Blank?page job posts drifted; governed AI drafted compliant posts from HR pay bands and legal boilerplate to Greenhouse

Overview

A fintech startup rewrote job descriptions from scratch for every role. Language drifted, pay ranges weren’t aligned to approved bands, and Legal flagged inconsistent disclaimers and remote?work terms. Intelligex implemented a governed AI assistant connected to a library of approved competencies, leveling criteria, and pay bands. Drafts ran through bias and compliance checks, routed to Compensation and Legal when needed, and published to Greenhouse only after approvals. Content became consistent and compliant, hiring managers stopped reworking drafts, and every posting carried a clear audit trail—while Greenhouse, HRIS, and collaboration tools stayed in place.

Client Profile

• Industry: Financial technology (consumer and enterprise products)
• Company size (range): Fast?growing startup with centralized People Ops and distributed hiring managers
• Stage: Greenhouse as the Applicant Tracking System (ATS); leveling and competencies stored in docs; pay bands maintained in HRIS and spreadsheets; ad hoc Legal review of postings
• Department owner: Human Resources & People Ops (Talent Acquisition, HRIS, and Total Rewards)
• Other stakeholders: Hiring Managers, Legal/Compliance, DEI, Employer Brand/Marketing, Security/IT, Finance, Internal Audit

The Challenge

Every new role started with a blank page. Hiring managers copied language from old postings or competitors and modified it to fit. Competencies varied across teams for the same level, pay ranges weren’t consistently tied to location or job family, and boilerplate clauses changed from post to post. Legal found incorrect at?will statements, missing equal opportunity language, and inconsistent data?handling requirements for regulated functions. DEI flagged phrasing that could discourage qualified applicants.

Reviews happened late. Talent partners requested changes after postings went live, Compensation chased down correct bands, and Legal asked for edits during offer stages. Employer Brand struggled to keep tone and format consistent. Drafts pinballed between email, docs, and the ATS, so no one could show a single, authoritative record of who approved what and when.

Why It Was Happening

There was no canonical content source or enforcement. Competencies and leveling lived in slide decks and wiki pages, pay bands in spreadsheets, and boilerplate in an old doc template. Greenhouse accepted free?form text, so the path of least resistance was to paste whatever seemed close. Owners of competencies, bands, and legal text were known, but their approvals were not tied to the creation flow. Without a governed library and gates, drift accumulated with each new post.

The Solution

Intelligex delivered a governed AI assistant embedded in the job creation flow. The assistant generated drafts using approved competencies, leveling criteria, and tone guidelines, selected the correct pay range by location and job family, and ran bias and compliance checks before routing for approval. Legal and Compensation approvals were triggered only when policy dictated (for example, new job families, regulated functions, or location?specific clauses). Approved drafts published to Greenhouse via API with a complete audit trail. The design used retrieval?augmented generation to ground drafts in the client’s library (see RAG concepts), integrated with Greenhouse through the Harvest API, and aligned bias/compliance reviews to current EEOC guidance on AI use in employment (EEOC AI).

• Integrations: Greenhouse for posts and updates; HRIS as the source of pay bands and locations; content library in the client’s wiki/drive; identity for role?based permissions; collaboration tools for draft review; audit logs to the compliance repository.
• Governed content library: Approved competencies by job family and level; tone and brand guidelines; legal boilerplate by jurisdiction; remote/hybrid standards; leveling rubric mapped to role scopes.
• Draft generation: Assistant assembled duties, qualifications, and nice?to?haves from the library; included pay ranges and benefits based on location and level; formatted to brand standards.
• Bias and compliance checks: Detectors for exclusionary phrasing and degree?only requirements; validation of EEO and at?will language; location?specific clauses for pay transparency and remote work; flagged items surfaced with suggested rewrites.
• Approval gates: Compensation approval when new or adjusted bands were used; Legal approval for new job families, regulated functions, or new jurisdictions; maker?checker for exceptions; approvals and comments recorded with the draft.
• Publishing controls: One?click publish to Greenhouse after approvals; environment scopes for internal vs public postings; version history preserved; rescind and replace flows maintained links to candidates.
• Dashboards and evidence: Queue of drafts, approval status, exceptions, and content health; library freshness and usage; exportable packets per posting with inputs, checks, approvals, and final content.
• Security and privacy: Role?based access; minimal exposure of comp data in drafts; secrets stored in the enterprise vault; immutable logs of prompts, sources, and outputs.

