AI Native ODA: The path to open digital autonomy

1. Problem statement

Autonomous, interoperable AI agents have the potential to transform telecom operations - reducing costs, improving customer experience, and enabling new growth. Demand for connectivity tailored to AI workloads creates new revenue opportunities (“networks for AI”), while AI-driven automation (“AI for networks”) is essential to deliver these services efficiently and profitably.

However, most AI deployments in telecoms today focus on isolated tasks (e.g., ticket triage, RAN optimization, chatbots). While these deliver incremental efficiencies, they create fragmented, siloed solutions that increase technical debt, duplicate logic, compromise human control, and limit scalability. As a result, they cannot coordinate decisions across customer, service, and network domains, preventing progress toward higher levels of autonomy. Achieving true autonomy requires end-to-end, cross-domain workflows that integrate data, decisions, actions and control points across highly complex, multi-vendor IT and network environments. This introduces a fundamental challenge: AI must operate in a decentralized, agent-driven model, while maintaining unified governance over security, guardrails, compliance and costs.

Without a common architectural foundation, CSPs will be forced to manage multiple proprietary AI platforms and incompatible agent frameworks, leading to fragmented automation, inconsistent outcomes, and increased risk. To overcome this, the industry requires a shared, open architecture that enables modularity, composability, and interoperability, allowing AI agents to reason over common data, share a common understanding of telecom business processes, and coordinate actions across domains while operating within a consistent governance framework. This goal can be described as "open digital autonomy". ODA (TM Forum's Open Digital Architecture) is already evolving into this unified foundation, but there is still work to be done. The principles set out below will guide a collaborative work program to build AI Native ODA, with fully developed support for AI-driven operations across telecoms IT and networks.

2. Guiding principles

2.1 General-purpose AI models are not sufficient for use in telecoms ecosystems

Specialized terminology, complex long-running processes, massively multi-vendor environments and the potential for catastrophic consequences of AI mistakes are not unique to telecoms, but all of these factors combined creates a uniquely demanding environment for AI adoption at scale. As a result, telecoms requires AI foundations that go beyond general-purpose approaches. ODA must provide this telecoms-specific foundation through shared semantics, interoperability, robust governance, consistent data access, lifecycle management and control mechanisms, enabling AI-driven operations that are trustworthy, efficient, scalable, and aligned to real business outcomes.

2.2 AI should be designed in, not bolted-on

Scaling beyond siloed AI use cases requires AI to be designed into the architecture, but this must be achievable through evolution not revolution. The opportunity and appetite for greenfield transformation in our industry is rare. Therefore, ODA must fully integrate autonomous agents and intent-driven, closed-loop automation into its broader standards and frameworks, ensuring consistent, reusable solutions for both IT and networks. AI native transformation is a journey, and ODA must provide not only a "To-Be" architecture blueprint, but also a toolkit to guide the transformation.

2.3 AI will work alongside and within existing systems

AI must be built on robust, composable IT foundations. Systems of record (ODA Components) remain essential, providing data to agents for reasoning and exposing their capabilities for agents to take reliable deterministic action. Open APIs remain essential to support this capability exposure (e.g., via MCP and other mechanisms). ODA Components may become agentic, incorporating reasoning and learning capabilities while retaining their standardized functional scope and data ownership. At the same time, external agents will work alongside ODA Components, replacing hardwired processes, rules-based automation and human dependencies with AI-driven autonomy.

2.4 Multi-vendor interoperability requires standards

Telecoms processes span multiple IT and network domains, cloud platforms and vendor software stacks. Agent interoperability across these boundaries requires shared semantic understanding, not merely shared context (existing protocols allow agents to share context and state, but this alone does not deliver trustworthy agent interoperability). Therefore, ODA must provide a common semantic grounding for agent interoperability, building on the structure provided by SID, eTOM, the Functional Framework, Open APIs, ODA Components, Capability Framework, Value Streams and other core ODA standards, enabling agents to reason accurately, express intent consistently, and take reliable, trustworthy actions.

