The recruiting copilot

The problem

A recruiting firm's tooling was the industry standard: a form-based applicant tracking system, fifteen screens deep. Recruiters spent their days clicking through dropdowns instead of talking to candidates. Worse, the system never told them what mattered — candidates quietly went stale mid-pipeline, follow-ups slipped, and every placement depended on a recruiter's memory.

What we built

A conversational copilot that opens every session with a briefing, not a blank chat box: "3 candidates are stuck in First Interview 6+ days. A new candidate matches your open Java role. One pipeline is still empty." Those priorities come from a deterministic engine — eight ranked task types computed from real pipeline state, each with a why and a one-click action.

From there, the recruiter works in plain English. Drop a resume PDF into chat and a structured, schema-validated candidate profile appears, instantly matched against open roles. Say "move her to shortlist" and a strict nine-stage state machine executes the transition — pipeline totals always reconcile, candidates never disappear. The agent drafts stage-appropriate emails from thirteen typed templates into an editable panel; the recruiter can say "make it more urgent" and watch the draft rewrite — but nothing sends without an explicit Approve click, and every action lands in an append-only audit trail.

The harness pattern

The AI is confined to bounded jobs — parsing, drafting, planning — while everything that must be correct is deterministic code. Typed tool contracts are the only way the agent can touch data, and every write requires human approval. The whole system is model-agnostic: the recruiter can switch the same workflow between Claude, GPT, Gemini, and Grok from a menu, live.

Notably, our first version used a multi-agent orchestrator. We measured it, found a single governed agent simpler and more reliable, and rebuilt. We charge for that kind of honesty.

Verifiable facts

• Every consequential action human-approved and logged — zero unaudited side effects, by construction.
• Thirteen typed email templates, all approval-gated.
• Runs on four AI providers, switchable live from the interface.
• Ships with a deterministic demo test suite — the demo works every time, not just in the recording.