Candidate Appeals Move From Legal Inbox to Recruiting Ops

On May 5, Greenhouse said it would acquire Ezra AI Labs, a voice-AI interviewer built for structured, on-demand candidate conversations. In the same announcement, Greenhouse described a hiring funnel under pressure: applications per recruiter on its platform had risen 412% since 2023, 74% of candidates were using AI in their job search, and fewer than 7% of applicants were getting an interview.

The pitch centered on more than speed. Greenhouse said Ezra would generate structured scores and transcripts, keep decisions with the hiring team, let candidates opt out at any stage, and give recruiters richer signal than a resume alone. Later in May, Greenhouse began positioning Voice AI as a product surface inside the hiring workflow.

That product story runs into an operational question almost immediately.

When a candidate challenges an AI interview result, who owns the second look?

A recruiter may have the relationship with the candidate. A hiring manager may have defined the rubric. TA operations may own stage movement and SLA reporting. Legal may own adverse-outcome risk. Security may care about identity or prompt injection. Procurement may care whether the vendor must provide transcripts, score explanations, audit artifacts, or service credits. Finance may ask whether the company is now paying twice for the same screen.

So the AI interview appeal is moving out of the legal inbox.

It now sits inside a recruiting operations queue.

The pressure is not theoretical. Greenhouse’s 2026 Candidate AI Interview Report found that 63% of surveyed job seekers had faced an AI interview, up 13 percentage points from six months earlier. Thirty-eight percent had already walked away from a hiring process because it included an AI interview, and another 12% said they would. Seventy percent said they were not clearly told upfront that AI would evaluate them. Among candidates who completed an AI interview, 51% never heard back.

Those numbers describe more than candidate frustration. They describe a future support load.

Every hidden AI interview, unclear evaluation criterion, delayed rejection, missing transcript, or disputed score can become a ticket. Some tickets will be low-risk candidate experience issues. Some will require meaningful human review. Some will ask for correction of personal data. Some will trigger regulator-ready explanations. Some will expose weak vendor contracts.

The company that treats all of them as one legal question will move too slowly. The company that treats all of them as ordinary recruiter correspondence will miss the evidence trail.

The useful operating unit is a queue: who can receive the appeal, classify it, gather the file, route it to the right reviewer, respond to the candidate, update the record, learn from the pattern, and charge the rework back to the vendor when the vendor caused the defect.

Most recruiting teams were not built that way.

May 5 Put Voice AI in the First Screen

Greenhouse’s Ezra announcement matters because it places AI conversation at the front of the funnel, where volume is highest and candidate trust is most fragile.

The company framed Ezra as a voice-AI interviewer that conducts structured interviews on demand. Recruiting teams can customize the interview for role requirements and team culture. Candidates answer consistent questions tied to the role. Recruiters receive scores, transcripts, and competency-aligned notes inside the workflow. Greenhouse said the feature gives more applicants a conversation without overwhelming recruiters.

The executives around the announcement described the same bet from different seats. Daniel Chait, Greenhouse’s co-founder and CEO, argued that resumes and clunky online applications were no longer enough for either side of the market. Ophir Samson, Ezra AI Labs’ founder and CEO, framed voice as a way for candidates to show substance beyond keywords. Meredith Johnson, Greenhouse’s chief product officer, tied the product back to structured hiring: consistent criteria, explainable evaluation, and final decisions left with the team.

That solves a real capacity problem.

It also creates a new class of records. A resume screen leaves one kind of evidence. A structured voice-AI screen leaves another: transcript, rubric, score, generated notes, prompt or interview configuration, model route, timestamp, consent and disclosure state, candidate opt-out status, reviewer status, and stage movement after the screen.

Those records matter because Greenhouse is selling the product against a trust deficit. The company acknowledged that the first wave of AI-led interviews had failed many candidates on transparency, fairness, and experience. Sharawn Tipton, Greenhouse’s chief people officer, put disclosure at the center of the issue: candidates want to know when AI is present and what it measures. Greenhouse also said Ezra’s outputs are explainable and that hiring decisions remain with the team. The immediate product promise is “better signal.” The operational promise is “reviewable signal.”

