AI Agent Registry

What counts as an agent

An assistant answers when asked (chat over RAG — e.g. the base Open WebUI experience). An agent additionally acts: it uses tools (queries systems, drafts into workflows, sends notifications), may run unattended on triggers/schedules, and therefore needs explicit governance. Assistants are governed by platform RBAC alone; everything that acts is registered here.

Lifecycle states: proposed → pilot → production → retired. All entries below are proposed as of 2026-07-03; state changes are steering decisions recorded in the entry.

Governance (gate before any pilot, per ../02-program/governance-model.md): every agent must have (1) a named owner role, (2) defined human approval points — no irreversible or customer-visible action without one in phase 1, (3) a KPI with a baseline, (4) a risk level with matching controls (least-privilege credentials, audit logging, prompt-injection posture per ../05-technical/security-and-operations.md). Corresponding demand entries: ai-use-case-pipeline.md.

---

Knowledge Assistant

• Department: company-wide · State: proposed · Rollout: M3 (wave 1) · Owner: Knowledge manager
• Purpose: answer questions over the curated knowledge base with citations (UC-01). Registered as an agent because it also files retrieval-miss reports.
• Inputs: user questions (GR/EN) · Outputs: cited answers; retrieval-miss log entries
• Tools: RAG (Qdrant hybrid + rerank) · Data: vault, approved SharePoint, SOPs — ACL-filtered per user
• Approval points: none for answers (read-only, cited); KB gap reports reviewed by knowledge manager
• Risk: Low · KPIs: weekly active users; answer helpfulness (thumbs); time-to-answer

ERP Query Agent

• Department: sales / management / warehouse · State: proposed · Rollout: M4 · Owner: IT integration owner
• Purpose: natural-language questions over the SoftOne semantic layer (UC-02) — stock, prices, orders, receivables — with freshness stamps.
• Inputs: NL questions · Outputs: figures/tables + generated SQL shown on demand
• Tools: NL-to-SQL over PostgreSQL read replica (../05-technical/integrations-m365-softone.md) — read-only credentials, no ERP writes
• Approval points: none for reads; result rows ACL-filtered by department (e.g. margins restricted)
• Risk: Medium (wrong-number risk → SQL shown, freshness stamped, eval-set SQL accuracy gate) · KPIs: self-served queries/week; SQL accuracy on eval set

Quote Drafting Agent

• Department: sales · State: proposed · Rollout: M4 pilot (one product family) · Owner: Sales operations lead (role)
• Purpose: turn an RFQ (email/PDF) into a draft quote: matched lines, prices, availability, flags for missing info (UC-03).
• Inputs: RFQ documents/emails · Outputs: draft quote lines + confidence flags; phase 2: draft quote object in Soft1
• Tools: RAG, Requisition Matching Agent, semantic layer read, n8n flow; phase 2 SoftOne API (drafts only)
• Approval points: sales rep reviews and finalizes every quote; nothing reaches a customer unreviewed
• Risk: Medium (pricing errors) · KPIs: quote turnaround time; % draft lines accepted unchanged

Requisition Matching Agent (IMPA)

• Department: sales / warehouse · State: proposed · Rollout: M4 (wave 1) · Owner: Sales operations lead
• Purpose: extract line items from vessel requisition lists and match to IMPA/catalog codes with confidence ranking (UC-04).
• Inputs: requisition PDFs/spreadsheets/emails · Outputs: ranked match table (item → IMPA → product code → stock)
• Tools: OCR/docling, embedding similarity + fuzzy match, semantic layer read
• Approval points: human confirms matches below confidence threshold; all matches reviewable before quoting
• Risk: Low · KPIs: top-1 match precision; minutes per requisition list

Tender Extraction Agent

• Department: tenders · State: proposed · Rollout: M5 · Owner: Tenders manager (role)
• Purpose: read tender packs and produce a draft compliance matrix — requirement clauses, deadlines, certificates demanded — with page-level citations (UC-05).
• Inputs: tender document packs (often scanned, GR/EN) · Outputs: compliance-matrix draft; deadline/cert checklist
• Tools: OCR pipeline, RAG over pack, LangGraph multi-step extraction, n8n intake flow · Data: tender packs (restricted collection), past matrices
• Approval points: tenders manager validates every matrix before use; agent output is never the submitted document
• Risk: High (missed requirement = lost/invalid bid; defense-sensitive data — restricted ACL, air-gapped inference) · KPIs: requirement recall vs human baseline; hours per tender

