 bars an employer from using an automated employment decision tool on a New York City candidate unless a bias audit was done within the prior year, a summary is posted publicly, and the candidate had notice at least 10 business days before the tool is used, with civil penalties running between $500 and $1,500 per day ⁴. Those are statutory maximums rather than sums anyone has been made to pay, and the rule reaches only tools that substantially assist or replace discretionary decision-making, which is a determination the employer makes first and defends later.

Illinois took a different route. Public Act 103-0804 amended the Illinois Human Rights Act, effective January 1, 2026, to make it a civil rights violation to use artificial intelligence that has the effect of discriminating on a protected basis in recruitment or hiring, to use zip codes as a proxy for a protected class, or to fail to notify an employee that AI is being used for those purposes ⁵. That is an effects standard. Texas went the other way: its Responsible Artificial Intelligence Governance Act, also effective January 1, 2026, prohibits deploying an AI system with the intent to unlawfully discriminate, and says a disparate impact alone does not establish that intent ⁶. There is no single sentence that means compliant with AI hiring law, so check the states your candidates sit in rather than the state on the letterhead, and take the reading to counsel before it becomes policy.

Some of this is already running without anyone calling it automation. A screening question with a disqualifying answer dispositions applications with no human involved, which is why the auto-reject question turns out to be a question about your own rules rather than about the vendor.

Decide it feature by feature, in writing

Go through the features you already own before buying another one. For each, write a sentence naming what it can change, who reads its output, and whether a candidate's stage can move without a person. File those sentences with a date and an owner. That document is the only thing that answers the question when it arrives, and it always arrives after the decision rather than before it.

Three questions do most of the sorting, and a vendor who cannot answer them quickly has told you something.

1. Can this feature change a candidate's stage on its own? If yes, it is a selection procedure. Ask what the fallback is when it is switched off. 2. What does it write into the candidate record? A number in the file is part of the decision trail whether or not anyone looked at it. 3. Can it be disabled per requisition, and who can do that? A setting only the vendor can change is not a control you hold.

The traps are mostly defaults. A feature that ships enabled, an ordering that quietly becomes the review order, an integration that starts writing scores into records the day it is connected. None of those involved a decision anyone made, which is precisely why they end up undocumented.

Something does come back, and it is not speed on its own. Automating logistics gives coordinators their week back and stops candidates leaving mid-process; it does not shorten a loop that is slow because feedback sits with an interviewer for six days. Teams that added AI at every stage this year and got slower usually find the time reappeared somewhere specific, and it is rarely the stage they automated.

See a sample report >