Customer Service

Pilot wave 1 (M5). Impact: high. See the hub.

Current state (assumptions to validate)

Customer service sits between shipping clients (owners, managers, agents — English), Greek retail/e-commerce PPE customers, and internal operations. The daily reality: a high-volume shared inbox mixing order-status requests (“where is our delivery, vessel sails at 18:00?”), requisition clarifications, complaints (short-shipped or wrong items, temperature-sensitive provisions), document requests (invoices, delivery notes, certificates), and e-commerce queries. Urgency is dictated by sailing schedules — a status question about a vessel departing tonight cannot wait in queue order.

Assumptions — to validate in discovery: team ~5–8, working in Outlook + SoftOne lookups + phone; no ticketing system; triage is whoever-opens-it; answers require checking SoftOne, warehouse, and logistics by phone or Teams; bilingual load with English formal correspondence taking longer to draft. Pain points: no urgency-based prioritization, repetitive lookups, inconsistent reply quality, knowledge trapped in individual mail histories.

Target AI-first operating model

By M12 the shared inbox is triaged by an agent: every incoming email classified (order status / complaint / document request / new RFQ → routed to sales / e-commerce), urgency-scored against sailing deadlines, and pre-answered with a grounded draft reply in the customer’s language pulling order status from SoftOne. Staff review, edit, and send — nothing goes out unapproved. Complaints get a structured intake summary and a case file. The team’s job shifts from typing lookups to managing exceptions and customer relationships.

AI use cases

Use case	Pain addressed	Data needed	Complexity	Impact	Pilot
Bilingual email triage & draft replies (EN/GR)	Volume, inconsistent quality, slow English drafting	Shared mailbox, reply templates, SoftOne order status	M	H	Y
Order-status answer generation from SoftOne	Repetitive lookups	SoftOne orders/deliveries, logistics schedule	M	H	Y
Urgency scoring against vessel sailing times	Deadline-critical mails buried in queue	Email content, port-call/ETA data from orders	M	H	Y
Complaint intake summarization & case drafting	Unstructured complaint handling	Complaint emails, delivery records, photos/scans	M	M	N
Document retrieval assistant (invoice/delivery-note/certificate copies)	Time spent fetching documents	SoftOne documents, SharePoint archives	L	M	N
FAQ/knowledge answers for e-commerce PPE queries	Repetitive product questions	Product catalog, datasheets, shop content	L	M	N

Process transformation opportunities

• Inbox to queue: replace first-come email handling with a triaged, urgency-ranked queue (agent-classified), including auto-routing of RFQs to sales.
• Complaint handling: structured intake → case record → root-cause tag feeding compliance/quality, instead of email threads.
• Self-serve status: standard proactive status notifications (order confirmed / picked / delivered to berth) so status emails stop arriving at all.

Required data sources

• Shared customer-service mailbox (Outlook), historical threads for template mining.
• SoftOne: orders, delivery status, invoices, customer records.
• Logistics delivery schedule; vessel ETA/ETD as recorded on orders.
• SharePoint: reply templates, product certificates; e-commerce shop content.

Potential AI agents

• Email Triage Agent — classifies, urgency-scores, and routes inbox items; human owns the queue.
• Reply Drafting Agent — grounded bilingual draft replies with SoftOne status; human approves every send.
• Complaint Intake Agent — summarizes complaints into structured cases; human validates and dispatches.

Automation opportunities

• n8n: mailbox polling → triage → Teams alert for sailing-critical items.
• Proactive status notification flow on SoftOne delivery-status changes.
• Daily open-cases digest to the team lead.

Required integrations

• M365 Outlook shared mailbox (read + draft creation), Teams; SoftOne read (orders, deliveries, invoices) — see integrations.

KPIs

KPI	Baseline	M12 target
Median first-response time	TBD (M2)	−60%
Sailing-critical emails answered within 1h	TBD	≥95%
Draft replies sent with minor/no edits	TBD	≥70%
Status-request email volume	TBD	−40% (proactive notifications)
Assistant weekly active users	0	100% of team

Risks

• Wrong status information sent to a customer → drafts always cite the SoftOne record; human approval mandatory.
• Misclassified urgent email delayed → conservative urgency defaults; sailing-related keywords always escalate.
• Tone/formality errors in English correspondence → template grounding; pilot review of samples before scale-up.

Training needs

• All CS staff: assistant basics, triage-queue workflow, editing drafts responsibly (M4–M5).
• Team lead: queue analytics, quality sampling of AI drafts (M6).

Deliverables

• Triage + draft-reply pilot (M5), production queue (M9).
• Bilingual reply-template library; complaint case structure.
• Proactive-notification automation; CS KPI dashboard.

12-month execution milestones

Month	Milestone
M1–M2	Discovery: mailbox volume analysis, category taxonomy, baselines
M4	Training; template library assembled
M5	Wave-1 pilot: triage + draft replies on the shared mailbox
M6	Pilot evaluation; SoftOne status grounding added
M7	Urgency scoring live; complaint intake pilot
M9	Email Triage + Reply Drafting Agents in production; proactive notifications
M12	KPI review; handover to team lead