Major

Automate Greenhouse Hiring-Stage Adverse-Impact Reviews

Turn Greenhouse hiring-stage data into a repeatable, reviewable adverse-impact monitoring workflow, with human oversight and clear limits on what the numbers mean.

Ritvij Gautam
HR compliance analyst reviewing anonymized Greenhouse hiring-stage audit data.

TL;DR: what you'll build. HR compliance teams using Greenhouse rarely check whether a hiring stage is screening out a protected group more than others, unless legal asks for it after the fact. This guide builds a quarterly, Major-hosted monitor: a deterministic workflow pulls Applications, Application Stages, and EEOC data from the Greenhouse Harvest API, calculates selection-rate ratios in code, and an agent writes a plain-language summary of what to look at. HR compliance reviews and approves every report before it goes anywhere.

The hiring audit problem and its process owner

Talent acquisition teams run candidates through Greenhouse stages every week: screen, phone interview, onsite, offer. Nobody routinely checks whether one demographic group clears a given stage at a meaningfully lower rate than another, because that analysis takes someone to export data, build a spreadsheet, and know the math.

The process owner here is HR Compliance, often working alongside Talent Acquisition operations. They are the ones accountable when a regulator or plaintiff's attorney asks whether the company's hiring funnel has been reviewed for disparate impact.

The EEOC has published guidance that is directly on point: Assessing Adverse Impact in Software, Algorithms and AI Used in Employment Selection tells employers that any automated or AI-assisted screening tool should be evaluated for disparate impact against protected groups under Title VII. The guidance is explicit that passing a four-fifths comparison does not, by itself, prove the hiring process is compliant.

That is the trap this article avoids. The goal is not to certify legality. It is to make a repeatable, documented review happen quarterly instead of only after a complaint lands, so HR compliance always has current evidence and can escalate early.

This use case sits in the AI agents for HR hub and the AI agents for recruiting function hub. If you already automated the Greenhouse-to-BambooHR handoff when a candidate is hired, see the sibling Greenhouse-to-BambooHR onboarding workflow for that different use case. This one is about reviewing the funnel before the hire, not onboarding after it.

What Greenhouse Harvest data can and cannot tell you

Greenhouse's Harvest API v3 is the only data source this workflow touches. Four resources matter for this review:

  • Applications: a candidate's record on a specific job, including the current stage id, a lifecycle status (in process, rejected, hired, or converted from a prospect), and rejection reason id, retrieved from the List applications endpoint.
  • Application Stages: the stage-history fact table. Each row is one visit to a stage with entered and exited timestamps, documented at List application stages. Greenhouse's own docs call this the canonical source for stage-history reporting.
  • Jobs: the hiring plan a job belongs to, used to scope the review to a specific requisition or department.
  • EEOC: row-level self-identification data (race, gender, veteran status, disability status) that candidates voluntarily submit. The List EEOC endpoint is explicit that this is sensitive PII gated behind a dedicated scope, separate from the aggregated, anonymized counts shown in Greenhouse's own in-app EEOC report.

Authentication uses OAuth 2.0. Harvest v3 supports a Client Credentials grant for custom, in-house integrations, where the token comes from Greenhouse's dedicated auth endpoint, or an Authorization Code grant for partner integrations distributed to multiple customers. Legacy Harvest v1/v2 used Basic Auth with a single API key and is being retired, so build any new integration on v3.

Pagination is cursor-based: pass the cursor string from the previous response's pagination header to get the next page, with a page size configurable from 1 to 500 results (default 100). Greenhouse's docs do not publish a specific rate-limit number for Harvest v3 in the pages reviewed, so build in backoff rather than assuming a fixed ceiling.

Permissions and the EEOC scope

Site admins create Harvest credentials in Greenhouse's Dev Center and assign access to specific endpoints or scopes, per the credential management guide. You can also restrict a credential to specific IP addresses.

Treat the EEOC scope as its own decision, not a default. Greenhouse's own documentation calls out that EEOC responses are sensitive PII tied to a specific candidate and application, gated on a dedicated scope with partner permissions typically restricted. Only grant it if your organization's legal and privacy policy supports pulling row-level demographic data into a reporting app, and scope the credential to read-only.

A webhook analogy helps here: Greenhouse does support webhooks for application and candidate events, but this workflow is a scheduled pull, not an event stream. A webhook is a doorbell, useful when you want to react the moment something changes. A quarterly compliance review wants a consistent snapshot of a whole period, so a scheduled poll across the quarter's applications is the right shape, not a doorbell ringing on every stage move.

Split the architecture: app, agent and workflow

Three pieces do different jobs here, and keeping them separate is what makes the review auditable instead of a black box.

The app is the control surface and system of record. It holds every run's inputs, the calculated rates, the flags, and the reviewer's decision, so a compliance officer can open any past quarter and see exactly what was pulled and concluded.

