Source. Screen. Schedule. Per-client.

1. The pain — expensive tooling, expensive compliance

Recruiting agencies in 2026 carry the most expensive tooling per seat in B2B services. Bullhorn licenses run $99-315/user/mo plus $1k-50k implementation, with 20% renewal increases reported across agencies. Greenhouse and Lever entry tiers start around $6k-12k/year and climb to $63-144k+ for enterprise. LinkedIn Recruiter Corporate is now $10,800-$15,000 per seat per year — five seats puts you at $54-75k/yr for LinkedIn alone, with InMail and Talent Insights billed extra. Add Gem for sequencing, an ATS, a scheduling tool, and a compliance vendor and you're at $1,500-3,000 per recruiter per month before payroll.

The compliance load is heavier than the license cost. NYC Local Law 144 requires an annual third-party bias audit for any Automated Employment Decision Tool, with 10 business days' candidate notice and an alternative-assessment option. After the New York State Comptroller's December 2025 audit of DCWP enforcement, the agency committed to proactive (not complaint-driven) enforcement starting in 2026. Penalties run $15,000-$45,000 per month per non-compliant AEDT, before per-use multipliers.

And Europe is moving in parallel. EU AI Act high-risk obligations apply from 2 August 2026 to AI systems used for recruitment, candidate filtering, and evaluation, with deployer fines up to €15M or 3% of global turnover. Meanwhile US precedent keeps stacking: iTutorGroup paid $365k in the EEOC's first AI discrimination settlement; Mobley v. Workday was certified as a collective action in May 2025 (NDCA). Agencies running undocumented AI screening are not in a defensible position.

3. The os.moda stack

1. 1 · haku (sourcing) sources from LinkedIn, GitHub, portfolios, conference attendee lists. Per-role search profile, not generic Boolean. Outputs a ranked shortlist with a one-line reason each — and a SQL trace showing exactly why each candidate ranked.
2. 2 · soot (enrichment/ETL) enriches and dedupes candidates against your existing ATS, prior placements, do-not-contact lists, per-client off-limits agreements. No more pitching the head of eng to her own VP.
3. 3 · tofu (voice + text screening) runs voice and text screening on Telegram, WhatsApp, or web — in the agency's tone of voice, with the hiring manager's rubric. Transcripts, scoring, structured fields write back to the candidate record.
4. 4 · naga (rubric + bias audit) holds the per-client rubric, bias-audit configuration, KEYD vault credentials. Runs the LL144-aligned audit as a recurring job and emits a publishable report — not a separate compliance project.
5. 5 · lantern (hiring-manager brief) writes the weekly hiring-manager brief: pipeline status, top three candidates, blockers, next-week plan. Per-client voice. Citations link back to candidate records.
6. 6 · Per-hiring-manager dashboard pipeline by stage, time-to-fill, screening notes, scheduled interviews, offer status. White-labeled. Replaces the per-client Notion / Slack / email sprawl.
7. 7 · SHA-256 audit ledger logs every sourcing decision, screening question, rubric application. EU AI Act Article 26 logs (6-month retention minimum) are a query, not a project.

5. The 3–5 year future

• 2027 · per-client agentic recruiter. Each hiring manager has a named agent with their tone, rubric, and pet peeves. The agent sources, screens, and books — the human recruiter does relationship and offer close.
• 2028 · compliance-as-a-runtime. The bias audit is the agent's output, not a quarterly project. Every screen is auditable in real time, audit report is a publishable URL versioned alongside the model.
• 2029-2030 · agentic onboarding. The same agent that ran the screen runs the first-week onboarding — paperwork, intros, learning plan — then hands off to the people manager with a structured 30-60-90 brief.