Agentic Fulfillment
Proposed reference design

Seven agents run the hiring workflow. People make the hiring decisions.

Each agent covers one stage of the process flow, from negotiating demand to republishing at-risk requisitions. Agents handle forecasting, outreach, scheduling, evidence gathering and recovery. Hiring, rejection and pay decisions stay with a named person, and every action is logged.

7agents, one per stage of the process flow
18process steps covered, demand to republish
10human decision points, all logged with reasons
4regimes mapped: EU AI Act, DPDP, US, ISO/NIST
Acts on its ownActs within limitsRecommends, person decidesShared control plane
01

Autonomy follows consequence, not convenience

The same agent can act freely on reversible, administrative work and still stop for a person when the action shapes someone's career or pay. Read the chart bottom to top: the higher the consequence, the more orange it gets.

Leadership
Agent acts on its own, audited afterPerson decides, agent prepares the evidence
The top row is entirely human. No agent in this design rejects a candidate, takes an adverse trust action, or sets pay. That is the line regulators draw for recruitment AI, and it is also what performs: AI-led information gathering with human decisions has been shown to raise offer and retention rates in field trials.
02

Follow one requisition through the system

An illustrative scenario: 40 Java developer openings in Pune. Press play, or step through. Watch which actions the agents take alone, where a person steps in, and what lands in the audit ledger.

LeadershipSelf-serve

Audit ledger

    Figures are illustrative to show behaviour, not forecasts or benchmarks.

    03

    One control plane under every agent, inside your boundary

    Agents never call your systems directly. Every tool call passes a policy check, every consequential action waits in a review console, and every event is written to a ledger. All of it runs inside your environment on your data.

    IT and security
    PeopleWho works with it
    Hiring managerNegotiates demand, approves requisition
    RecruiterConfirms shortlists, owns adverse calls
    PanelInterviews, decides hire
    ApproverSigns off offer terms and spend
    CandidateTold when AI is involved, can reach a person
    ExperienceWhere people meet it
    Demand workspaceLive what-if on fill odds
    Review consoleApprove, edit, reject, escalate, stop
    Candidate channelsEmail, SMS, WhatsApp, voice, video
    Signal canvasFill risk across requisitions
    OrchestratorMoves the case file
    Fulfillment orchestratorHolds one case file per requisition and candidate, hands work between agents in a fixed format, enforces stage rules
    AgentsSpecialists, narrow scope
    Control planeBuilt once, enforced outside the model
    Tool gatewayDeny by default, short-lived scoped access
    Guardrail serviceSeparate model screens inputs and outputs
    Human reviewQueues, reasons, overrides
    Audit ledgerWho, what, when, versions, reviewer
    Consent and noticeDPDP notices, withdrawal, deletion
    Evals and bias monitorQuality and adverse impact by role
    Stop controlsPause one agent or all
    Systems of recordStay the source of truth
    ATS
    HRIS
    Comp bands
    Calendars
    Talent pool
    Identity checks

    Your data stays put

    Agents read from your systems of record and write back through them. By design, candidate data does not train shared models. Memory expires and is deleted when consent is withdrawn or the purpose ends.

    Every agent has an identity

    Each agent has its own permissions and its own trail, so any action can be traced to the agent, the policy version and the person who approved it.

    Built to be handed over

    Co-created with your team, documented as it is built, and designed for your ownership: playbooks, rubrics and thresholds live in your configuration, not in vendor code.
    04

    Inside each agent: the same eight-part build

    Every agent is assembled from the same parts: a goal, memory with an expiry date, a plan, guardrails on the way in and out, scoped tools, an autonomy gate, and a record. Pick an agent.

    Product and engineeringSelf-serve
    05

    Compliance designed in, stage by stage

    Recruitment AI is treated as high-risk by default. Each cell shows how strongly a control applies to an agent. Hover or tap a cell for the reason.

    Legal and privacy
    Select a cell to see why the control applies.
    Core control for this agentAppliesNot material
    In forceEU ban on emotion inference at work; NYC LL144 bias audits
    2 Aug 2026EU duty to tell people they are talking to AI
    13 Nov 2026DPDP consent manager regime starts
    1 Jan 2027Colorado disclosure and human-review path
    13 May 2027DPDP notice, consent and rights fully enforceable
    2 Dec 2027EU high-risk duties for recruitment AI
    06

    Red lines, held regardless of what competitors ship

    Some capabilities are left out on purpose. They carry legal exposure today and erode candidate trust, and the same outcome can be reached with evidence and a human decision.

    LeadershipLegal and privacy
    No emotion inferenceNothing is scored from tone of voice or facial expression.
    No auto-rejectA person confirms every exit, with an appeal route.
    No AI negotiatorThe agent never holds or bargains the pay ceiling.
    No hidden scoresNo cross-employer score gates a candidate; evidence, not a number.

    Sources

    Arrow keys move between chapters. Esc to exit.