HeyMilo is an AI-assisted recruiting platform that can screen candidates through forms, SMS, resume analysis, voice, and video, then return reports to an applicant tracking system. It is most relevant to high-volume teams with a standardized first-stage screen. It should not be treated as an autonomous hiring authority, a validated substitute for every interview, or proof that a candidate cheated.

The product has expanded beyond the voice-first description in the previous article. This review reflects public materials available on September 13, 2026 and separates company claims from independently documented integration and employment-law requirements.

Direct answer: what does HeyMilo do?

HeyMilo’s current product page describes configurable agents for sourcing, eligibility screening, voice and video interviews, SMS and forms, scenario assessments, scheduling, reporting, and suspected-cheating signals. The same page says reports can contain scores, rationales, competency breakdowns, transcripts, recordings, and shareable PDFs.

These are vendor-described capabilities. Buyers should verify which modules are generally available, included in a quote, supported in their region, and enabled for the chosen ATS. A public demo does not establish accuracy, reliability, accessibility, or legal suitability for a particular role.

The strongest use case is a repeatable first screen

The product is most plausible where recruiters ask the same job-related questions many times and need candidates to respond outside office hours. Examples include basic eligibility, license or shift confirmation, structured experience questions, and a documented handoff for recruiter review.

The platform is less suitable when the stage requires relationship building, sensitive discussion, negotiation, live clarification, or observation that cannot be fairly captured through a remote automated interaction.

A deployment should begin with a workflow decision, not a channel decision:

Hiring needLowest-risk starting mode
Objective eligibility questionsForm or SMS
Short job-related explanationsVoice with transcript
Demonstration or presentationVideo with clear rubric
Ambiguous or sensitive discussionHuman interview
Disability or technology barrierAccessible alternative

Using video because it is available can add candidate burden without adding valid evidence.

Structured questions still need human design

HeyMilo says teams can configure questions and evaluation criteria. That is useful because consistency starts with a written job analysis and rubric. It does not make the resulting score objective by default.

Before activating AI scoring, define:

  • the job behavior each question measures;
  • acceptable evidence and anchored score levels;
  • attributes the model must ignore;
  • conditions that trigger human review;
  • prohibited automatic actions;
  • an accommodation and appeal path;
  • the period after which the rubric must be revalidated.

Generated follow-up questions should stay within the approved competency and avoid protected, medical, family, or other non-job-related topics. Review transcripts and recordings rather than relying on a summary when the result affects progression.

Scores are recommendations, not decisions

The product page presents detailed reasoning for candidate scores. An explanation can help a recruiter inspect the output, but fluent reasoning does not prove that the score is correct or legally job-related.

The US Equal Employment Opportunity Commission’s history of its Artificial Intelligence and Algorithmic Fairness Initiative states that existing employment-discrimination laws apply when software participates in hiring. The federal ADA guidance on hiring technology also warns that automated systems can screen out qualified people with disabilities or seek disability-related information.

Recommended production policy:

  • no automatic rejection based solely on an AI score;
  • a trained reviewer can inspect source evidence and override the result;
  • overrides and reasons are logged;
  • employers measure selection outcomes and error types by role;
  • candidates receive a clear way to request an alternative process.

Human review must be substantive. A recruiter clicking approve on every recommendation is not an effective control.

ATS integration is a material product feature

HeyMilo publishes an integration catalog covering more than 20 ATS partners and describing triggers, candidate intake, screening, and report writeback. The page is a vendor inventory, so integration depth must be verified per system.

Greenhouse independently documents its HeyMilo assessment integration. The documented workflow adds HeyMilo to an interview plan and returns assessment status and results to the Greenhouse candidate record.

For any ATS, test:

  1. the exact stage or event that triggers an invitation;
  2. duplicate events and candidate identity matching;
  3. job and rubric version changes;
  4. missing email, phone, or consent;
  5. report and recording permissions;
  6. retry and reconciliation after an outage;
  7. deletion in both systems;
  8. credential rotation and revocation.

Keep the ATS as the authoritative source for the application and disposition.

Candidate experience needs its own acceptance gate

HeyMilo says it operates across multiple languages and reports a 4.6 out of 5 candidate-satisfaction score. That figure is a vendor metric; the public page does not provide the survey population, response rate, geography, role mix, or question wording.

Measure experience in the buyer’s pilot:

  • invitation delivery and start rate;
  • completion and abandonment by mode and device;
  • median time to complete;
  • technical-support contacts;
  • accommodation requests and resolution;
  • candidate complaint rate;
  • post-process survey response and score;
  • differences by language, location, disability pathway, and role where lawful.

Tell candidates that AI is involved, what data is collected, how the output is used, who reviews it, how long it is retained, and how to request help or an alternative.

