Share with your CTO
A TM Forum Catalyst project, showcased at DTW Ignite in June 2026, is betting that the missing layer in telecom AI isn’t a better model but a governed execution framework that makes agentic AI, meaning AI that acts rather than just advises, safe enough to touch live BSS and OSS systems. The Essential Framework for Telecom Agentic AI proposes a reusable reference architecture built on TM Forum Open APIs, a Cedar policy engine, and a hard human-confirmation gate before any write action executes. Projected gains run 15-20% on customer service cost and 30-40% on productivity across care and network operations.
What this means for your business
Most telecom AI programs have quietly accumulated the same technical debt: each new use case gets its own bespoke integration into billing or provisioning systems, which means every proof of concept that graduates to production is a one-off. The operators most exposed here aren’t the ones who’ve moved slowly on AI. They’re the ones who’ve moved fast without an execution layer underneath, and now own a portfolio of demos that can’t share infrastructure, policy, or audit trails.
The framework’s most consequential design choice is the Subscriber Context Token, a structured credential minted from an authoritative subscriber profile that agents read from instead of hitting the database directly. This is a meaningful architectural boundary, not a formality. It’s the difference between an AI agent that can act on customer data and one that can exfiltrate it. Pairing that with a Cedar policy engine that evaluates every tool call, not just the session, closes the gap where most enterprise AI governance frameworks fall apart, which is that session-level permissions don’t constrain what happens inside the session. TM Forum, whose business model depends on operators adopting its standards, has an obvious incentive to frame this as a TM Forum-native problem requiring TM Forum-native APIs, and that tilt probably overstates how cleanly the architecture maps onto operators running hybrid ODA and legacy stacks. The governance primitives, though, are sound regardless of the standards wrapper.
The proposed “agentic Capabilities” extension to ODA Canvas component definitions is the part of this worth watching most closely. If BSS vendors actually adopt the spec, it means AI-callable capabilities become a declared, discoverable property of a component rather than a custom integration every buyer has to negotiate. That shifts procurement leverage: operators can require tool-catalog compliance in RFPs the same way they require API conformance today. The vendors who get ahead of that declaration will own the top of the agent discovery stack. The ones who wait will find themselves effectively invisible to orchestration layers they didn’t design for.
Concept deep-dive: Cedar policy engine
Cedar is an open-source policy language, originally developed by AWS, designed to evaluate authorization decisions at the level of individual actions rather than just roles or sessions. Think of it as a traffic cop stationed at every single API call, not just the front door. In an agentic context, where an AI might chain dozens of tool calls inside a single session, session-level access controls are insufficient. Cedar’s per-invocation evaluation is what makes a hard audit record credible rather than approximate.
Based on reporting from Turning agentic AI into a safe, scalable operational capability, originally published 2026-07-30 10:43:00.
