AI Use-Case Pipeline

Living register. Funnel states: idea → scored → pilot → production → retired. All entries below are scored unless noted. Update the row in the same change as any decision. Entry/approval mechanics: _index.md.

Scoring method

Priority = Impact × Feasibility, adjusted by modifiers.

• Impact (L/M/H): hours saved × people affected, revenue protected (tenders won, quotes faster), or error/risk reduction.
• Feasibility: data readily available? systems integrable now (../05-technical/integrations-m365-softone.md)? output verifiable by a human in seconds?
• Modifiers: −complexity (H complexity needs a stronger case), −risk (customer-facing, compliance, or financial-action use cases need approval gates before they can pilot), +platform reuse (uses components we already built).

Pilot suitability criteria: a motivated champion exists; a measurable KPI can move within 6–8 weeks; failure is cheap and non-customer-visible; needed data already flows into the platform. Wave-1 pilots are picked for momentum, not maximum theoretical value.

Master register

Complexity/Impact: L/M/H. Phase = suggested program milestone (M#). Departments: SAL sales, TEN tenders, PRC procurement, CS customer service, WHS warehouse/logistics, FIN finance, HR, QUA quality/compliance, IT, MGT management, MKT marketing/e-shop.

ID	Use case	Dept	Process	Pain point	AI solution sketch	Data needed	Systems	Cx	Impact	Risk	Pilot fit	KPI	Phase
UC-01	Internal knowledge assistant	All	Find company knowledge	Answers live in heads, folders, old emails	RAG chat over vault + SharePoint with citations	Vault, SharePoint, SOPs	Open WebUI, Qdrant	L	H	L	High	Time-to-answer; weekly active users	M3
UC-02	ERP query assistant (NL over SoftOne)	SAL/MGT/WHS	Look up stock, prices, orders	ERP queries bottleneck on a few power users	NL-to-SQL over semantic layer, freshness-stamped answers	ERP extracts	SoftOne→PG, LiteLLM	M	H	M	High	Self-served queries/week; deflected requests	M4
UC-03	Sales quote / RFQ assistant	SAL	Quote assembly	Quoting a multi-line RFQ takes hours of lookup	Draft quote lines from RFQ text + catalog/price data, human finalizes	Catalog, price lists, quote history	Semantic layer, RAG	M	H	M	High	Quote turnaround time; quotes/rep/week	M4
UC-04	IMPA-code product matching	SAL/WHS	Vessel requisition intake	Manual matching of requisition lists to IMPA/catalog codes	Extract lines from PDFs/emails → fuzzy+semantic match → confidence-ranked list for review	IMPA catalog, product master, past requisitions	RAG, semantic layer, OCR	M	H	L	High	Match precision; minutes per requisition	M4
UC-05	Tender compliance-matrix extraction	TEN	Tender qualification	Reading 100+ page tenders for requirements is slow and missable	Extract requirement clauses → compliance matrix draft with clause citations	Tender PDFs, past matrices	OCR, RAG, LangGraph agent	H	H	M	High	Hours per tender; missed-requirement count	M5
UC-06	Procurement supplier-quote comparison	PRC	Supplier selection	Comparing multi-format supplier quotes by hand	Normalize quotes into a comparison table + flag anomalies	Supplier quotes, product master	OCR, n8n, semantic layer	M	M	L	Med	Comparison prep time	M5
UC-07	Customer email triage & draft replies (bilingual)	CS	Inbox handling	Shared inboxes overflow; repetitive replies	Classify/route + draft GR/EN replies for approval	Shared mailboxes, KB, order status	Graph API, n8n, RAG	M	H	M	High	First-response time; % drafts sent ≤ minor edits	M4
UC-08	Meeting summarization	All	Meetings	Decisions/actions evaporate	Whisper → summary → vault note draft	Recordings, calendar	Whisper, n8n, vault	L	M	L	High	% meetings with notes; action follow-up rate	M3
UC-09	SOP generation from interviews	QUA	Process documentation	SOPs outdated or nonexistent	Interview transcript → structured SOP draft in vault format	Transcripts, SOP template	Whisper, LLM, vault	L	M	L	High	SOPs published/month	M4
UC-10	Monthly management report pack	MGT/FIN	Reporting	Manual copy-paste reporting eats days	Semantic-layer figures → drafted narrative pack	ERP KPIs, prior packs	Semantic layer, n8n	M	M	M	Med	Days to produce pack	M6
UC-11	Inventory insight assistant	WHS/PRC	Stock management	Slow-movers, stockouts spotted late	NL queries + scheduled anomaly digests on stock data	Stock, orders, lead times	Semantic layer	M	M	L	Med	Stockout/slow-mover incidents flagged early	M6
UC-12	Management dashboard / KPI narrative	MGT	Performance review	Dashboards lack the “so what”	Auto-narrative on KPI movements with drill-down Q&A	KPI tables	Semantic layer, Grafana	M	M	M	Med	Exec usage; prep time for reviews	M6
UC-13	HR onboarding assistant	HR	Onboarding	New hires ask the same 100 questions	RAG assistant over HR policies + onboarding checklists per role	HR docs, policies	RAG, n8n	L	M	L	High	Ramp-up time; HR interruptions	M5
UC-14	Technical documentation assistant (PPE/product specs)	SAL/MKT	Product support	Spec lookup across supplier PDFs is slow	RAG over datasheets/certs; draft spec answers with citations	Datasheets, certificates, e-shop data	RAG, OCR	L	M	L	High	Spec-answer time; e-shop content reuse	M5
UC-15	Compliance documentation assistant (ISO 22000/21001)	QUA	Audit readiness	Evidence gathering before audits is frantic	RAG over QMS docs; gap-check drafts against clause lists	QMS docs, standards clause lists	RAG	M	M	M	Med	Audit-prep hours; findings count	M6
UC-16	Process audit assistant	QUA/MGT	Internal audit	Audit trails assembled manually	Draft audit checklists + evidence summaries from logs/docs	SOPs, records	RAG, semantic layer	M	L	M	Low	Audit cycle time	M7
UC-17	AI helpdesk assistant (internal IT)	IT	IT support	Small IT team drowning in how-to tickets	RAG over IT KB + guided fixes; escalation with context	IT runbooks, ticket history	RAG, n8n	L	M	L	High	Ticket deflection rate	M5
UC-18	Customs / bonded-warehouse document drafting	WHS/FIN	Customs paperwork	Repetitive transit/customs docs, error-sensitive	Draft documents from order + template with field validation, human files	Order data, templates, past filings	Semantic layer, n8n	H	M	H	Low	Doc prep time; rejection rate	M7
UC-19	Receivables chasing draft letters	FIN	Collections	Chasing overdue invoices is manual and inconsistent	Aging-driven bilingual reminder drafts, tone-tiered, approval before send	Receivables aging, customer history	Semantic layer, n8n, Graph	L	M	M	Med	DSO; % overdue contacted on schedule	M5
UC-20	Supplier document/cert expiry watch	PRC/QUA	Supplier compliance	Expired certs found at the worst moment	Extract expiry dates from certs, watchlist + alerts	Supplier certs	OCR, n8n	L	L	L	Med	Expired-cert incidents	M6
UC-21	Tender opportunity scanning (public sector)	TEN	Business development	Missed relevant tenders on public platforms	Scan feeds/portals, classify fit, weekly digest	Tender feeds, win history	n8n, LLM	M	M	M	Low (external data)	Relevant tenders surfaced/qualified	M6
UC-22	E-shop product description generation	MKT	Catalog content	Thousands of PPE items with thin GR/EN copy	Generate bilingual descriptions from specs as unpublished drafts	Product specs, brand tone guide	Semantic layer, e-shop API	L	L	L	Med	Products with full copy; content cost	M6

