Use cases by industry2 use cases
AI for recruitment and staffing agencies, built for the EU AI Act
An IT contracting agency with 40,000 candidates in Bullhorn and 15 recruiters pays job boards to source people it already has. Contractors whose assignments end next month, and the near-miss candidates from last quarter's shortlists, sit in the database while recruiters start from scratch. Staffing candidate redeployment watches end dates and new job orders, matches existing candidates on skills, availability, location and rate, and drafts the outreach; the recruiter decides who is contacted.
Screening is the crowded part. Bullhorn Amplify already sources, screens and submits for Bullhorn users, and Bullhorn reports that its customers see 51% more submissions, its own figure. Bullhorn's 2026 GRID survey says 30% of staffing firms already use agentic AI. If you run Bullhorn and your process is standard, start there.
Europe adds a legal frame. AI used to screen or rank candidates is high-risk under the EU AI Act; the Digital Omnibus moved those obligations to 2 December 2027, and GDPR Article 22 already restricts decisions based solely on automated processing. Explainable candidate screening is built for that: a rubric the recruiter defines, quoted evidence for every criterion, blind review, an audit log, and a person making every decision.
What I would build for recruitment and staffing
Candidate screening that quotes the CV behind every score, built for the EU AI Act
Scores each application against a rubric the recruiter wrote, quotes the CV line behind every score, catches hidden instructions, and never rejects anyone on its own.
ATS (Greenhouse, Personio or Bullhorn) / Gmail / Slack / The language model / Audit store (Postgres)
Redeploying contractors and near-miss candidates before paying to source new ones
Finds contractors rolling off and past near-misses who fit a new job order, drafts personal outreach in the recruiter's mailbox, and updates Bullhorn with every reply.
Bullhorn, Vincere or JobAdder / Outlook / WhatsApp Business Platform / SMS gateway / Postgres with pgvector
The staffing work recruiters would hand over first
Orders from vendor portals
Job orders arrive through SAP Fieldglass, Beeline and client emails. Pulling them in, checking rates and compliance requirements, and packaging submissions in each portal's format is admin that eats recruiter time.
Credentials that expire
Healthcare staffing lives on licenses and certificates with expiry dates. Tracking them per candidate and warning before an assignment starts keeps an expired certificate from reaching a client.
Assignment limits in Germany
German temporary agency work law limits how long a worker can be placed with one client. Tracking durations per worker and client, and warning well before the limit, is plain date logic that belongs in the system, not in a recruiter's memory.
Timesheets against purchase orders
Matching timesheets, rates and client purchase orders before invoicing catches disputes while they are still cheap to fix.
Where the AI Act and GDPR draw the line
The recruiter decides
The system ranks, explains and drafts. Who is contacted, shortlisted, submitted or rejected is a recruiter's decision, recorded in the ATS.
Evidence instead of a match score
Every criterion shows the sentence in the CV that supports it. A black-box percentage cannot be explained to a candidate or a regulator.
Proxies for protected traits
Criteria such as graduation year, career gaps or postcode can stand in for age, parental status or origin. Rubrics are reviewed for that, and outcomes are monitored across groups.
Re-contact needs a lawful basis
Old candidate records need a retention rule and a lawful basis before anyone messages them again, and contact frequency is capped so the database stays an asset rather than a spam list.
Hidden instructions in CVs
White text telling an AI screener to rank the candidate first is a known trick. CV text is treated as data, never as instructions, using the same prompt injection defenses as any agent reading outside content.
Software these builds usually connect to
The systems do not get replaced. The build sits across them, reads from them through their APIs, and writes results back.
Bullhorn / Vincere / JobAdder / Mercury xRM / Firefish / Greenhouse / Personio / SAP Fieldglass / Beeline / Microsoft Outlook / WhatsApp Business
Frequently asked questions
Is AI candidate screening legal under the EU AI Act?
Yes, but it is high-risk. AI used to screen, rank or evaluate candidates falls under Annex III, and after the Digital Omnibus those obligations apply from 2 December 2027: risk management, documentation, human oversight, logging and transparency. GDPR Article 22 applies now and restricts decisions based solely on automated processing. A system where the recruiter decides, with evidence shown for every criterion, fits both.
What are AI agents for recruitment agencies actually good for?
Mostly the work around the placement rather than the judgment: finding candidates already in your database who fit a new order, drafting personal outreach, pulling orders from vendor portals, tracking credentials and assignment limits, and reconciling timesheets. Screening can be assisted, with explanations and a recruiter deciding, but it carries the most legal weight and should come after the admin wins.
Is Bullhorn Amplify enough, or do we need a custom build?
Start with Bullhorn Amplify and Bullhorn Automation, which cover sourcing, screening, engagement and submissions on Bullhorn data. Custom work fits when matching needs data Bullhorn does not hold, such as timesheets, client feedback or a second ATS, when your rubric and explainability requirements are your own, or when candidates live in several systems after an acquisition.
Also browse: HR and recruiting · Outlook and Microsoft 365 · WhatsApp · all use cases