Implementation

• Discovery: Collected existing postings, competency frameworks, tone guides, and legal boilerplate; inventoried pay bands and location rules; mapped current Greenhouse workflow; reviewed DEI and compliance concerns; gathered audit requirements for approvals and versioning.
• Design: Defined the content schema (competencies, responsibilities, qualifications, benefits); authored retrieval rules and grounding sources; mapped pay band logic by job family and location; defined bias and compliance checks; set approval criteria and exception paths; planned audit logs and dashboards.
• Build: Indexed the content library; connected to HRIS for bands and locations; implemented draft generation grounded in approved content; configured bias/compliance validators; built approval workflows and maker?checker rules; integrated publish/update to Greenhouse; enabled logging and evidence exports.
• Testing/QA: Ran in shadow mode, generating drafts alongside manual ones; compared outputs to legal and brand standards; validated pay ranges and location clauses; tuned detectors and phrasing; piloted with select job families and hiring teams.
• Rollout: Enabled the assistant for high?volume roles first; expanded to regulated and location?sensitive postings after approvals proved reliable; kept manual authoring as a controlled fallback; tightened gates as adoption grew.
• Training/hand?off: Delivered guides for Talent and hiring managers on generating, editing, and approving drafts; published style and compliance checklists; updated SOPs for posting and exceptions; transferred ownership of the library, checks, and dashboards to HRIS and Legal under change control.
• Human?in?the?loop review: Established a content council to review flagged language, evolving legal requirements, and library gaps; recorded decisions with rationale and effective dates; updates flowed back into competencies, boilerplate, and checks.

Results

Drafts started consistent and stayed compliant. Hiring managers selected the role and level, and the assistant produced a posting aligned to the competency framework, correct pay range, tone, and location clauses. Bias and compliance checks were addressed in context, and approvals were triggered only when policy required them. Legal and Compensation stopped chasing late edits, and Talent shifted time from rewriting to recruiting.

Governance became straightforward. Each posting carried a record of sources, checks, and approvals, so auditors and leadership could see how content met policy. Updates to competencies, benefits, or legal text flowed into new drafts automatically after change control. Greenhouse remained the ATS, HRIS remained the source for bands, and the team kept its collaboration tools—the new layer connected them with generation, checks, and approvals.

What Changed for the Team

• Before: Job descriptions started from scratch. After: Drafts were generated from approved competencies, leveling, and tone.
• Before: Pay ranges and clauses were inconsistent. After: Bands and location terms pulled from HRIS and legal boilerplate automatically.
• Before: Bias and compliance were caught late. After: Checks flagged issues during drafting with suggested language.
• Before: Legal and Compensation reviewed every post. After: Approvals triggered only when policy dictated, with clear rationale captured.
• Before: Brand voice varied by team. After: Formatting and tone aligned to one standard without manual edits.
• Before: Little audit trace. After: Each posting carried sources, checks, approvals, and versions in one record.

Key Takeaways

• Ground AI on approved content; use your competency library, leveling, and boilerplate to prevent drift.
• Automate pay and location logic; pull ranges and clauses from HRIS and legal sources rather than manual edits.
• Check bias and compliance before publish; make issues visible during drafting with suggested rewrites.
• Gate selectively; trigger Legal and Compensation approvals only when policy requires them.
• Keep humans in the loop; maintain content owners, change control, and exception paths.
• Integrate, don’t replace; keep Greenhouse and HRIS—add generation, validation, and approvals around them.

FAQ

What tools did this integrate with? Drafts published and updated via Greenhouse using the Harvest API. Pay bands and locations came from the HRIS. The assistant grounded on the client’s wiki/drive library and followed retrieval?augmented generation RAG patterns. Bias and compliance checks aligned to guidance from the EEOC, and audit logs flowed to the compliance repository.

How did you handle quality control and governance? Competencies, boilerplate, and prompts lived under change control with owners and rationale. Drafts recorded sources, checks, and edits. Maker?checker approvals applied to new job families, regulated functions, and new jurisdictions. Bias and compliance detectors flagged issues with suggested fixes, and exceptions required reason codes and approvals.

How did you roll this out without disruption? The assistant generated drafts in shadow mode alongside manual ones for select roles. Legal, Compensation, and Talent compared outputs and tuned the library and checks. Publishing remained manual until confidence grew. Coverage expanded in waves, and manual authoring remained a fallback for edge cases.

How were pay bands and locations handled? The assistant queried HRIS for the correct band based on job family, level, and location. Location rules added pay transparency and remote/hybrid clauses where required. Any deviation from the default band triggered Compensation review with rationale captured in the record.

How did bias and compliance checks work? Drafts were scanned for exclusionary phrasing, degree?only requirements without business need, and missing EEO and at?will language. Location?specific clauses were validated. Flags appeared in context with suggested rewrites, and Legal reviewed only when policy dictated or an exception was requested.

Did this replace recruiters or hiring managers? No. Hiring managers and Talent remained authors and approvers; the assistant provided grounded drafts and guardrails. Owners edited content, accepted or rejected suggestions, and approved exceptions. The goal was consistency and compliance, not automation without oversight.

How were updates to competencies or legal text propagated? Changes to the library went through change control. After approval, new drafts used the updated content automatically. Existing live postings could be refreshed through a controlled update flow that preserved candidate links and audit history.

What about employer brand and tone? Brand guidelines were part of the library. Drafts followed the approved voice and formatting, and Employer Brand reviewed templates periodically. Variations for technical, go?to?market, or operations roles were supported through structured style profiles, all under governance.