Arena Analytics

The prediction is the product. Arena does not sell access to a candidate network or a job board, and there is no version of the service that functions without the model: what a customer buys is a per candidate score for a named outcome, generated by machine learning models combining Arena proprietary data, external data and applicant data.

That is the distinction separating Arena from the marketplace vendors in this category, where the durable asset is the network and the model is an efficiency layer over it. The moat is the dataset question applies in part, since roughly one million unique applicants a year pass through the deployed base and that volume is not easily replicated, but here the dataset is an input to a model the buyer pays for rather than the thing being sold.

Arena scores candidates and, through Talent Rerouting, redirects lower fit applicants toward roles where they are predicted to succeed rather than discarding them. That is a real design choice, it reduces the cost of a wrong prediction to the person, and it is credited here. What is not published is the decision boundary that matters most: whether a low retention prediction can remove a candidate from consideration without a human reviewing the file, and if so at what score.

The asymmetry is the same one the index applies to payer side prior authorization, where recommending is safe and declining is not. A wrong positive prediction costs an employer a hire that does not last. A wrong negative prediction costs a person a job they would have kept. Publishing the screening threshold, and whether any candidate is ever eliminated without human review, would move this to A.

Arena names its bias mitigation technique specifically, adversarial fairness, which is a recognized method in the fairness literature rather than a marketing coinage, and it names the three input classes feeding the models as proprietary data, external data and applicant data. That is more than most workforce vendors disclose. Below that level the disclosure thins.

There is no model card, no published methodology or peer reviewed description, and no feature list, with inputs described only as scores of data points specific to each role and organization. Because models are fitted per role and per organization, a buyer also cannot tell how much predictive signal comes from Arena's cross customer data versus what is learned locally, which is the question that determines whether early performance holds. A published technical description of the model family and feature classes would move this to A.

Nothing identifies any party in the chain: no model or model family, no hosting arrangement and no sub processor list was located in two passes, and no retention position was found. The clinical form of this axis does not apply, since the company processes applicant and employee records rather than patient data, and the grade is for the domain equivalent rather than for an absent patient data posture. That equivalent is significant and unaddressed.

The company holds identified employment histories and prediction scores on roughly a million job applicants a year, including large numbers of people who were never hired and who have no continuing relationship with the employer that scored them.

That group is the sharpest version of the problem this index keeps finding: a rejected applicant has no contract, no account, and usually no knowledge that a score about them exists anywhere, so every mechanism that normally lets a data subject act, request, correct or delete is unavailable because they cannot know to use it.

Nothing states how long applicant records are retained, whether a person can see or contest the prediction attached to them, or whether data from one customer's applicant pool trains models serving another customer, which for a vendor selling across competing employers in the same labour market is a live commercial question too. Ask for all three in writing rather than assuming them.

Arena has named healthcare customers and a decade of quantified outcome claims, which puts it ahead of most of this category, and its post hire reports are a genuine accountability mechanism because the vendor commits to measuring what actually happened to the people it scored. The problem is that the published figures do not reconcile. The Retention Prediction page states a 21 percent reduction in first year turnover.

Company materials elsewhere describe a 35 percent turnover reduction and 19 million dollars in savings at a 20,000 employee hospital system, a median 38 percent turnover reduction with a claimed 100 percent success rate across deployments, and a 77.6 percent reduction in 90 day turnover at Mount Sinai. Those are different metrics over different periods, and none is presented with a denominator, a comparison group or a time frame that would let a buyer reconcile them.

No independent or peer reviewed evaluation was located. Graded C on the index precedent that scale of deployment does not substitute for evidence of benefit. Publishing the post hire reports in aggregate, with method and denominators, is the single change that would most improve this grade.

Arena processes applicant and employee records rather than patient data, so PHI stewardship does not apply in the provider sense. The C is not for the absent PHI posture.

It is for the domain equivalent, which is significant and unaddressed. Arena holds identified employment histories and prediction scores on roughly one million job applicants a year, including large numbers of people who were never hired and who have no continuing relationship with the employer that scored them. That group is the sharpest version of the problem, because a rejected applicant has no contract, no account and usually no knowledge that a score exists.

Nothing located states how long applicant records are retained, whether a person can see or contest the prediction attached to them, or whether data from one customer's applicant pool trains models serving another customer. Those three disclosures are the employment data analogue of what this index requires on PHI elsewhere, and buyers should require them in writing rather than assume them.

No patient data is in scope, so a HIPAA posture is not applicable in the provider sense and is rated accordingly rather than penalized. No BAA language was located on the vendor site, and that is the expected outcome rather than a gap, since an employer sharing its own workforce records with a vendor is generally not making a covered entity disclosure of PHI.

The document governing this relationship is the data processing agreement rather than a BAA, and buyers should read it for the retention and secondary use terms named in the stewardship row instead of looking for HIPAA coverage that does not apply here.

No SOC 2 Type II, ISO 27001 or equivalent attestation was located, and the site carries no trust center. The footer links only to terms of service and a privacy policy. This is graded rather than left unrated because applicability is not in question: Arena holds identified applicant and employee records and connects directly into Workday, ADP, UKG, Oracle and other systems of record, which is among the more sensitive integration surfaces a health system exposes to a third party.

The absence mirrors a pattern the index has documented in healthcare cybersecurity, where vendors that assess risk in others publish no attestation of their own. A published SOC 2 Type II would move this to B, and a trust center carrying subprocessor and residency detail would take it to A.

Arena is not a medical device and makes no clinical claim, so no FDA pathway applies and the missing clearance is not the reason for this grade.