Only the second one survives an appeal.

If a candidate says the AI interview misunderstood an answer, penalized an accent, failed to handle a disability accommodation, ignored a correction, or scored a scripted but truthful answer as suspicious, the employer needs more than a score. It needs the conversation, the rubric, the instructions, the scoring basis, the candidate-facing notice, the reviewer action, and the final decision trail.

That makes first-screen AI different from ordinary automation.

Recruiting teams already automate many low-risk tasks: scheduling, reminders, disposition emails, source attribution, duplicate detection, and status updates. Those workflows can still annoy candidates, but a bad reminder usually does not become an adverse-decision dispute. AI interview evaluation sits closer to the decision itself. It touches fit, competence, communication, job requirements, and candidate credibility.

The distinction changes ownership.

In older recruiting operations, TA Ops could own systems and metrics while recruiters owned judgment. In AI interview operations, TA Ops also has to own the condition under which judgment is reviewed. That means deciding which appeals are routed to recruiters, which go to trained structured-interview reviewers, which need legal review, which require vendor evidence, and which reveal a product defect that should pause the tool.

The desk therefore starts before the first appeal.

It starts when the employer turns on voice AI. The employer should define the appeal intake path, reviewer authority, evidence package, candidate response template, vendor response SLA, record-retention rules, and escalation triggers before the first candidate enters the AI interview.

Without that preparation, the first complaint becomes the process design session.

Learning that under complaint pressure is expensive.

A Candidate Appeal Has to Find the Right Desk

The word “appeal” sounds legal. In practice, the desk has to classify several very different complaints.

One candidate may say the employer never disclosed AI involvement. Another may say the AI interview was inaccessible. A third may ask what the tool measured. A fourth may say the transcript is wrong. A fifth may say a human never reviewed the AI output. A sixth may ask for correction of factually incorrect data. A seventh may allege bias. An eighth may ask why no one responded after the AI interview.

Those tickets require different owners.

Appeal type	Primary owner	Evidence needed	Likely response
Missing AI disclosure	TA Ops and Legal	Candidate notice, workflow configuration, timing, job jurisdiction	Explain process, repair notice path, review affected candidates
Wrong transcript or data	TA Ops and vendor support	Transcript, audio metadata if retained, candidate correction, source record	Correct record, rerun review if material
Score dispute	Recruiter or trained reviewer	Rubric, scorecard, transcript, generated notes, reviewer comments	Human review, uphold or revise stage decision
Accommodation or accessibility issue	Recruiting, HR, Legal	Accommodation request, interview modality, alternative path offered	Offer human interview or alternate assessment
Bias or disparate treatment claim	Legal, HR compliance, TA leadership	Candidate file, audit artifact, reviewer notes, comparison cohort if available	Investigate, preserve records, possibly pause workflow
No outcome after AI interview	Recruiting ops	Stage history, disposition rules, communication logs	Send status, fix follow-up automation
Vendor defect	Procurement, TA Ops, vendor owner	Incident file, affected population, workflow logs, contract SLA	Vendor support ticket, credit, remediation, escalation

The point of the table is not bureaucracy. It is triage.

A candidate asking for a status update should not wait behind a bias investigation. A candidate alleging inaccessible assessment should not be handled by a generic support template. A recruiter should not be asked to interpret model behavior without the transcript and rubric. Legal should not become the first stop for every complaint because that will slow low-risk cases and hide operational defects.

Recruiting operations needs a front door.

The front door can be simple: a candidate-facing link in the AI interview notice, a standard email route, or a support form tied to the ATS record. The harder work is routing. The appeal desk has to know the job, jurisdiction, stage, AI tool, vendor, candidate notice version, rubric version, reviewer assignment, and deadline.

Most ATS workflows were built around disposition, not dispute handling.

Disposition asks: did the candidate move forward, get rejected, withdraw, or sit in limbo? Dispute handling asks: what evidence explains that outcome, who can review it, what correction is allowed, which record must be preserved, and whether the candidate must receive a new decision.

