Eightfold AI

The API documentation states the integration purpose in its first sentence on authentication: these APIs can be used to "inject and fetch data from Eightfold and other ATS, HRIS and LMS ecosystems".

ATS-system-scoped endpoints at /core/ats-systems/{systemId}/ats-candidates and /ats-positions take POST, PUT and PATCH, so records are written and not only read, and the entity schema carries external identifiers throughout (atsEntityId, atsJobId, atsTitle, extDemandId, extOrgUnitId), which is what two-way synchronization with a system of record looks like. A calendar endpoint sits alongside them.

Coverage spans applicant tracking, HRIS, learning management, calendar and identity, and the vendor names Oracle Fusion Cloud Recruiting as a certified partnership, with distribution through the Microsoft commercial marketplace. The public API itself is the route for adding anything else.

Sourceapidocs.eightfold.ai/docs/getting-started and eightfold.airead 2026-09-08

Work passes between named agents with different jobs. Candidate Agent carries one continuous thread across role discovery, questions, application initiation, scheduling, reminders and status updates, then hands off into the interview process, where AI Interviewer conducts the structured interview and AI Interview Companion supports the human-led round.

The portfolio spans Talent Agents, AI Interviewer, AI Interview Companion and Career Coach across both Talent Acquisition and Talent Management, with TalentForge available to build further agents on the same intelligence layer. The handoff is documented as a lifecycle sequence rather than a named orchestration runtime, and no branching, conditional routing or customer configurable flow builder is published.

Sourceeightfold.ai/products/talent-agents and the Candidate Agent and interview journey releasesread 2026-09-08

Two facts are worth separating. Eightfold's corpus of roughly 1.6 billion career trajectories and 1.6 million skills is the vendor's own data, absorbed into its models, and the trust page confirms model improvement runs on anonymized aggregated data rather than identifiable customer data.

The grounding that matters to a buyer is the Digital Twin of the customer's own workforce: their skills, capabilities and the work their people actually perform, maintained against enterprise HR data and queried by the agents at run time. It persists between runs, scales far past a context window and stays queryable, and it is what makes internal mobility, succession and resource-management recommendations possible.

The API shows this is a maintained store: profile, position, course and insights entities are individually addressable, listable and filterable through a documented query syntax.

Sourceeightfold.ai/products, the interview journey release and apidocs.eightfold.airead 2026-09-08

The trust page states human in the loop, not automated hiring decisions: AI supports decisions and people make them, and the AI Interviewer does not make hiring decisions. The EU AI Act readiness program names human oversight as a requirement for high-risk hiring use, and AI subsystems are held inside the same permission boundaries as the platform, so an agent can only reach what its role permits.

That reserves the final decision for a person, but there is no review and approve step over the agent's own actions, and the agent does act on its own, since AI Interviewer conducts structured candidate interviews at scale with no documented gate before it interviews a real person. No approval queue, confidence threshold, escalation rule or hold is published. Marketing lines such as helping recruiters move up to 5x faster "all with your oversight" are positioning, not a mechanism.

Sourceeightfold.ai/trust/security and eightfold.ai/products/ai-interviewerread 2026-09-08

Both certifications and access controls are documented on the vendor's trust page. Access rests on SSO and SAML, MFA, and role and attribute based access control. Privileged actions are logged with acting admin, target user, role, tenant and timestamp, and a separately documented SCIM API handles user provisioning.

Authorization is enforced at a checkpoint that cannot be bypassed across 2,000+ production endpoints, with deny by default allowlisting and object level checks blocking cross tenant access. Multitenant isolation is taken from the authenticated session, enforced at the data layer and fails closed, with customer managed keys supported, and least privilege applies, with privileged support access time bound, approval based and logged.

A shared-responsibility table names what the customer configures: access, roles and permissions within their tenant, user provisioning and SSO. The compliance set covers FedRAMP Moderate authorization against 325+ NIST 800-53 controls, ISO 42001, ISO 27001, ISO 27017 and ISO 27701 certification, SOC 2 Type II and SOC 3 attestation, and Data Privacy Framework certification. Reports can be requested through a trust portal at trust.eightfold.ai with no NDA to begin. Published remediation timelines and continuous automated authorization testing sit underneath.

Sourceeightfold.ai/trust/security and apidocs.eightfold.ai/docs/scim-api-documentationread 2026-09-08

Eightfold keeps a real audit record with named fields: privileged actions are logged with acting admin, target user, role, tenant and timestamp, continuous monitoring and anomaly detection run across the environment, and bulk-access guardrails apply rate limiting and monitoring to stop authorized users extracting data at scale. The API adds async transaction ids that let a caller track an operation after the fact. What is logged is administrative action, not agent behavior.

Nothing documents a per-agent run trace, a decision record, or any field distinguishing a change made by Candidate Agent or AI Interviewer from one made by a recruiter, so a buyer cannot reconstruct why an agent did what it did. For a platform whose agents interview real candidates, that gap matters. The published bias audits evaluate model output; they are not a record of what happened.

Sourceeightfold.ai/trust/security and apidocs.eightfold.airead 2026-09-08

Scope is stated, and it is the vendor's own selling point: Candidate Agent carries a single memory of each candidate across role discovery, questions and application, with no re-introductions, which the vendor contrasts with point tools that leave a dozen different memories of the same person. Lifetime is not stated. The trust page says the customer controls retention and deletion and that data is deleted on request and on defined schedules under the DPA, but that governs customer data generally rather than an agent memory store with its own retention setting.

