genalpha-bss · Proof of Concept

Autonomy Accelerated

Selling the network edge as an AI platform — the full lead-to-assure loop, AI at every step, running on a vendor-neutral BSS with a thin autonomous OSS.

This PoC re-creates, on genalpha-bss, the pattern shown in the TM Forum "Autonomy Accelerated" Catalyst: a 5G AI slice sold and assured end to end. The premise is physics. Light through fibre is fast, but not fast enough for a sub-10-millisecond round trip; route that workload to a distant cloud and the latency budget is gone before any code runs. The fix is GPU compute at the edge, paired with an SLA-backed slice — so AI inference happens next to where the data is born.

The scenario: a football tournament. VIP fans in AI glasses, fed by edge GPUs — real-time player overlays, AI commentary, highlight reels within minutes of the final whistle. None of it reachable from a centralised cloud inside the latency constraint.

The loop, and who acts at each step

The telling part isn't the use case — it's watching the whole lead-to-assure loop run with AI at every step, and no swivel-chairing between systems.

Prepare the approach. A B2B sales agent (Claude, over MCP) drafts the sales approach and turns a plain-language ask — "a 5G slice with AI glasses for the stadium-north tournament" — into a structured network intent. AI agent → BSS
Capture intent, no forms. The intent lands in the BSS through an AI chat seam; a TMF921 intent expression is created — place, latency budget, bandwidth, AI token volume. BSS · TMF921
Autonomous feasibility. The orchestration layer checks feasibility on its own: a sub-20ms round trip cannot be served from regional cloud, so it is only feasible where an edge GPU pool covers the venue. The network doesn't just say yes — it proposes edge AI inferencing as an upsell, more than the customer asked for. OSS · TMF921 + TMF685
AI-assisted quote, token pricing. The BSS prices the proposal into a TMF648 quote: the slice as a recurring charge, and the AI service metered in tokens — an included allowance, then a fixed price per million. An LLM drafts the executive narrative. BSS · TMF648 + TMF635
Sold → provisioned. Accepting the quote places the product order; the thin SOM provisions the slice live onto a fibre delivery path. BSS → OSS · TMF622 + TMF641
Mid-match: a fibre cut. A critical alarm hits the delivery path. Assurance streams the KPI and opens a service problem. OSS · TMF642 / TMF656
The system works out the fix on its own. Every slice on the failed path is re-homed from fibre to the edge site, the SLA is restored, an ITSM ticket is auto-generated to document it, and the problem closes itself — no human in the loop. OSS → ITSM · self-heal

The real story isn't any single agent. It's that an AI-driven sales flow in the BSS, an autonomous OSS, and an ITSM ticketing step interoperate as one workflow — the standardised agent-to-agent tissue (MCP / A2A) the industry is converging on. Here the BSS is simply a set of tools an agent can hold, authenticated as a tenant-scoped machine identity like every other caller.

What runs it

Step	Component	API
Intent from language	intelligence	/ai/v1/intentDraft
Intent + feasibility + upsell	service-orchestration	TMF921 · TMF685
Slice & token-metered AI	product-catalog · usage	TMF620 · TMF635
Priced quote + AI narrative	quote	TMF648
Order & provisioning	product-ordering · SOM	TMF622 · TMF641/638/640
Alarm → problem → self-heal	assurance	TMF642 · TMF656
Auto-ticket the incident	trouble-ticket	TMF621
Agent-to-agent tissue	mcp-server	MCP / A2A

Drive it yourself, as an AI agent

The integrations/mcp-server exposes the loop as MCP tools. Point Claude Desktop or Claude Code at it and ask, in plain language:

"A stadium wants a 5G slice with AI glasses for their tournament — draft the intent, check feasibility, quote it and place the order."

Claude chains five tools — draft_intent → submit_intent → create_quote → accept_quote — and closes the commercial loop. Everything below is a real run through the gateway:

1. draft_intent  -> stadium-north  8ms  AI 80 Mtok
2. submit_intent -> feasibilityChecked | proposes 5g-slice + edge-ai-inferencing
3. create_quote  -> 5890 EUR/mo  (slice 4900 + edge-AI 990; 50M tokens incl, 4 EUR/M over)
4. accept_quote  -> accepted | product order placed → slice provisioned

   [ fibre cut on fibre-route-stadium-north ]
   -> slice re-homed to edge:gpu-site-stadium-north — SLA restored
   -> service problem auto-resolved; ITSM ticket "SLA protected: … re-homed to edge" filed

Honest scope

This is a PoC on a demo stack, and it says so. The edge GPU pool, the slice activation and the "fibre cut" are modelled (the SOM is a thin stand-in for a real network controller; TMF640 activation is mocked). What is real is the shape: standardised TM Forum APIs as the connective tissue, an intent-driven autonomous feasibility-and-upsell step, token-based pricing rated by the usage engine, a genuine self-healing assurance loop, and an MCP surface any AI agent can drive — all multi-tenant, all event-backed, all verified by an end-to-end browser test (ops/e2e/ai_slice_test.js). Swap the mock activator and alarm source for a real OSS and the commercial and assurance logic is unchanged.

genalpha-bss · vendor-neutral BSS on TM Forum ODA · github.com/kundanvarma/genalpha-bss