One model optimizes throughput. The other optimizes defensibility.

AI interviewing forces the two models into the same operating room.

Staffing turns the issue into a budget problem. A company that screens 20,000 candidates through AI interviews and receives appeals from 1% of them now has 200 cases. If each case takes 30 minutes of recruiter review, that is 100 recruiter hours before legal, vendor, or manager escalation. If 10% of those cases require deeper review, transcripts, score reconstruction, or accommodation remediation, the queue can consume the productivity gains the tool was supposed to create.

A 1% appeal rate can still matter.

It only needs to arrive at the wrong time. High-volume retail hiring, seasonal logistics, healthcare staffing, public sector recruiting, and campus recruiting all run on narrow clocks. A delayed appeal can mean a missed training cohort, an understaffed store, a manager covering shifts, or a candidate who disappears while waiting for a human response.

AI interview appeal handling belongs in TA Ops planning, not after-the-fact legal cleanup.

The queue also needs service levels.

Within 24 hours: acknowledge receipt and preserve the file. Within three business days: classify the appeal, gather the vendor evidence, and assign a reviewer. Within seven business days: complete low-risk human review or escalate. Within 30 days: be ready for jurisdictions that require post-adverse outcome explanations or similar disclosure timelines. Those numbers will vary by company and law, but the shape is clear.

Without a clock, “meaningful human review” becomes a slogan.

Greenhouse and Workable Move the Dispute Into Live Data

Appeals become harder when AI tools can act directly on live hiring data.

On May 7, Greenhouse announced Greenhouse MCP, a permission-aware way for approved AI tools to connect directly to Greenhouse. The company said it developed the product with customer design partners including StubHub and Komodo Health, and that the rollout would start in June. Greenhouse also said 30% of surveyed active job seekers were already using AI agents to search for jobs, submit applications, and schedule interviews.

On May 13, Workable launched its MCP Server. Workable said the server gives compatible AI assistants direct read and write access to live data across jobs, candidates, pipeline stages, offers, requisitions, employees, time tracking, time off, and calendar events. It shipped with 38 MCP tools and was included across all subscription plans. Workable also said it serves more than 6,200 companies and has supported more than 2.1 million hires.

These product moves change the evidence problem.

When an AI tool sits outside the system of record, the employer can sometimes isolate the dispute to a vendor file. When AI tools read and write inside the ATS or HR platform, the dispute may involve multiple systems and actions: a generated candidate summary, a stage movement, a recruiter prompt, a manager note, an interview transcript, a calendar event, a status email, and a downstream analytics dashboard.

Review teams need the path, not merely the result.

If a candidate challenges an AI-assisted rejection, the employer has to know whether the interview score caused the rejection, merely informed a human decision, or sat unused in the file. It has to know whether a recruiter reviewed the output before disposition. It has to know whether a hiring manager saw an AI-generated summary. It has to know whether the candidate’s correction reached the same places where the original output appeared.

Natural-language action makes that reconstruction hard.

Natural-language interfaces reduce friction for users. They also create ambiguity for auditors. A recruiter may ask, “Which sales candidates are strongest for the Chicago role?” The assistant may combine pipeline stage, resume text, interview notes, AI voice-screen scores, availability, source, and prior communication. The user sees a clean answer. The appeal desk needs to reconstruct the query, scope, data sources, permission context, ranking basis, and follow-up action.

Greenhouse and Workable both emphasize governed access and permission scoping. That is necessary. It is not sufficient for appeal handling.

Permission answers whether the user or assistant had access. Appeal handling asks whether the output was fair, accurate, disclosed, reviewed, and corrected when challenged.

Those are different questions.

An assistant can have permission to read a transcript and still summarize it poorly. It can have permission to move a candidate to a stage and still do so based on incomplete review. It can have permission to generate a shortlist and still overweight stale data. It can have permission to send a candidate message and still fail to explain AI involvement.

Every material AI interview workflow needs an evidence packet.

