Evidence-led demos for legacy BSS problems.
Exos can turn ERPNext, CaveauAI, RAG, MCP, and Microsoft-ready agent tooling into a working standards map: record to TMF resource, ODA component, eTOM process, MCP tool, harness test, and human approval rule.
Data, proof, and a path to governed action.
The demo should not ask a buyer to believe an AI claim. It should show the source record, the retrieved evidence, and the approval boundary.
Order fallout - equipment unavailable
Operational data: Greenfield Apartments order blocked because Central Depot Oslo has 0 DOCSIS 3.1 modems; Bergen Depot has 12.
AI/RAG use: RAG retrieved Support Sherpa verified run 212 and the scenario code that says a blind resubmit would fail again.
Standards map: TMF622 Product Ordering / TMF641 Service Ordering -> Fulfillment, Order Handling
End result: Propose a stock-reallocation ticket. Human approval required before any ERPNext write.
Enterprise bill shock
Operational data: Meridian Media Group AS invoice jumps from NOK 12,000 baseline to NOK 44,000, including NOK 32,000 roaming overage.
AI/RAG use: RAG retrieved invoice seed data, verified run 211, scenario JSON, and roaming guidance for dispute context.
Standards map: TMF666 Party Bill / TMF678 Customer Bill -> Billing, Bill Inquiry Handling
End result: Explain the bill with line-item evidence and draft a goodwill-credit review ticket for approval.
Transient OSS timeout
Operational data: Northstar Fiber order failed because upstream provisioning missed a 120-second acknowledgement window; stock and serviceability passed.
AI/RAG use: RAG distinguishes transient retry-safe fallout from equipment-blocked fallout.
Standards map: TMF641 Service Ordering -> Fulfillment, Order Handling
End result: Propose recreate-telecom-order only after evidence shows retry is safe; approval gates the write.
Overnight SLA sweep
Operational data: The order book is scanned for Fallout and near-breach items against a 4-hour jeopardy threshold.
AI/RAG use: RAG retrieved the verified SLA sweep result: autonomous inform only, citing real ORD-70021.
Standards map: TMF629 Customer Management / SLA context -> Assurance, Service Quality Management
End result: Send an FYI digest to Teams/Copilot. No customer-visible message and no write action.
Broadband fault diagnosis
Operational data: Sure Guernsey Broadband Demo reports hourly disconnects with xDSL sync at 2,100 Kbps versus 18,000 Kbps expected.
AI/RAG use: RAG retrieved the support scenario and open ERPNext issue context.
Standards map: TMF621 Trouble Ticket / service assurance -> Assurance, Problem Handling
End result: Draft wholesale fault ticket with measured evidence, then wait for operator approval.
From question to approved action.
The workflow is designed so the model can explain, cite, and propose - but customer-visible writes stay behind human approval.
Ask
Operator asks in Exos chat or a Microsoft agent surface: "Why did this order fail?" or "Explain this high bill."
Retrieve
CaveauAI searches Exos private knowledge plus TM Forum, ERPNext/Frappe, BSS, support, and playbook corpora.
Ground
The answer cites exact records: orders, invoices, issues, depot stock, support runs, and standards references.
Map
The evidence is mapped to TMF API/resource, ODA component, eTOM process, MCP tool, and harness expectation.
Gate
Read-only answers can complete. Any customer-visible write becomes a pending approval action first.
Six slices that turn legacy BSS pain into evidence-led outcomes.
Bill Shock Copilot
- Legacy problem
- Billing teams lose time proving what changed and whether the customer needs an explanation, dispute, or credit review.
- Data shown
- Sales Invoice, prior-month baseline, usage line items, related Issue, regulatory/roaming guidance.
- AI/RAG work
- RAG retrieves invoice lines, compares periods, explains the high-bill reason, and classifies whether a credit action is financial risk.
- Microsoft-ready path
- Teams/Copilot Studio front door -> Streamable MCP read tools -> pending TMF Agent Action for credits.
