Pipeline
Publishing the Evaluation Process Inside the Job Posting
A job posting should say how applicants will be evaluated: publish the stages, what each one examines, where AI touches the file, where a named human makes the call, and the elapsed time. One short block near the end of the posting. Then run exactly that. A described process you quietly skip becomes a complaint you have to answer, and where disclosure rules apply, a misstatement. Where AI is involved, say what it does rather than that AI is used.
The takeThe commitment is the feature, not the risk. Publishing the process is the cheapest way to stop a hiring loop growing a fifth round unnoticed, because the posting is a public artifact and the fifth round would have to be added in front of everyone who already applied. Candidate trust is the second-order benefit. If the honest objection is that the process changes too often to publish, that is worth knowing before a candidate finds it out.
Where Olive fits
Open a role and see what the work shows
Under the automated-decision rules, a number standing for a candidate explains nothing. Olive produces no composite at all: a person writes each of the six findings, every finding carries the excerpt it rests on, and a released report exports with its rubric, scorer and bank versions attached.
Rank your shortlistWhat belongs in the block, and in what order?
Five things, in the order a candidate meets them: how many stages there are, what each stage looks at, which of them involve an automated tool, who makes the decision at the end, and the elapsed time from application to offer. Put it under a plain heading near the end of the posting, after the role and before the pay range or the legal block.
A worked version, for a four-stage loop:
- Application review: a recruiter reads every application. A matching tool orders the queue; it removes nothing.
- Screen, 30 minutes: two people, same questions, notes against a written rubric.
- Assignment, 90 minutes: a task from the role, done on your own clock, graded blind by two reviewers.
- Panel, 90 minutes: three interviewers, one hiring manager, decision within five working days.
- The hiring manager decides. Recruiting compiles, and does not vote.
- Application to offer: about four weeks.
Six lines carry more information than most careers pages carry in total, and every one of them is a fact rather than a promise. The order matters because a candidate reads to the first thing that disqualifies the job for them, and for most people that is the assignment or the elapsed time rather than the number of interviewers.
Say the elapsed number you actually hit three times out of four, not the best case. A posting that says two weeks and takes six has spent the credibility it just bought, and the candidate who withdrew on day fifteen was reading the posting rather than the pipeline.
Why naming the decider does the most work
Because that is where the objection concentrates. In Pew's survey of 11,004 US adults, 71% opposed AI making a final hiring decision against 7% in favour, while opinion on AI merely reviewing applications was far softer: 41% opposed, 28% in favour, 30% unsure 1. Naming a person at the end of the process answers the objection most readers actually hold.
Two caveats belong on that number. It was fielded in December 2022, before mainstream assistant use, and it records what respondents imagined rather than any tool they had met. The same survey found 47% thought AI would do better than humans at treating all applicants the same, against 15% who said worse, so the public is not uniformly hostile to AI in hiring; it objects to AI holding the decision.
The research on how automated assessment feels points the same way. A review of seven studies with more than 1,300 participants between them found overall fairness perceptions mixed, but perceptions of behavioural control and social presence mostly negative: candidates feel less able to influence the outcome and feel the human element is missing 2. It is a narrative review with scenario-based samples, so it supports a direction rather than a magnitude. Its most useful finding for a job posting is the negative one: explaining how the algorithm decides did not reliably improve perceptions. Name the person, not the model.
That is a different job from the notice a jurisdiction may require when AI screens applications, which is a legal duty with its own wording and timing.
Publish only the process you will actually run
Publish the version you ran last time, not the one on the process map. A stage you describe and skip is a promise a candidate can quote back, and where disclosure duties apply it is a statement about your process that turns out to be wrong. The test is small: could the recruiter working this requisition read the block aloud and recognise their own week?
Most teams fail that test in one specific way. The published loop describes the standard process, and the actual loop grows an extra conversation whenever a hiring manager is unsure, which happens on roughly the candidates who are closest to the bar. Publishing does not stop that; it makes it visible, which is the point. If the extra conversation is real, publish it as a stage. If it is a symptom of a decision rule nobody wrote down, that is a different problem, and what belongs in an AI hiring policy is where it gets fixed.