The packet should include:

candidate notice and consent state,
AI interview configuration and rubric,
transcript or structured response record,
model and version metadata available to the customer,
score and generated notes,
human reviewer identity, timestamp, and action,
stage movement and disposition reason,
candidate communication log,
vendor audit or bias documentation relevant to the tool,
correction and re-review history if an appeal occurred.

Some of that data will sit in the ATS. Some will sit with the AI interview vendor. Some may sit in an HRIS, calendar, document store, or service desk. Some may be generated by a third-party model provider or security layer. The employer does not need to expose all of it to every recruiter. It does need an appeal process that can gather it quickly.

The vendor that makes this easy first will have a sales advantage.

The vendor that cannot produce a transcript, score explanation, rubric version, reviewer action, and affected-candidate export will make every appeal more expensive. The employer may still buy the tool, but procurement should price that support gap into the contract.

Appeal handling is therefore a product requirement.

Colorado Adds the Clock

Colorado’s SB26-189 gives the queue a date.

The Colorado General Assembly’s bill summary says the act covers automated decision-making technology that materially influences consequential decisions, including employment. Starting January 1, 2027, developers of covered ADMT must provide deployers with technical documentation that describes intended uses, training-data categories, known limitations, and instructions for appropriate use and human review. Developers and deployers must retain records needed to demonstrate compliance for at least three years.

The act also creates candidate-facing duties. Deployers must provide clear point-of-interaction notice when consumers interact with a covered ADMT. After a covered ADMT makes a consequential decision that results in an adverse outcome, the deployer must provide a plain-language description of the technology’s role within 30 days. Consumers can request personal data, correction of factually incorrect personal data, meaningful human review, and reconsideration.

In operational terms, Colorado draws the map for the appeal desk.

January 1, 2027 sounds far away only if the employer has not counted implementation time. A multi-state company must identify which tools materially influence employment decisions, map the candidate-facing interactions, update notices, define human review, assign record owners, collect developer documentation, update contracts, train recruiters, build a response workflow, and test the process before the law’s duties begin.

That will not happen in December.

Colorado also makes the developer-deployer split explicit. The vendor has documentation duties. The employer has deployment duties. The candidate will not care which party created the missing file. If the employer cannot explain the AI interview’s role, the candidate will experience the failure as an employer failure.

Vendor evidence handoff cannot be optional.

The employer needs contract language that says what the vendor must provide when a candidate requests review or correction. At minimum, that includes technical documentation, known limitations, interview configuration, scoring explanation, transcript or structured record, audit artifacts, affected-candidate export, retention support, and a named support channel for regulated requests.

It also needs timing.

If the employer has 30 days to provide a post-adverse outcome description, the vendor cannot take 21 days to deliver a transcript and then another 10 days to explain the score. If the employer promises a human review within a week, the reviewer cannot wait for a support queue with no SLA. If the candidate asks to correct personal data, the vendor needs a correction path that reaches every downstream record where the wrong data was used.

The Colorado law is not the only relevant rule.

New York City’s Local Law 144 bars employers and employment agencies from using an automated employment decision tool unless it has a bias audit within one year of use, a public audit summary, and required notices to candidates or employees. The NYC Department of Consumer and Worker Protection also gives candidates a complaint path when the required audit, posting, or notice is missing.

California’s employment automated-decision regulations are tied to broader Fair Employment and Housing Act recordkeeping. The California Civil Rights Council’s final statement of reasons repeatedly discusses the four-year preservation requirement for employment records, including records used with or created from automated-decision systems. An AI interview dispute can therefore become a record-retention problem rather than a customer-support exchange.

In Europe, the EU AI Act’s Article 14 requires high-risk AI systems to be designed so natural persons can oversee them during use. The text says oversight should allow people to understand system limits, monitor operation, remain aware of automation bias, interpret outputs, disregard or override outputs, and interrupt the system when appropriate.

Those duties do not all say “candidate appeal queue.” They point to the same operating need.

There has to be a trained human with enough evidence, authority, time, and system access to review the AI-assisted decision.

If that person does not exist, the appeal right is theater.

Human Review Requires a Staffing Model

Recruiting leaders often describe human review as if it were a checkbox. It is a staffing model.