2.5 AI operations require strong, unified control over security, compliance and costs

AI must operate with full human governance, control and accountability for security, compliance and costs. Therefore, ODA must provide standardized AIOps architectural patterns and reference implementations for operating and managing agents in a telecoms environment, securing their access to data and their interactions with other agents, components and humans, enforcing guardrails, making AI decisions observable and explainable, and allowing token consumption and costs to be proactively managed.

2.6 Agent ecosystems operate through peer-to-peer collaboration and self-organization

While a central orchestrator may define high-level goals and boundary policies, agents must be empowered to self-organize and act without a central brain to orchestrate processes. Within ecosystems, independent agents collaborate across organizations with no central orchestration authority. Governance is exercised through shared semantic contracts, interoperable policy enforcement, and mutual trust mechanisms, not hierarchical control. Therefore, ODA must provide ontology-driven semantics and a flexible framework for agent orchestration to enable interoperability, policy enforcement and trustworthy autonomous collaboration.

2.7 Agentic workflows are triggered by events in real-time

Agents respond to real-time events from operational systems, coordinating their response through intent negotiation. Situations without design-time workflows are resolved through autonomous agent collaboration, with outcomes recorded for future pattern recognition. Therefore, ODA must support decoupled, Event-Driven Architectures enabling agents to be triggered in real-time by events from the network and IT systems, and to respond by negotiating intents that solve unscripted problems for which there are no pre-wired processes.

2.8 AI is evolving too fast for traditional standardization to keep up

Telecoms must be flexible to use the latest AI advances. Therefore, ODA must be able to rapidly incorporate and adapt to new protocols and technologies, defining telecoms extensions rather than creating telco-specific alternatives.

2.9 Agents are diverse, but reusable patterns will be standardized

Developers will design agents for all kinds of tasks, large and small. This implies that agents will not be standardized in the same way as ODA Components (with a normative, non-overlapping functional mapping), but ODA can still usefully provide a standardized agent taxonomy, a standardized library of reusable Agent Skills, and emerging best practices for dividing common telecoms tasks among agents.

2.10 Success with AI depends on humans

AI enhances - rather than replaces - human judgment, creativity, and strategic intent, but it has a profound effect on the ways humans work and the skills they need. Therefore, ODA should provide best practice guidance to help address the people, culture and organizational challenges in transforming to AI native operations.

2.11 The goal is business value, not autonomy for its own sake

AI will not be applied to every problem in telecoms; only to those where the business case for autonomy is clear and AI delivers significant benefits over rules-based automation. TM Forum's industry missions will clarify, prioritize and champion these use cases, and ODA's maturity measurement tools will not only determine standardized levels of autonomy, but will also establish the correlation between autonomy and business value.

2.12 Agents will become the main consumers of ODA's standards, frameworks and guides

Agents will assist humans in understanding and adopting ODA, consuming ODA's machine-readable and unstructured assets faster and in higher volume than humans. Therefore, ODA's assets should be packaged as data products that can be consumed by agents; readily available for ingestion as knowledge sources for the AI tools used by architects, software developers and business planners, helping them to design and build systems that conform to ODA's standards. For example, ODA should provide tools to support the AI-enabled software engineering lifecycle (design, build and operate), consistently transforming business requirements into ODA-compliant code.

3. What AI Capabilities does ODA already have?

ODA is already evolving to provide the architectural patterns, standards, governance, ontologies and best practices for enabling AI-driven intent-based autonomy across IT and network domains. For example, ODA already provides:

• A standardized component-based software architecture as an essential foundation for AI
• Common language for grounding AI with telecoms understanding, including SID, eTOM, business capabilities, value streams, and the TM Forum Intent Ontology
• Support for agent integration with existing systems via standardized Open APIs, with MCP support in ODA Component specifications and the ODA Canvas
• Reference architectures for intent‑driven closed‑loop automation across autonomous domains
• ODA Canvas extensions to support automated lifecycle management of AI agents alongside ODA Components
• ODA Canvas Operators that enable centralized management and control over AI, including:
  • Secure agent interactions with other agents, ODA Components and humans
  • Consistent enforcement of guardrails
  • Managed agent access to data products (with Data Product Lifecycle Management), enabling Canvas-configured agent access to knowledge sources, memory and state information
  • Managed agent access to models (Model-as-a-Service), enabling cost control and enforcement of AI sovereignty
• A reference implementation supporting demonstrations of Level 4+ autonomy for High-Value Scenarios enabled by autonomous agents and ODA Components running on this AI‑Native Canvas foundation
• Telecoms extensions for the open source A2A protocol
• Reference architecture, use cases and implementation guidance for the use of digital twins in telecoms
• Tools and metrics for measuring AI readiness, maturity and standardized levels of autonomy
• Guidance and best practice for AI and data governance, as well as people, organization and culture transformation for AI.