The workflow does the deterministic work: paginate through Greenhouse, join applications to their stage history and EEOC responses, validate the cohort is complete enough to calculate, and compute the actual selection-rate arithmetic. None of that needs a model, and all of it should run the same way every time.

The agent has exactly one job: take the workflow's calculated flags and missing-data caveats and turn them into a plain-language summary a non-statistician can read. It never recalculates a rate, never decides whether a flag means discrimination occurred, and never touches a candidate record.

Architecture diagram showing Greenhouse Harvest API feeding a Major workflow and app, with an agent writing the summary and HR compliance approving the output

Quarterly Greenhouse adverse-impact monitor: workflow, app, agent and human review

This is the same reason Major pushes repeatable work into an app rather than having a model re-derive the math every run: the selection-rate formula does not change quarter to quarter, so it should be code the agent runs, not a calculation the agent re-reasons through each time.

Build the quarterly monitor on Major

Here is the concrete build, in order.

  1. Connect Greenhouse as a Major connector using a Harvest v3 OAuth 2.0 client-credentials grant, scoped to read-only access on Applications, Application Stages, Jobs, and (only if your policy allows it) EEOC. Store the client key and secret in Major's credential proxy, never in workflow code. You should see the connector listed as active with the specific endpoints it can call.
  2. Build the app with a managed database holding four tables: review runs (quarter, run date, status), stage cohorts (job id, stage id, cohort window, group counts), selection flags (cohort id, group, selection rate, comparison rate, four-fifths ratio, data-quality status), and report reviews (reviewer, decision, timestamp, notes). You should see these tables in the app's data model with the right columns before moving on.
  3. Prompt Major to build the app's UI: a run list, a cohort detail view, and a flag review screen where HR compliance can read the agent's summary next to the underlying counts.
"Build an app with a table of quarterly review runs, each linking to its stage cohorts and flags. Each flag should show the stage name, group counts, selection rate, and a status of pending, approved, or needs-more-context. Include a text field for the reviewer's notes."
  1. Add the agent and restrict its permissions to read from the app's calculated tables and write only a draft summary field, never the underlying numbers and never anything in Greenhouse.
"Given a list of flagged stage comparisons with group counts and selection rates, write a two-paragraph plain-language summary per flag. State the comparison, the ratio, and any data-quality caveat (small sample, low EEOC response rate). Do not conclude whether this is discrimination. End every summary with a note that EEOC guidance treats the four-fifths comparison as a screening signal, not a legal determination."
  1. Set the workflow trigger to run quarterly on a fixed schedule, with a manual rerun option restricted to an authorized reviewer role. You should see a scheduled trigger plus a manual "Run now" action gated by permission.
  2. Test the workflow against a sandbox or a single department's data first, and confirm the stage cohorts, selection-rate math, and the agent's summary all match a hand-calculated example before turning it on org-wide. You should see a completed test run with reviewer sign-off recorded in the report reviews table.
  3. Roll out with role-based access: Talent Acquisition can see cohort counts but not row-level EEOC data, HR Compliance can see everything and approve or reject, and the agent's write access is limited to its own summary field. You should see distinct permission levels reflected in who can open which screen.

A worked example, for illustration only

Say a quarter's onsite-interview stage had 200 White candidates and 60 Black candidates reach that stage. Of those, 90 White candidates (45%) advanced past onsite, versus 15 Black candidates (25%). The ratio of 25% to 45% is about 0.56, below the four-fifths (0.80) threshold EEOC's Uniform Guidelines Q&A describes as a rule of thumb for a closer look.

This is a fictitious number set built to demonstrate the math, not a finding from any real dataset. A ratio below 0.80 is a prompt for HR compliance and counsel to look closer at that stage, not proof of anything on its own. Sample size, interviewer variance, and job requirements all matter before anyone draws a conclusion.

Governance and IT control

Every piece of this workflow should be inspectable, not just functional.

  • Scoped credentials. The Greenhouse connector's OAuth credential should carry only the endpoints this workflow needs. If EEOC access is granted, it should be its own credential, separately revocable from the Applications and Stages credential.
  • Role-based access in the app. Only HR Compliance roles should see row-level EEOC data if it is pulled at all; Talent Acquisition sees aggregated counts.
  • Audit logs. Every workflow run, every agent summary, and every reviewer decision should be logged with a timestamp and actor, so a later audit can reconstruct exactly what the report said and who approved it.
  • SSO. Access to the app should go through your organization's single sign-on, not a separate login, so offboarding an employee immediately removes their access to candidate data.
  • Human approval before distribution. No report leaves the app until an HR compliance reviewer approves it. If a Slack notification is used at all, it should fire only after approval, and it should never carry candidate-level or row-level demographic detail, only the fact that a quarterly report is ready for review in the app.