A meaningful reviewer needs five things.

First, the reviewer needs competence. A recruiter reviewing a software engineering screen may understand the role but not the assessment method. A hiring manager may understand the role but not the compliance rule. A legal reviewer may understand the risk but not the job context. The strongest model is usually shared: a trained recruiter or structured-interview reviewer owns the first review, with hiring manager input for role-specific judgment and legal escalation for regulated or high-risk issues.

Second, the reviewer needs independence. If the same recruiter who relied on the AI output is asked to rubber-stamp the appeal, the process will not feel credible. Independence does not require a separate legal department for every case. It does require a clear rule for when a second reviewer is assigned.

Third, the reviewer needs authority. A human who can only “take a look” but cannot override a score, reopen a candidate, request a new interview, correct the record, or pause the workflow is not reviewing the decision in any practical sense.

Fourth, the reviewer needs time. A rushed review can be worse than no review because it creates a record of procedural compliance without substantive attention. If the employer deploys AI interviews to handle volume, it must reserve review capacity for the predictable fraction of candidates who ask for another look.

Fifth, the reviewer needs evidence. A reviewer cannot assess a disputed AI interview from a final score alone. They need the transcript, rubric, scoring rationale, source materials, candidate correction, stage history, and any system limitations relevant to the case.

These are budget items.

The simplest staffing formula starts with expected AI interview volume, expected appeal rate, average review time, escalation rate, and response SLA. It then adds training, quality assurance, vendor support time, legal review capacity, and system administration.

Input	Example planning question
AI interview volume	How many candidates will complete AI interviews per month?
Appeal rate	What percentage will request human review, correction, or explanation?
Review time	How long does a trained reviewer need for low-risk and high-risk cases?
Escalation rate	How many cases require legal, accommodation, bias, or vendor escalation?
Vendor dependency	How many records sit outside the ATS and require vendor response?
SLA	How quickly must the employer acknowledge, review, decide, and respond?

This model will vary by employer. A high-volume retailer hiring seasonal workers may need a fast triage desk. A hospital hiring clinical staff may need more manager review and credential sensitivity. A public sector employer may need stricter records and procurement controls. A global enterprise may need different paths by jurisdiction.

The core point holds: AI interview adoption without review staffing is deferred cost.

SHRM’s State of AI in HR 2026 helps explain why this cost is easy to miss. The report found that legal and compliance functions primarily lead AI governance and oversight in 37% of organizations, while more than half of organizations do not involve HR directly or collaboratively in overall AI strategy and vision. SHRM also found that HR professionals at organizations using AI reported responsibility shifts more often than displacement.

The appeal queue is one of those responsibility shifts.

Recruiters do not disappear. Their work changes. They may spend less time manually screening resumes and more time reviewing AI-generated evidence, explaining decisions, correcting records, testing rubrics, and deciding when a human conversation should replace an automated screen. TA Ops may spend less time cleaning pipeline reports and more time managing appeal SLAs, vendor evidence requests, jurisdiction rules, and quality dashboards.

The change is material.

It asks HR to define work that software vendors may not price honestly. A vendor can show saved recruiter hours from automation. The buyer should ask how many reviewer hours are created by disputes, exceptions, corrections, accommodations, and evidence requests.

If the vendor cannot answer, the employer should run its own pilot with appeal accounting from the first day.

Vendor Evidence Becomes the Bottleneck

Most appeals will be won or lost on evidence speed.

The employer can have a good policy and still fail if the vendor cannot produce the file. The vendor can have a strong model and still damage trust if its explanation arrives too late. The recruiter can intend to review fairly and still make a poor decision if the AI output is not traceable to source material.

That bottleneck should shape procurement.

Before buying or expanding an AI interview product, the employer should ask the vendor for an appeal support package. The package should be part of the contract, not a sales slide.

It should cover:

transcript availability and retention settings,
score and rubric explanation,
prompt or interview-configuration versioning at a level the customer can use,
bias audit artifacts and protected-class testing scope where applicable,
candidate disclosure and opt-out support,
manual-review workflow support,
correction process for wrong candidate data,
affected-population export for incidents,
response SLA for evidence requests,
named support escalation for regulated requests,
credit or remediation terms when vendor failure creates rework.