4. Priorities for the AI Native ODA Roadmap

While the architectural patterns for AI are becoming clearer, there is still work to do. The AI Native ODA Roadmap will prioritize the following[1]:

• Full support for event-driven architectures in the ODA Canvas and async Open APIs, with native publish / subscribe capabilities enabling ODA Components and autonomous agents to operate seamlessly across both synchronous and event-driven interaction models
• Telecoms BSS / OSS ontologies based on ODA's frameworks (e.g., SID, eTOM), providing common semantic understanding enabling accurate agentic reasoning, decision-making and explainability
• Standardized architectural patterns and reference implementations for agent-to-agent communications, building on existing and emerging agent communication protocols
• Agent-ready Open APIs - turning ODA standards into a resource any AI agent can use, enabling agents to understand what each API does and when to use it without human prompting
• A standardized BSS / OSS orchestration layer for AI-native networks, with AI-native Network Functions and Autonomous Networks Solution Packages managed as ODA components on the ODA Canvas leveraging TMF Intent APIs
• AI & data sovereignty as a design requirement for ODA, making its frameworks 'sovereign aware' (e.g., sovereign orchestration patterns and jurisdiction profiles defined as standardized, deployable ODA Canvas operators)
• ODA as a data product for RAG / graph ingestion into developer environments and CSP knowledge systems
• Expansion of the ODA Canvas to support the full software development lifecycle, including design-time, build-time and run-time environments
• Tools for developers that ensure ODA conformance throughout the DevSecOps lifecycle (e.g., an Agent Skills Library, continuous conformance testing as part of DevSecOps, agents that fix as well as test compliance to TMF standards...)
• A shared network digital twin repository in the Innovation Hub as a model of a reference network topology for testing innovations, pre-populated with synthetic but realistic network state data
• Intervention & control for agentic systems, enabling humans to constrain, correct, and terminate AI agents at any point, for any reason, irrespective of their level of autonomy under which they operate
• Guardrail reference implementations
• Establishing mechanisms for coherent governance when agentic systems cross from a Canvas-managed environment into an open ecosystem
• Expansion of observability in the ODA Canvas to include audit trails as proof of trustworthiness, and tracking / retrieval of agent reasoning and decisions
• Extension of the Autonomous Networks High-Value Scenarios into IT and business (Party Management, Core Commerce Management, Engagement Management), including solution packs, autonomy level evaluation tools and metrics
• Transformation guides covering both AI native transformation and the use of AI to accelerate traditional digital transformation
• Best practice guidance to help address the people, culture and organizational challenges in transforming to AI native operations.

At the same time, the ODA roadmap will continue to mature and extend the existing ODA assets to ensure solid IT foundations for AI, while at the same time looking further ahead at the future impact of new technologies1:

• Completion of the ODA Component specifications in the Party, Core Commerce, Production and Engagement Management Function Blocks, including support for v4 / v5 API optionality, CTKs for all ODA Components, and taxonomy for other types of software application running on the ODA Canvas
• Further maturing and extending the ODA Canvas, including support for carbon management, cost management, and multi-Canvas federation across distributed clusters
• Open API specialization for ODA Component capability exposure and cross-enterprise use cases
• Extension of the ODA use case library and other ODA assets to improve support for B2B2X scenarios
• Resolving inconsistencies between existing frameworks that have been developed independently and now need to come together in a single ODA toolbox, for example:
  • Formally integrating assets originally developed outside core ODA (e.g., by the Autonomous Networks project and Digital Ecosystem Management project) into ODA, reconciling any differences in concepts and terminology
  • Bringing all maturity models into a common hierarchy
• Anticipating the impact of future technologies such as quantum computing (designing "Quantum ODA")