Sourceeightfold.ai/products/candidate-agent and eightfold.ai/trust/securityread 2026-09-08

The vendor's trust page says data is "stored in the region you choose", with controls against unintended cross-region processing so data stays within contractual and regulatory boundaries. A public subprocessor list backs it, naming each provider, its purpose and its region.

Customer managed keys are supported under multitenant isolation enforced at the data layer and failing closed, and FedRAMP Moderate authorization, assessed against more than 325 NIST 800-53 controls by an accredited third party with continuous monitoring, gives a named government grade environment alongside the commercial one. The shared-responsibility table on the same page puts data handling and retention settings on the customer's side of the line. The region is an option offered to the buyer, not just a fact about where the vendor runs.

Sourceeightfold.ai/trust/security and eightfold.ai/subprocessorsread 2026-09-08

A named agent portfolio ships out of the box, and each agent has its own product page: AI Interviewer, which conducts structured functional and coding interviews autonomously at scale; AI Interview Companion, which supports human-led rounds in real time with guidance, consistent evaluation and automatic documentation; Candidate Agent, which runs the applicant journey from role discovery through application; and Career Coach on the talent-management side.

Each does a whole job for a different user at a different point in the lifecycle: remove the Interview Companion and AI Interviewer still conducts interviews, and remove Candidate Agent and the interview agents still work. Alongside them sit a dynamic role library and an adaptable skills vocabulary as ready-to-use assets, and TalentForge as the route to build further agents on the same intelligence layer.

Sourceeightfold.ai/products/talent-agents, /ai-interviewer, /ai-interview-companion and /candidate-agentread 2026-09-08

Agents engage at the career-site front door, through conversational candidate journeys, and across interview touchpoints, described as contextual, multilingual and available around the clock, with a planned-events entity and a calendar endpoint on the API.

No webhook, event subscription or push mechanism appears anywhere in the developer documentation, and the vendor's own recommended integration pattern points the other way: the best-practices section advises sending updates for all open positions in a daily batch so none are missed. Async operations return a transaction id the caller polls for status, again a pull pattern. So work reaches the agents when a candidate arrives or a recruiter invokes them, not when an external event fires. Channels are career site, conversational interface, interview surfaces, email and calendar, with no messaging, voice or collaboration-tool presence documented.

Sourceapidocs.eightfold.ai/docs/getting-started and the Candidate Agent releaseread 2026-09-08

No model picker, admin model setting, routing description or bring-your-own-model capability appears on the trust and security page (including its security-architecture and responsible-AI sections), the four agent product pages or the public API documentation. The subprocessor list names OpenAI, LLC as Eightfold's generative AI services provider, so one model provider is disclosed, and the choice of model stays with Eightfold.

The vendor presents the platform as an AI-native engine trained on its own corpus, with AI subsystems held inside the platform's permission boundaries and the AI supply chain (training data, embeddings, feature stores and inference paths) access-restricted and monitored: vendor-operated machinery rather than something the customer selects.

Sourceeightfold.ai/trust/security, eightfold.ai/subprocessors and apidocs.eightfold.airead 2026-10-05

A public, ungated developer portal at apidocs.eightfold.ai documents a REST API in two versions, v2 served from apiv2.eightfold.ai/api/v2. It covers OAuth 2.0 authentication with a step-by-step configuration guide, a separate SCIM API, an authorization guide, an interactive API reference, a changelog, and an llms.txt publishing an index of endpoints in OpenAPI.

Roughly thirty entity and action pairs are listed with URLs for both versions (Profile, Employee, Position, ATS Candidate, ATS Position, Course, Events, Insights, Talent Pool) with full write coverage including POST, PUT, PATCH and DELETE.

Async operations return a transaction id trackable through a status endpoint, and the docs publish rate limits and thresholds per endpoint, pagination rules, a filterQuery syntax with named parameters for candidates, positions, demands and org units, and a complete error-code table. The vendor states the purpose outright: core entities are exposed so users can build independent systems on top of the Eightfold ecosystem. TalentForge, a layer for building custom apps and agents inside the product, is separate from this outside access.

Sourceapidocs.eightfold.ai/docs/getting-startedread 2026-09-08

No customer-facing evaluation surface is documented. Two things come close. Eightfold publishes independent bias-audit results for both the Matching Model and AI Interviewer under NYC Local Law 144 as public pages; these evaluate the shipped model's output against a defined standard, but they are the vendor's own annual regulatory disclosure, not something a customer can run against their own configuration.

Continuous automated authorization testing against a defined permission table checks the permission setup, not what an agent did, and it is Eightfold's own development practice. Nothing published lets a customer test a prompt, an agent configuration or a workflow change before or after deployment, and no scored test cases, simulation or optimization loop over agent behavior is documented.

Sourceeightfold.ai/trust/security, /trust/bias-audit-results and /trust/ai-interviewer-bias-audit-resultsread 2026-09-08

Nothing on the trust page, the agent product pages or the public API documentation describes a hosted or local browser, a desktop session, or remote or local computer control. Agents act through Eightfold's own interfaces and reach outside systems through the REST API and ATS, HRIS and LMS integrations, which are programmatic.

The closest case is AI Interviewer, which conducts structured candidate interviews at scale and evaluates answers on content rather than on video, voice or appearance, so it runs a real time multimodal conversation with a person. That is an interaction channel, not an agent operating software it does not control.

Sourceeightfold.ai/products/ai-interviewer and eightfold.ai/trust/securityread 2026-09-08