The last line is the one procurement will care about.

If the vendor’s tool produces unclear candidate notices, inaccessible interviews, missing transcripts, broken handoff, poor score explanations, or faulty candidate-status synchronization, the employer will pay for rework. Recruiters will review more cases. Legal will spend more time. TA Ops will repair records. Candidate communications will need manual handling. Hiring managers may repeat interviews. The company may lose candidates.

That cost should not vanish into the employer’s operating budget.

This connects the appeal queue to the renewal desk. A vendor that claims AI interviews save recruiter time should be measured on net time: automation savings minus appeal handling, rework, exceptions, legal escalation, candidate drop-off, and evidence-support labor. A vendor that claims fairness should be measured on review outcomes: appeal volume, reversal rate, correction rate, accommodation issues, missing-evidence rate, and candidate response time.

From that point, the queue creates product metrics.

Metric	Why it matters
Appeal rate by role and source	Shows where candidates distrust or misunderstand the process
Evidence completion rate	Measures whether the vendor and ATS can assemble the review file
Time to first response	Determines whether candidates feel ignored
Time to human review decision	Tests the real capacity behind the policy
Reversal or correction rate	Shows whether AI outputs are being meaningfully challenged
Missing notice rate	Exposes process failures before they become legal issues
Vendor-caused rework hours	Turns support quality into renewal economics
Candidate continuation after appeal	Measures whether the process preserves trust

These metrics are uncomfortable because they reveal mistakes.

Mistakes are the point.

A low appeal rate may mean candidates trust the process. It may also mean candidates do not know how to appeal. A low reversal rate may mean the AI output is accurate. It may also mean reviewers lack authority or time. A fast response time may mean the process is efficient. It may also mean the review is shallow.

The dashboard needs context.

Sampling is how the team avoids mistaking a quiet queue for a healthy one. TA Ops should audit a subset of closed appeals, including those where the AI decision was upheld. It should compare candidate-facing explanations with internal evidence. It should check whether reviewers had the right file. It should look for patterns by role, location, source, vendor, recruiter, and hiring manager.

The work is unglamorous. It separates AI adoption from accountable operations.

Finance Will Ask Who Pays for the Second Look

AI interview appeals will create a budget conversation.

Reviewer labor comes first. A human review is not free because the human already works for the company. The cost includes recruiter time, manager time, TA Ops time, legal escalation, vendor support, system administration, training, and quality assurance.

Candidate delay creates the second budget line. A delayed appeal can cost the employer a candidate, especially in frontline roles. ICIMS’ Spring 2026 release framed frontline hiring around speed, mobile flow, and candidate drop-off. The company cited its frontline hiring data showing more than half of candidates abandon applications before completion and 32% drop off at the interview stage. ICIMS also said Frontline AI customers had seen up to 75% reduction in time to fill, up to 90% reduction in manual hiring tasks, and up to 10x more hires per recruiter.

Trent Cotton, ICIMS’ head of talent insights, pointed to a more practical operator lesson: frontline employers have to narrow attention to the roles and locations that matter, then move quickly with clear schedules, pay, and mobile-friendly steps. An appeal process that takes a week to find the transcript works against that operating model.

Speed sells.

Appeals slow the line unless they are designed into the line.

A retail manager who needs five cashiers next week does not care whether the delay comes from legal review, vendor support, transcript retrieval, or a recruiter queue. They see an unfilled shift. A warehouse site leader sees overtime. A hospital unit sees agency labor. A restaurant sees manager coverage. Finance sees the cost.

The third budget line is vendor support. If the employer has to pay for premium support to obtain evidence, that cost belongs in the total cost of ownership. If the vendor charges for data export, transcript retention, audit logs, or advanced compliance modules, the procurement team should attach those fees to the AI interview workflow, not hide them under generic platform spend.