The C reflects a regime that applies directly and is more consequential for this buyer than the device question would have been. A tool that scores candidates and shapes who gets hired is an automated employment decision tool under a growing body of regulation, including New York City Local Law 144, which requires an annual independent bias audit with published results, Illinois legislation governing AI in video interviews, and the Colorado framework for consequential decisions. No Local Law 144 audit or equivalent filing by Arena was located.

In covered jurisdictions that obligation sits with the employer, not the vendor, which is precisely why a buyer should obtain the audit before deployment. The duty does not transfer with the software.

This is the strongest bias disclosure located in this category so far, and it is still short of what a buyer needs. Arena names the technique rather than gesturing at fairness, describing a proprietary bias mitigation model built on adversarial fairness. It states that it removes protected characteristics and the signals related to them, which is the harder half of the problem, because proxy variables rather than explicit attributes are where disparate impact usually enters a hiring model.

It also commits to a measurable standard, stating that its recommended candidate groups exceed the EEOC four fifths threshold and often surpass 95 percent. Every one of those claims is self attested. No independent bias audit, no breakdown by protected class, no methodology and no adverse impact ratio by role or customer was located. Graded B because naming a method and a threshold is materially better than the category norm, which is silence. Publishing an independent audit, which employers in several jurisdictions must obtain in any case, would take this to A.

Naming a recognised technique rather than coining one is what carries this grade. The bias mitigation approach is identified as adversarial fairness, which is an established method in the fairness literature rather than a marketing term, so a reader can consult that literature, understand what the method does and, importantly, read its known limitations rather than accepting the vendor's characterisation. A named method can be argued with; a coined one cannot.

The three input classes feeding the models are also named, covering proprietary data, external data and applicant data. Held at C because nothing below that level is disclosed. There is no model card, no published methodology or peer reviewed description, and no feature list, with inputs described only as scores of data points specific to each role and organisation.

Because models are fitted per role and per organisation, a buyer also cannot tell how much predictive signal comes from cross customer data and how much is learned locally, which is the question that determines whether early performance holds as the local sample grows. No accuracy figure, no validation and no warranty, indemnity or remediation commitment was located.

The output decides whether a person is hired, so the error a candidate cares about is the false negative and it is unmeasurable by construction, since a rejected applicant's counterfactual performance is never observed. Ask for adverse impact testing results by protected class, and how the local and cross customer contributions are separated.

No EHR integration exists and none is required, because this product reads and writes to workforce systems rather than clinical ones, so the axis is assessed on the interoperability surface that actually applies. On that surface the coverage is good.

Arena names nine integrations spanning the major applicant tracking and human resources systems used in healthcare, including Workday, ADP, UKG, Paycom, iCiMS, TalentReef, Jobvite, Oracle and symplr, which means most health systems can adopt it without changing their system of record.

What is not published is the depth of any individual integration: whether scores write back into the applicant record, whether exchange is bidirectional and real time or batch, and how much implementation work each connector carries. Named integration logos are a claim of coverage rather than of depth, and buyers should ask which of the nine are productized and which are project work.

A second pass again found nothing published by the vendor on hosting model, tenancy, region or residency, and the earlier assessment's treatment of the aggregator derived server location as secondary sourced and unconfirmed stands.

Two facts from the second pass shape what the question actually is.

The first is the data path. The platform integrates with applicant tracking systems, so candidate records move from an employer's recruiting system into the vendor's environment for analysis and back. That makes the residency question one about what leaves the employer's estate rather than about a self contained application, and it means the answer matters to the employer's own privacy commitments to candidates as much as to the vendor's.

The second is scope. The technology is described as guiding hiring, promotion, training and compensation decisions across millions of job seekers, and as now supporting workforce boards and intermediaries. Public workforce entities bring requirements a private employer does not: state procurement standards, and in many jurisdictions public records obligations that can reach records held by a contractor. A single hosted estate serving both private employers and public bodies should be able to say how those are handled.

Tenancy deserves its own question here for a commercial reason as much as a regulatory one. The customer base is employers competing for the same workers in the same labour markets, and the underlying asset is identified employment history at scale.

Ask for the hosting region, the tenancy model, the subprocessor list, and what is transferred through the applicant tracking integration in each direction.

Arena publishes a pricing page and it contains no price. The page presents two segments, medium sized businesses and enterprise, each resolving to a contact form, with no rate, no unit and no stated mechanism. That is a contact gate presented as pricing, and it grades below vendors in the index that publish either a figure or the basis on which they charge. Two related disclosures are also absent.

The site offers an ROI calculator that frames the purchase in terms of avoided turnover cost, and the index applies added scrutiny to savings linked commercial models because the vendor is then measuring the number its own fee depends on. Arena also advertises a retention guarantee for candidates it rates as likely, but the terms, the remedy and the measurement method behind that guarantee are not published. Publishing the pricing unit alone, whether per hire, per applicant scored or per employee, would move this to B.

Coverage across care settings is genuinely broad and is not limited to nursing. Named deployments span academic and community acute care at Mount Sinai and MultiCare, senior living at Sunrise, Benchmark and Validus, and hospice and home based care at Curo, with company materials describing 486 healthcare facilities processing roughly one million applicants a year as of 2022.

Because the product predicts role level outcomes rather than clinical ones, it reaches the whole frontline workforce, including the environmental services, food service and support roles that carry the highest turnover and are largely ignored by the clinician focused vendors in this category.

Two caveats hold this at B. The facility count is four years old and no current figure was located, and Arena has extended into hospitality and quick service work, so healthcare specificity is a claim buyers should test against how the model is trained for their own roles.