Agent-based automation must bridge AI and structured knowledge to improve inferences and optimize outcomes
Based on an interview with James Crawshaw, Principal Analyst and Practice Lead for Service Provider Transformation at Omdia.
Dave Duggal, founder and CEO of EnterpriseWeb, notes that at the core, the recent buzz around Agentic-AI is about three things:
1) transforming user experiences with conversational AI;
2) reinventing static, unresponsive manually-designed business processes into dynamic, contextual workflows dynamically assembled by agents and;
3) leveraging agent-based automation to unleash productivity.
Duggal is a pioneer in the field with his own enterprise-grade solution for agent-based, ontology-driven automation. The outspoken entrepreneur has been evangelizing these ideas for a decade. The company is a pioneer in Telecom virtualization and automation. It led the first ETSI NFV proof-of-concept and the first ETSI Zero-touch Network and Service Management proof-of-concept.
Last year, EnterpriseWeb won a Light Reading Leading Lights award for the first industry presentation of “Telco-grade Generative AI for Intent-based Orchestration” in collaboration with Microsoft. It demonstrated how developers can “talk to the network” to design, deploy and manage complex services – no-code and zero-touch.
Duggal is happy to see generative AI and now “Agentic” AI drive interest in conversational AI and autonomous networking. “GenAI has raised expectations of what human-computer interactions could be and is helping to enable the next generation of agent-based automation.”
He explained that, “the current wave of agent-based automation, Agentic AI, is closely associated with generative AI. It places Large Language Models (LLMs) at the center of the solution to support both natural language interactions and intelligent behavior. However, LLMs aren’t good enough to support complex, enterprise-grade business processes or system workflows. They simply aren’t accurate or consistent enough to be trusted.”
Duggal is far from alone in this critique. In academia, LLMs have been derisively labelled as “Stochastic Parrots” in that they don’t have the power to reason, they aren’t governed by any logical and transparent model. Rather, Generative AI approximates human intelligence by processing ungodly amounts of data and consuming incredible amount of compute and energy just to return a probabilistic response. LLM outputs range from incredibly responsive to wildly incorrect, irrelevant and possibly offensive “hallucinations”. Mind you, Duggal notes, “hallucinations aren’t technically bugs, they are just unexpected LLM outputs that are unrelated to the user inputs. No amount of training, tuning, prompts or tokens can guarantee a factually accurate or even a high-relevant LLM output.”
Duggal believes that, as with GenAI hype in general, “there are no practical enterprise-grade agentic automation solutions”. He noted that most of what is referred to as “agentic” with LLMs and function calling ends up being simplistic prototypes – coded agents representing a targeted workflow or Finite State Machine with manual integrations to a static set of functions. In development terms, he believes they are “closer to microservices or actors than a rich conceptualization of an intelligent agent. The reasoning and planning capabilities are limited.”
Duggal warns, “you can’t extrapolate from the initial Agentic implementations and assume you’ll ever realize truly complex enterprise-grade processes.” According to Duggal, the current Agentic approach is “naïve” and doesn’t come close to supporting the inherent complexity of telecom networks.
Duggal acknowledges that some informal tasks can leverage unmodeled, unstructured LLM-centric approaches. However, he says those responses won’t be consistent, predictable or explainable. He asks “How will telcos manage collections of agents and consistently enforce IT governance and business compliance?”. Duggal believes it is unwise if not outright irresponsible to simply “trust” an LLM. Duggal argues that while “some use-cases may have lower-requirements for correctness and safety, telco operations still require deterministic transactions and system controls to ensure reliable and available networks.
Inference by definition is “fuzzy”. It fills gaps in knowledge, but it doesn’t replace knowledge. In Duggal’s view the best software solutions will be a hybrid approach, bringing together unstructured probabilistic approaches from GenAI and classic AI/ML with structured domain knowledge. He feels Telcos should focus on how to combine AI with domain knowledge to improve inferences and optimize decisions. Duggal explains that, “knowledge is generally modeled in ontologies, which are graphs of linked concepts, types and policies that describe domain semantics”.
Duggal uses the human brain as an analogy. “The combination of ontologies and AI is like using the two hemispheres of our brains, one rational and one creative. They work together so we can reason based on existing knowledge, while also being able to make inferences to fill in gaps in what we know and allow us to make predictions about future events.”
As to agent-based automation, Duggal posits that workflows can be deconstructed into a set of loosely-coupled tasks, like network functions in a service chain. Instead of a static flowchart or pipeline, you are left with a logical model of a process where tasks are connected by policies. Agents can then dynamically assemble and configure tasks based on conditions. Since the process is dynamic, agents can leverage domain knowledge and real-time operational context, along with inferences, to optimize individual tasks and next best actions.
According to Duggal, this is exactly what EnterpriseWeb’s platform does.
EnterpriseWeb’s offering for Telcos, Netwrx.ai, features a telecom ontology (i.e., a harmonized, standards-based, graph-connected industry model). EnterpriseWeb’s agents leverage the ontology to interpret events, evaluate conditions, automate decisions and optimize actions. The agents can perform an advanced form of ‘Function Calling’ to leverage internal tools and external systems, services, databases and devices.
In contrast to Agentic approaches, Duggal boasts that EnterpriseWeb’s approach supports “highly-reflective and deeply recursive reasoning”. Put simply, the agents can leverage the telecom ontology to access a wide variety of information and reach out to distributed solution elements, so the agents are situationally aware. The agents can also generate service templates, deployment workflows, Kubernetes operators and management plans directly, without an LLM. Instead, EnterpriseWeb’s agents use the ontology to drive precise outputs.
Duggal points out that the agents are handling complex implementation details behind the scenes on behalf of developers. Developers can simply “Talk to the Network” to design, deploy and manage complex services. The agents use the ontology to deterministically translate the developer’s “intent”.
EnterpriseWeb: Telco-grade GenAI for intent-based Orchestration (demo #2, 092523)
EnterpriseWeb: GenAI for Business ETL and Code Support 062624 v2
EnterpriseWeb: Secure Dev Centric Networking CAMARA APIs and Intel TDX