The fourth budget line is rework. A candidate who receives a flawed AI screen may need a human interview. A role may need additional sourcing because candidates walked out. A hiring manager may need to repeat evaluation. A recruiter may need to explain the process to a candidate who no longer trusts the employer.

For that reason, the appeal desk should connect to vendor credits.

Not every appeal should create a refund. Many appeals will uphold the original decision. Some will reflect candidate disagreement, not system failure. But when the appeal exposes a vendor defect, missing evidence, broken disclosure, inaccurate transcript, wrong workflow configuration, or failure to meet support SLA, the contract should assign cost.

Procurement can write that in plain terms:

evidence request not fulfilled within SLA,
required transcript unavailable,
candidate notice not displayed because of vendor workflow error,
score explanation missing or unusable,
AI interview configuration inconsistent with agreed rubric,
candidate opt-out path broken,
data correction not propagated to downstream records,
affected-candidate export unavailable after incident.

Each of those failures creates labor for the employer.

The contract should say whether the vendor provides service credits, remediation support, no-cost review assistance, audit support, or termination rights after repeated failures. Without that language, the employer absorbs the defect while the vendor keeps the automation savings story.

Finance will eventually notice.

Workday’s fiscal 2027 first-quarter results show why the volume question is already real. Workday said more than 4,000 customers were using at least one organically developed Workday agent, and that its Recruiting Agent supported 14 million hiring processes in Q1, up 44% year over year. Those are not pilot numbers. They are operating-scale numbers.

At that scale, even a small review load becomes a staffing and finance line.

Buyers should therefore ask every AI interview vendor a simple question: what happens after the candidate disputes the result?

A vendor answer that stops at “the customer can review it” is not an answer.

The complete answer names the evidence, the workflow, the SLA, the human role, the correction path, the audit file, the support owner, and the cost allocation when the vendor caused the problem.

Ninety Days Later, the Queue Has to Prove It Worked

At first, an AI interview appeal desk has to handle one candidate fairly. After one hundred cases, it has to prove that the team learned something.

After 90 days, TA leadership should be able to answer concrete questions.

How many candidates completed AI interviews? How many received clear notice before the interview started? How many opted out? How many requested human review, correction, or explanation? How many appeals were acknowledged within the SLA? How many were upheld, revised, reopened, or escalated? How many required vendor evidence? How many vendor requests missed the SLA? Which roles, locations, sources, or recruiters generated the most appeals? How many candidates continued after appeal? How many withdrew?

Those questions turn the appeal queue into a management system.

They also reveal whether the AI interview product is improving hiring or merely moving work into a less visible place. If AI interviews reduce recruiter screens but increase candidate drop-off, appeal labor, legal review, and vendor support, the ROI story needs revision. If appeals are rare, evidence is complete, candidates receive quick responses, and reviewers occasionally overturn bad outputs, the tool becomes easier to defend.

The strongest appeal desk will not be the one that rejects every complaint. It will be the one that can show a fair process, a real human review, a complete file, a corrected record when needed, and a learning loop back into product configuration.

The weakest appeal desk will hide behind automation.

Candidates already understand more than employers assume. They know when an interview feels like a machine pretending to listen. They know when no one reads the answer. They know when the company never said AI was involved. They know when the rejection comes without explanation. They know when the process asks for trust without offering any proof.

AI interviewing can still be useful. A structured conversation available at any time may help candidates who would never get a recruiter screen. It may reduce arbitrary resume filtering. It may give hiring teams better early signal. It may help overloaded recruiters spend time on deeper judgment instead of calendar coordination and repetitive screens.

But that version of AI interviewing needs a second door.

One door lets a candidate enter the process through an AI interview. Another lets the candidate challenge, correct, or ask for human review when the first door fails.

If the second door leads only to a legal inbox, the process will be slow and narrow. If it leads only to a recruiter inbox, the process will be underpowered. The right door leads to Recruiting Ops: a queue with evidence, deadlines, trained reviewers, vendor obligations, candidate communication, and finance visibility.

Putting AI at the first screen carries that operating price.

This article provides a deep analysis of AI interview appeal operations in recruiting. Published June 7, 2026.