Email Triage Agent

• Department: customer service · State: proposed · Rollout: M4 (wave 1) · Owner: Customer service lead
• Purpose: classify and route shared-mailbox email; draft bilingual replies for routine requests (UC-07).
• Inputs: inbound email (approved shared mailboxes) · Outputs: category/priority tags, routing, reply drafts in the drafts folder
• Tools: Graph API (scoped mailboxes), prime-fast for triage, prime-greek/prime-large for drafts, RAG, ERP read for order status, n8n
• Approval points: agent never sends — humans send drafts; auto-routing reviewed weekly in pilot. Inbound content treated as untrusted (injection posture).
• Risk: Medium · KPIs: first-response time; % drafts sent with ≤ minor edits; misrouting rate

Meeting Scribe Agent

• Department: company-wide · State: proposed · Rollout: M3 (wave 1) · Owner: Knowledge manager
• Purpose: transcribe meetings (Whisper, on-prem), summarize decisions/actions, draft a vault meeting note (UC-08).
• Inputs: Teams recordings/uploads · Outputs: transcript, summary, vault note draft (git branch)
• Tools: Graph API, Whisper large-v3, n8n, vault git
• Approval points: meeting owner approves the note before merge; recording requires participant awareness (GDPR notice)
• Risk: Low–Medium (privacy — retention limits per ../05-technical/rag-and-data-pipeline.md) · KPIs: % meetings with published notes; owner edit distance

SOP Drafting Agent

• Department: quality/compliance · State: proposed · Rollout: M4 · Owner: Quality manager (role)
• Purpose: turn process-interview transcripts into structured SOP drafts in vault format (UC-09).
• Inputs: interview recordings/transcripts, SOP template · Outputs: SOP draft (vault branch)
• Tools: Whisper, prime-large, vault git, n8n
• Approval points: process owner + quality manager approve before the SOP becomes citable KB
• Risk: Low · KPIs: SOPs published/month; review cycles per SOP

Supplier Comparison Agent

• Department: procurement · State: proposed · Rollout: M5 · Owner: Procurement lead (role)
• Purpose: normalize multi-format supplier quotes into a comparison table; flag price anomalies vs history (UC-06).
• Inputs: supplier quotes (PDF/Excel/email) · Outputs: comparison table, anomaly flags
• Tools: OCR/docling, extraction prompts, semantic layer read (price history), n8n
• Approval points: buyer makes the award decision; agent output is decision support only. Supplier docs treated as untrusted input.
• Risk: Medium · KPIs: comparison prep time; extraction field accuracy

Receivables Reminder Agent

• Department: finance · State: proposed · Rollout: M5 · Owner: Finance lead (role)
• Purpose: generate tone-tiered bilingual reminder letters from receivables aging (UC-19).
• Inputs: aging report (semantic layer), customer history · Outputs: reminder drafts queued for approval
• Tools: semantic layer read, prime-greek/prime-large, n8n approval flow (Teams adaptive card), Graph send-on-approval
• Approval points: finance approves every letter before send (Teams approval button); escalation-tier letters require finance lead specifically
• Risk: Medium–High (customer relationship) · KPIs: % overdue accounts contacted on schedule; DSO trend

Onboarding Agent

• Department: HR · State: proposed · Rollout: M5 · Owner: HR lead (role)
• Purpose: guide new hires through role-specific onboarding: answers over HR policy, checklist nudges, first-week scheduling help (UC-13).
• Inputs: new-hire role/start date, questions · Outputs: answers with policy citations, checklist reminders (Teams)
• Tools: RAG (HR collection — restricted ACL), n8n scheduled nudges, Graph calendar read
• Approval points: HR approves checklist templates; agent never answers individual salary/contract questions (hard-scoped, deflects to HR)
• Risk: Medium (HR data — strict collection ACL) · KPIs: ramp-up time; HR interruption count

Report Narrator Agent

• Department: BI / executive · State: proposed · Rollout: M6 · Owner: Program lead / finance lead (role)
• Purpose: draft the monthly management pack narrative from semantic-layer KPIs; answer drill-down questions in review meetings (UC-10, UC-12).
• Inputs: monthly KPI tables, prior packs · Outputs: narrative draft with every figure traceable to a query
• Tools: semantic layer read, ERP Query Agent, n8n schedule · Data: ERP KPIs (restricted — management group)
• Approval points: finance validates figures, program/GM approves narrative before circulation
• Risk: Medium (a wrong number in front of the board) · KPIs: days to produce pack; correction count post-approval

Helpdesk Agent (IT)

• Department: IT · State: proposed · Rollout: M5 · Owner: Platform engineer
• Purpose: first-line internal IT support over runbooks and ticket history; guided fixes; escalates with collected context (UC-17).
• Inputs: employee IT questions · Outputs: guided solutions, pre-filled escalation tickets
• Tools: RAG (IT KB), n8n ticket flow
• Approval points: no system changes ever — advice and ticket-drafting only
• Risk: Low · KPIs: ticket deflection rate; escalation quality rating