Wave-1 pilot candidates (~M3–M4)

UC-01 Internal knowledge assistant — the platform’s proof of life. Touches every employee, zero write-risk, and it is the RAG core every later use case reuses. If this doesn’t earn daily users, nothing downstream will.

UC-04 IMPA-code product matching — the most “Prime Products-shaped” quick win: requisition lists arrive daily, matching is pure lookup drudgery, output is a ranked list a human confirms in seconds, and precision is trivially measurable. High visibility with the sales floor.

UC-07 Customer email triage & draft replies — shared-mailbox pain is acute and bilingual drafting plays to the Krikri + Llama pairing. Human-approval-before-send keeps risk contained while the KPI (first-response time) is one management already watches.

UC-03 Sales quote / RFQ assistant — directly revenue-adjacent and builds on UC-02/UC-04 components (semantic layer + matching). Piloted with one product family first; drafts only, finalization stays with the rep.

UC-08 Meeting summarization — cheapest pilot on the list, feeds the knowledge base (../06-knowledge-base/obsidian-vault-blueprint.md) with real content, and gives management a felt win in week one. Low risk, high habit-forming value.

(UC-02 ERP query assistant starts as build-in-M4 rather than a formal pilot: it is infrastructure for UC-03/10/11/12 as much as a use case, and depends on the SoftOne extract verification — see dependency flag in ../05-technical/integrations-m365-softone.md.)