HR AI crossed from reviewing what candidates submit into running the live conversation itself.

The week in hr-recruiting

The directional move this week is AI stepping out of the review seat and into the conversation. Across 26 products updated, 11 sparks landed at 10 vendors, and the ones that matter share a shape: the model no longer grades an artifact a candidate uploaded, it captures the interaction as it happens. Eightfold AI collapsed a role's separate interview rounds into one agent-run conversation with 360 Interview, and Spark Hire put an AI notetaker inside the live interview that drafts evaluation notes from what was actually said. Both move an agent up the funnel into a stage that used to require a person in the room, while leaving the final decision human.

The second pattern is the suite land-grab and the enterprise controls that gate it. Against 58 improvements sector-wide, the substantive non-spark work is unglamorous: permissions, auditability, and compliance plumbing. Workable pushed past the ATS into native employee surveys, a category customers used to buy standalone, and spent its other releases giving its hiring agent editable criteria and entity-scoped admin access rather than more autonomy. Underneath the launches, Deputy and Tanda ran quiet regulatory-fit and access-control trains that never make a headline but keep the products sellable to larger, multi-site buyers.

Leaders

Eightfold AI shipped the week's most consequential spark: 360 Interview folds the scheduled panels that make enterprise hiring take weeks into a single adaptive conversation. It is the logical continuation of Candidate Agent taking the pre-application stage in July — the agents are working through the funnel rather than sitting beside it.

Spark Hire moved its AI from reviewing submitted video to capturing the interview directly with AI Notetaker, then followed within days with role-aligned ratings that score answers against the job description and scorecard. The notetaker produces the transcript; the criteria make it reasoning. It is the clearest example this week of one release deliberately making the next possible.

Workable turned on native engagement, pulse and onboarding surveys, landing results in the same system as its performance reviews. Paired with a 30-tool MCP server and edit-and-replay controls on its generally-available hiring agent, this is a company trying to be the system of record for the whole employee lifecycle, with hiring as the entry point.

Workyard extended its AI Time Assistant to reallocate hours across jobs and cost codes from a plain-language request, rendering each change as a struck-through diff awaiting approval. With clock times, breaks and now project time in one surface, the assistant now covers the full pre-payroll cleanup pass a bookkeeper used to do by hand rather than one slice of it.

Wagepoint put an AI review layer at the payroll approval gate: AI Payroll Summary flags unusual changes in a run before the approve click, turning a line-by-line read into an exception review. It is the company's first model-backed feature inside the pay-run flow, aimed squarely at the accountant whose name is on the payroll.

Wildcards

Envoy converted the roster and directory it already held into a physical-security product. Threat Intelligence launched continuous monitoring of verified global and local risks on a Threat Dashboard, extended by HRIS-synced home addresses and per-executive VIP protection. A workplace-management vendor became a security one without collecting any new data — and the follow-up releases, scoped blocklists and map audit logs, are already hardening it for that register.

Ever Gauzy is the counter-example: eight releases in about 26 hours, almost none of which change the product. After a month of genuine AI, OCR and Documents work, this window was spent fixing its own build pipeline — Azure Artifact Signing for Windows desktop builds shipped gated, then chased through three patches after the signing config turned out silently inert. A maintenance interval that is self-correcting, each version bump repairing the last.

Themes that compounded

• AI moved into the live conversation, not just the artifact review, at Eightfold AI and Spark Hire.
• The highest-trust AI is landing at the approval gate — Wagepoint before payroll sign-off, Workyard before timecards close — as a proposal a human accepts, never an automated write.
• Skills and hiring data are being pushed out through agent-access surfaces: Workable's expanded MCP server and CodeSignal's MCP that lets a manager query candidate skills from Claude or ChatGPT without opening the dashboard.
• Enterprise controls, not new capability, dominated the improvement column — entity-scoped permissions at Workable, pay-data restrictions at Deputy, scoped blocklists at Envoy.
• Verification keeps automating toward the exceptions: Checkr added fraud detection on the employment documents that survive automated review, hunting the failure mode that shifts from slow to falsified.

Watch this week

The agentic push has a clear next target. Eightfold AI's own framing points at scheduling, panel debriefs and offer-stage coordination as the remaining steps between an agent-run loop and a human decision, and Spark Hire's role-aligned scoring logic is positioned to reach back into screening. On the review-gate side, Wagepoint's anomaly summary is a candidate for an exportable, client-shareable record for the firms that sell on it. And Envoy's security pivot most likely gains alerting rules and roles next, following the same scope-and-audit pattern it just applied to blocklists.