A deep agent for hiring that fixes what AI broke: it separates real candidates from the AI-written flood, verifies each one is a real, consistent human, and gives every score a reason, while a human signs every reject and every offer, by design.
What is the afka Recruiting agent? A single AI Brain for an SMB hiring team, run as three coordinated agents (sourcing, screening, and onboarding & engagement) on one shared talent graph. It turns a one-line brief into a ranked, reasoned shortlist: de-noising the AI-written application flood, flagging fakes and identity drift as signals for human review (never an auto-reject), interviewing on the candidate's terms (text, voice or video, no timer), carrying new hires through onboarding and early engagement, and keeping one governed candidate record with the why. It is built for the 2026 reality that AI broke the pipeline, and for compliance: reject, screen-out and offer are never finalized by the AI at any autonomy level.
Everyone sells "source and screen faster." But speed is commoditized. The real 2026 pain is a flood of AI-written applications, fake candidates, and a funnel real people bail on. afka fixes that, and keeps a human on every decision that matters.
Everything below runs in your own ATS and inbox, as sourcing, screening, and onboarding & engagement on one talent graph. Candidate-impacting actions are gated by design, and every score is written with its reasoning to an append-only audit log.
Flags AI-written and templated applications, spots duplicates and the same human applying twice, and surfaces the genuinely placeable few, with the evidence and a reason, never a bare score. In 2026 a real candidate is the scarce resource, and it finds them.
Checks identity continuity across channels and integrity signals on video evidence to catch deepfakes and real-time assistance. Every flag routes to you for review, and it never auto-rejects anyone on a suspicion.
Async evidence interviews where the candidate picks text, voice or video, with no timer, testing real reasoning over gameable scripts. It is transparent that it is an assistant and a human decides, which reverses the roughly 40% who bail on AI-only rounds.
Ranks with written reasoning, keeps every candidate updated so no one is ghosted, carries new hires through onboarding and early engagement, and learns from who actually gets hired and stays on the shared talent graph, recalibrating on retention, not vanity speed.
The same four questions get answered on every action and written to the audit log: who started it, which agent ran it, what data it touched, and why it was allowed.
Scores all 41 against your profile, flags 26 as AI-templated with the giveaway evidence, and catches 2 duplicate identities. Nothing rejected.
The 4 worth your minute come with written reasoning and quotes from the actual resume, not a black-box number you have to trust.
One video interview trips a liveness check. It routes to you with the evidence: a signal, never a verdict, and never an auto-reject.
Human reviewYou advance, pass, or interview. Every step, and the reason, lands in one candidate record and the append-only audit log.
It triages the overnight flood while you sleep, then sends one message at 07:30, with the reasoning for every candidate it surfaces.
Hiring is regulated as high-risk under the EU AI Act, NYC Local Law 144 and EEOC scrutiny. afka is built so the AI never makes the call that affects a person's livelihood. Every afka agent is built to the same operating standard.
Reject, screen-out and offer are never finalized by the agent, at any autonomy level. It prepares the decision with the evidence, and a person signs it. This gate does not have an "off."
Fake, deepfake and identity-drift signals route to human review with the evidence, never an automatic rejection. A suspicion never ends someone's candidacy on its own.
An adverse-impact self-check runs on every shortlist, and a breach pauses the role's scoring for your review. It never scores on protected attributes, and stores only candidate-provided evidence.
Because your afka agents share one memory, the signal to open a role reaches recruiting from the teams already feeling the strain.
A sustained support surge becomes a staffing signal ("open a CX role?") with the workload evidence attached, before burnout hits.
A jump in won deals flags the need to hire for delivery capacity, before the work outruns the people who do it.
When a role like a past one opens, the strong-but-declined candidates from last time come back up automatically, no re-sourcing from scratch.
Every touch and its reasoning live in one candidate record, so a handoff to the hiring manager is context, not archaeology.
Being specific about the edges is part of the standard. Here is where the agent deliberately stops.
Reject, screen-out and offer are never finalized by the AI at any autonomy level. A human signs every decision that affects a candidate. There is no setting that changes this.
Fraud, deepfake and identity signals are routed to you for review with the evidence, never used to auto-reject. A suspicion is a prompt to look, not a verdict.
It never scores on protected attributes or their proxies, runs an adverse-impact check on every shortlist, and stores only the evidence a candidate provided, not a background-data profile.
Every active candidate gets a transparent status update on the clock. No one is left waiting in silence, and the agent always says it is an assistant when asked.
Connect your ATS and inbox, brief it like a colleague, and wake up to the few real candidates worth your minute: de-noised, reasoned, and with every decision that matters signed by a human.