If a tool in your stack is Plex or UKG rather than Greenhouse, Major's public case study on Jefferson Southern shows the same governance pattern applied to a Tier 1 automotive supplier: live OEE dashboards, part-shortage tracking, and 24/7 automated shift workflows, all built on Major with the same permissioned, auditable app layer.

Failure handling and edge cases

A few things will go wrong in production, and the workflow needs defined behavior for each.

  • API errors and rate limits. Greenhouse does not publish a specific Harvest v3 rate-limit number in the docs reviewed. Build exponential backoff on any 429 or 5xx response, and alert the workflow owner if a run fails to complete within its window rather than silently truncating the quarter's data.
  • Partial syncs. If the workflow is interrupted mid-pagination, it should resume from the last successful cursor rather than restarting, and the run should be marked incomplete in the review runs table until every page is confirmed.
  • Duplicate events. Because this is a scheduled pull rather than a webhook stream, duplicate application rows are unlikely, but the workflow should still upsert on application id rather than insert blindly, in case a rerun overlaps a prior run's window.
  • Small cohorts. A stage with only a handful of candidates in a given group cannot support a meaningful selection-rate comparison. The workflow should flag these as data-quality: small sample rather than calculating and displaying a misleading ratio.
  • Missing or incomplete EEOC data. Candidates can decline to answer any EEOC question. If response rates for a cohort fall below a threshold your HR and legal team sets, the report should say so explicitly rather than presenting an incomplete picture as a complete one.
  • What the agent must never do. The agent never recalculates a selection rate itself, never flags or unflags a candidate, never writes to Greenhouse, and never issues a compliance conclusion. Every output it writes is a draft that a person reviews before anyone outside the app sees it.
Diagram showing data-quality gates for small cohorts and missing EEOC data before any selection-rate comparison reaches the HR compliance reviewer

Exception routing for small cohorts and missing EEOC data

What to measure

Track a small set of process metrics the HR compliance owner already cares about:

  • Review cycle time: days from quarter close to a signed-off report.
  • Stages flagged per quarter: how many stage comparisons fell below the four-fifths threshold and required a closer look.
  • Data-quality exceptions: how many cohorts were too small or had EEOC response rates too low to calculate confidently.
  • Manual touches per run: how many times a human had to intervene in the extraction or calculation before the workflow completed cleanly.
  • Time to reviewer decision: how long a flagged report sits before HR compliance approves, requests more context, or escalates.

Before you go live

  • [ ] Greenhouse OAuth credential scoped to read-only on Applications, Application Stages, Jobs, and (if approved) EEOC
  • [ ] EEOC access decision documented and signed off by legal/privacy, separate from the base credential
  • [ ] Test run completed against a single department with hand-verified selection-rate math
  • [ ] Small-cohort and missing-data thresholds agreed with HR and legal
  • [ ] Role-based access configured so only HR Compliance sees row-level demographic data
  • [ ] Human approval gate confirmed before any report leaves the app
  • [ ] Audit logging verified for every run, agent summary, and reviewer decision
  • [ ] Owner named for quarterly review and for responding to a failed or incomplete run

More hr AI automation use cases

Start building on Major

Build a Greenhouse hiring-stage monitor where repeatable calculations run as code, each review retains its evidence and history, and an HR owner approves every interpretation.

Related articles

About the author

Ritvij Gautam is Head of Growth at Major.

Last reviewed November 8, 2026 by Ritvij Gautam.

Frequently asked questions

Does Greenhouse's Harvest API support automated hiring-stage reporting?
Yes, within limits. The Harvest API v3 exposes Applications, Application Stages, Jobs, and a separately scoped EEOC resource for self-identified demographic data. Any integration should confirm exact field availability, OAuth scopes, and account-level access in Greenhouse's current documentation before building, since not every field discussed in planning is guaranteed to be enabled on a given account.
How does the four-fifths rule apply to hiring selection rates?
The four-fifths rule compares the selection rate of one group to the highest-selecting group; a ratio below 0.80 is a traditional screening signal under the Uniform Guidelines. EEOC's own Q&A guidance treats it as a rule of thumb for closer review, not a safe harbor. Context, sample size, and job relevance all affect whether a flagged ratio actually indicates a problem.
Can an AI agent decide whether a hiring process is discriminatory?
No. An agent can calculate ratios, summarize flags in plain language, and surface data-quality caveats, but determining whether a practice is unlawfully discriminatory requires qualified HR, legal, and often statistical review. The agent in this workflow is restricted to drafting a summary; a human always makes the interpretation.
What Greenhouse API authentication should a new integration use?
New integrations should use Harvest API v3 with OAuth 2.0, either a Client Credentials grant for custom in-house integrations or an Authorization Code grant for partner integrations distributed to multiple customers. Legacy Harvest v1/v2 Basic Auth API keys are being retired, so avoid building new integrations against that version. Scope any credential to only the endpoints the integration needs.