- End result
- Answer now; write later. The customer explanation is cited, and the credit ticket is approval-ready.
Order Fallout War Room
- Legacy problem
- Legacy BSS fallouts look identical in queues even when one needs stock reallocation and another is safe to retry.
- Data shown
- ERPNext order, depot inventory, supplier stock, serviceability, fallout reason, audit log.
- AI/RAG work
- RAG compares current fallout with verified prior scenarios and selects cause-specific next action instead of generic retry.
- Microsoft-ready path
- Foundry/Copilot agent calls MCP tools; Exosphere maps to TMF622/TMF641; approval policy gates execution.
- End result
- The operator sees why the order failed, what evidence supports it, and what approval will do.
Partner + Supplier Assurance
- Legacy problem
- Supplier, depot, subcontractor, and SLA context often sits outside the customer-care view.
- Data shown
- Depot stock, purchase order, purchase receipt, field tech capacity, service impact, SLA jeopardy.
- AI/RAG work
- RAG builds a supplier-aware narrative and ranks impact by customer, SLA, and recoverability.
- Microsoft-ready path
- MCP read tools for supplier/depot/field data; Teams digest for operators; no write unless an action is approved.
- End result
- Fulfillment, partner management, assurance, and care become one evidence chain.
Catalog Modernization Map
- Legacy problem
- Product catalogs in legacy BSS are hard to map to TMF resources, bundles, pricing, and service implications.
- Data shown
- ERPNext Item/Product Bundle, Product Offering, service spec, eligibility rule, migration notes.
- AI/RAG work
- RAG matches product terms to TM Forum resources and flags unmapped or ambiguous catalog claims.
- Microsoft-ready path
- Copilot explains catalog changes; MCP read tools fetch catalog records; harness tests validate mapping claims.
- End result
- Green/yellow/red coverage for product catalog readiness instead of a one-off spreadsheet.
Approval-First BSS Actions
- Legacy problem
- AI demos are easy until a tool can change a customer bill, order, ticket, or communication.
- Data shown
- Action type, impacted customer, source evidence, risk tier, approver, execution log, rollback note.
- AI/RAG work
- RAG provides the why; policy decides if the action is read-only, inform-only, propose-and-wait, or never.
- Microsoft-ready path
- Foundry approval controls, Copilot action governance, Entra identity, and Exos pending TMF Agent Action records.
- End result
- Operators can trust the demo because customer-visible writes are never invisible.
Legacy BSS Knowledge Twin
- Legacy problem
- System knowledge lives in vendor docs, tickets, Confluence, SQL, screenshots, and memory.
- Data shown
- Migration playbooks, TM Forum docs, ERPNext records, support tickets, case studies, old interface notes.
- AI/RAG work
- RAG assembles an evidence graph so each modernization claim shows source, confidence, and missing proof.
- Microsoft-ready path
- CaveauAI retrieval profiles, Document Explorer classification, MCP tool harness, and Copilot-ready citations.
- End result
- A standards map that tells the buyer what is proven, what is assumed, and what still needs evidence.
Built for the Microsoft conversation without hiding the controls.
The first interaction can stay in Exos chat today, then be packaged into Microsoft surfaces through MCP and enterprise governance.
Copilot Studio
Operator-facing agent in Teams or web chat, connected to Exos MCP through Streamable MCP.
Azure AI Foundry
Agent orchestration and approval policy story for enterprise governance and sensitive actions.
Microsoft 365 Admin Center
MCP connector publication path for governed enterprise discovery and adoption.
Entra + API Gateway
Identity, tenant controls, rate limits, logs, and enterprise security around every tool call.
The demo line is simple: every claim has evidence.
ERPNext supplies the operational truth, CaveauAI retrieves the supporting knowledge, MCP exposes governed tools, the harness proves expected behavior, and human approval decides whether an action becomes real.
Discuss an evidence twin demo