Persol Holdings Co. is piloting a shift-matching large language model for Kansai warehouse night crews, ranking substitute workers against forklift certifications, hazardous-material tags, and prior site ratings while requiring human dispatchers to approve every assignment. The test covers three logistics campuses in Osaka and Hyogo ahead of peak autumn e-commerce volume and follows revised labor-agency guidance treating AI staffing suggestions as advisory, not autonomous hiring.
What the artifact does
Dispatchers type natural-language requests—“need two reach-truck operators tomorrow 22:00 with cold-chain experience”—into a console connected to Persol’s temp database. The model returns ranked lists with explain strings citing certification IDs and commute bands; it does not SMS workers directly. Low-confidence ranks display amber banners; empty pools return explicit “no qualified match” instead of hallucinating credentials.
Persol logs prompts and approvals for ninety days on domestic cloud tenants segregated from marketing data. Workers opt into skill tags during onboarding; the pilot forbids using health or injury history as ranking features—a line compliance copied from internal bias checklists.
Claim versus evaluation
Persol claims the model cuts average fill time for night shifts by double-digit minutes in simulation replay of August call logs; it has not published controlled trials or third-party audits. Warehouse clients asked whether explain strings hold up when dispatchers override suggestions; Persol tracks override reasons manually during September to tune prompts.
Ministry of Health, Labour and Welfare labor-standards briefings remind agencies that temp dispatch remains licensed activity—AI cannot sign contracts. Persol’s UI watermark states dispatchers retain legal responsibility, a detail labor lawyers said must appear on printouts auditors see.
Who has power
Client warehouse managers can veto ranked lists but cannot access raw model weights. Temp workers see only assigned shifts, not their rank position—reducing gamification but limiting transparency advocates wanted. Personal Information Protection Commission rules on automated decision-making still map to human final approval; Persol declined to seek formal guidance until pilot metrics arrive.
Technical limits
The model runs on a hosted Japanese-capable foundation with retrieval over certification tables refreshed nightly. Latency targets stay under five seconds; beyond that dispatchers revert to spreadsheet habits the pilot aims to kill. Cold-chain tags mis-synced in one August test produced a false match caught by a dispatcher; Persol added mandatory registry refresh before night peaks.
Integration stops at Persol’s scheduling portal—no direct hooks into client WMS systems in phase one. APIs may follow if clients demand scan-to-staff workflows; engineers said that expands liability if barcodes encode productivity scores.
What careful readers still will not know
Persol has not named the foundation-model vendor or disclosed false-credential near-miss rates beyond internal summaries. Until those numbers publicize, treat speed claims as operational hypotheses. For night crews, the near-term change is faster phone trees when someone calls in sick—not robot bosses on the floor.
Industry context
Rival staffing firms experiment with rule-based matchers; none market LLM explainability in Japan-wide pilots yet. Japan Warehousing Association safety bulletins stress certification checks regardless of software; Persol invites association observers to a closed October walkthrough.
Pilot milestones
September metrics cover override rates and time-to-fill; October decisions determine whether Kansai expands to chilled-food sites with stricter tags. Clients must keep manual backup lists during the pilot—Persol’s contract assigns liability if dispatchers blindly accept top ranks without reading explain strings.
Worker communication
Temp staff receive FAQ sheets clarifying the model does not score personality or social posts—only declared skills and site feedback voluntarily entered after shifts. Union outreach groups asked for aggregate bias reports; Persol said it may share internal dashboards with MHLW upon request but not publicly during competitive season.
Client onboarding
Warehouse safety officers must upload certification schemas before the model activates; Persol blocks generic “warehouse worker” tags that caused ambiguous ranks in early tests. Clients who refuse structured tags stay on manual dispatch—an escape hatch labor lawyers insisted remain visible in contracts.
Cost accounting
Persol bills the pilot as a per-dispatcher seat uplift, not a per-hire commission, so finance teams can compare against overtime spend when night fills fail. If fill-time gains do not materialize by November, two Hyogo clients may revert to legacy software without penalty clauses—a flexibility Persol offered to win the pilot slate.