Accessibility cannot be inferred from conversational design

A voice or video interaction may create barriers for candidates with speech, hearing, vision, cognitive, motor, bandwidth, device, or language constraints. A text option is not automatically equivalent if the questions, time limits, or scoring change.

Test keyboard navigation, screen-reader behavior, captions, transcript correction, time extensions, device switching, low-bandwidth recovery, and human alternatives. Include people who use assistive technology in acceptance testing.

Accommodation requests should not be exposed to hiring managers beyond what is necessary to provide the adjustment. The alternative process should produce comparable job-related evidence without penalizing the candidate.

Anti-cheating signals are high-risk claims

HeyMilo says it uses active proctoring, an AI classifier, and other signals to produce a likelihood or trust score. That is a detection claim, not proof of misconduct. False positives can arise from assistive tools, eye movement, a second screen, network problems, rehearsed answers, non-native speech, or an unfamiliar environment.

Do not auto-reject on a flag. Require human review of the underlying evidence, disclose prohibited behavior before the assessment, allow a candidate explanation, and provide a fair retest or live verification path. Track confirmed and overturned flags by role and candidate group.

Procurement should ask for validation data, thresholds, calibration, known limitations, model-change controls, and the exact recordings or telemetry used. A single “cheat score” is not a defensible conclusion.

Responsible-AI disclosures are useful but vendor-authored

HeyMilo’s responsible AI page describes human review, transcripts and recordings, configurable model providers, deletion, limited prompt-log retention, a data-protection function, and SOC 2 claims. These statements help frame diligence, but buyers should request the current audit report, system scope, exceptions, and remediation evidence.

The page says standard configurations have undergone EU AI Act classification analysis. That is not a regulatory approval. Employment AI may be high risk depending on use, and obligations can differ between provider and deployer. Counsel should review the configured workflow and jurisdictions rather than relying on a general alignment statement.

The subprocessor chain must match the configuration

HeyMilo publishes a subprocessor page with service purposes and processing locations. It says some AI providers process data transiently and do not train on it. Verify those terms in the DPA and provider-specific configuration.

Ask for a field-level data-flow map covering resumes, job descriptions, contact data, audio, video, transcripts, derived scores, fraud telemetry, reports, logs, and backups. Confirm:

  • which modules invoke which provider;
  • storage and processing region;
  • default and configurable retention;
  • model-training and human-review terms;
  • encryption and key ownership;
  • support-access controls;
  • deletion from active systems and backups;
  • notice before a material subprocessor or model change.

Audio and video are sensitive operational records even when they are not labeled biometric data.

Pricing for core recruiting is quote-based

HeyMilo’s public recruiting pages invite buyers to book a demo and do not present a complete standard rate card for the core screening platform as of September 13, 2026. The separately published HeyMilo Red pricing applies to an interview-scanning product and should not be assumed to price the recruiting-agent suite.

Request a quote broken down by candidates, interview minutes, SMS, phone, video storage, ATS connector, implementation, languages, AI modules, support, and overage. Ask whether failed or abandoned interactions are billable and whether historical recordings incur continuing storage cost.

Compare total cost per completed, reviewable screen and per qualified candidate advanced, not price per invitation.

A consequence-based pilot

Start with one or two high-volume roles and a low-consequence use such as structured evidence collection, not automatic disposition.

  1. Build the rubric from a documented job analysis.
  2. Run security, privacy, accessibility, and integration tests in a sandbox.
  3. Score a historical set with known human-reviewed evidence.
  4. Launch prospectively with parallel human review.
  5. Compare completion, agreement, false-positive, override, support, and progression rates.
  6. Review differences across relevant candidate groups where lawful.
  7. Test outage recovery, deletion, export, and model rollback.
  8. Expand only after written acceptance thresholds are met.

Do not optimize solely for recruiter hours. Faster screening can create harm if it reduces valid candidate access or increases unexplained rejection.

Best fit and poor fit

HeyMilo is most plausible for high-volume employers and staffing teams with stable first-screen questions, a mature ATS, strong recruiting operations, and capacity to monitor model-assisted decisions. Multichannel delivery may help when a single interface excludes part of the applicant pool.

It is a weaker fit for executive search, sensitive relationship-led roles, unstructured hiring, low applicant volume, or teams seeking an AI system to make final decisions. It should not be deployed where the organization cannot offer accommodations, review underlying evidence, or investigate errors.

What remains unknown

Public sources do not establish independent scoring validity by role, bias-audit methods and results, anti-cheating precision, a standard core-platform price, customer-level ROI, or production reliability for every listed connector. HeyMilo’s volume, cost-reduction, and satisfaction numbers remain vendor claims without a published independent methodology.

The product can standardize and scale early evidence collection. Whether it improves a particular hiring process must be demonstrated with the employer’s candidates, rubric, ATS, and governance controls.