Where requisitions genuinely differ, publish per family rather than per role: one block for engineering, one for go-to-market, one for operations. Three maintained blocks beat thirty stale ones, and the template is the thing that survives a recruiter leaving.
One more reason to write the elapsed time down honestly: adding AI at several stages does not shorten a loop on its own, and the published number is the first place a longer one shows up. If the honest answer has moved, work out where the time is going before republishing a number you will miss again.
How should the AI stage be worded?
Name the function, the stage, and the human who reads the output. A resume-matching tool orders the first pass, and a recruiter reads every application it surfaces as well as a sample of the rest: wordy, and correct. AI is used to improve the candidate experience is short and, when anyone asks what it does, unanswerable, which is the version that draws both complaints and legal attention.
The wording matters most at the end of the process, because it is what a rejected candidate quotes back. In a vignette experiment with 921 working-age Austrians, a rejection from an AI with no explanation scored lowest on all four measures, from 1.49 for recommendation intention to 1.86 for outcome fairness on a five-point scale, while an AI rejection that carried an explanation drew the same ratings as a human rejection without one 3. These are hypothetical rejections imagined by an online panel, and every condition sat below the scale midpoint, so this ranks unhappy outcomes rather than promising a good one.
That result is a design instruction for the posting rather than for the rejection email. The explanation a candidate eventually gets can only describe stages the process actually has, and the block in the posting is where those stages get named. Teams that write it well find the rejection conversation almost writes itself.
Keep one sentence back for the tools you have not decided about. Saying that no automated tool makes a rejection decision at any stage is a strong claim and worth publishing only if the ATS settings actually support it.
Common questions
How specific does the published timeline need to be?
Specific enough to be wrong. A range in working days for each stage, plus a total, is the useful form: screen within a week of applying, assignment returned within five days, decision within five days of the panel. Publish the number you hit about three times in four rather than the best case, and treat a missed number as a process defect rather than a communications problem.
What if different requisitions run different processes?
Publish per job family, not per role. One block for engineering, one for go-to-market, one for operations, each maintained in the posting template. Three accurate blocks are worth more than thirty that drifted, and a candidate comparing two of your postings will notice the contradiction faster than any internal reviewer.
Does publishing the process cost applications?
Some, and the survey evidence on adjacent questions suggests the cost is concentrated among people who object to automated evaluation in general. Pew found 66% of US adults saying they would not want to apply to an employer using AI to help make hiring decisions, but that is stated intention in a survey rather than measured drop-off, and nobody in the sample was standing in front of a job they wanted. Treat it as a reason to describe the AI stage precisely, not as a forecast of your funnel.
Should the block name the actual interviewers?
Name roles, not people. Two engineers and the hiring manager is durable; two named engineers is stale the moment somebody changes team, and it turns a process description into a maintenance task. The exception is the decider: naming the role that owns the decision, usually the hiring manager, is the single most useful line in the block.
What happens if the process changes mid-requisition?
Tell the candidates already in the pipeline before you update the posting. A stage added after somebody applied is the case that generates complaints, and an email explaining the change costs nothing. If the change is structural rather than a one-off, update the template so the next requisition starts honest.
References
- 1. Americans' views on use of AI in hiring (chapter of 'AI in Hiring and Evaluating Workers: What Americans Think') pewresearch.org Supports the claim that public objection concentrates on AI holding the decision rather than on AI being involved, and the applicant-willingness figure in the FAQ.
- 2. Robots are judging me: Perceived fairness of algorithmic recruitment tools pmc.ncbi.nlm.nih.gov Supports the claim that candidates object to lost influence and a missing human, and that explaining the algorithm does not reliably improve perceptions.
- 3. Rejected by an AI? Comparing job applicants' fairness perceptions of artificial intelligence and humans in personnel selection frontiersin.org Supports the claim that an unexplained automated rejection is the worst case measured, and that the explanation does most of the work.
3 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.