The term ‘AI agent’ gets applied to an extraordinarily wide range of capability in telecom today, from a chatbot that drafts a suggested response for a human to approve, to a system that independently diagnoses a network fault, decides on a remediation, and executes it without anyone signing off. Those two things share a label but differ enormously in what they actually do and what risk they carry. TM Forum’s Autonomous Networks levels framework, now the industry-standard reference for this distinction, gives a structured way to place any specific AI agent claim on a spectrum from assistance to full autonomy, rather than taking the ‘AI agent’ label at face value.
The Core Distinction: Suggestion, Execution, and Autonomy
At the lowest level of automation, a system observes and suggests: it analyses data and presents a recommendation, but a human decides whether and how to act. This is assistance, not autonomy, and it’s where a large share of customer-facing and network-operations AI still sits. The next level up is execution under supervision: the system carries out a predefined action, but only after a human approves that specific instance, or within tightly bounded conditions that a human has pre-authorised. Genuine autonomy, in the TM Forum sense, means the system senses a condition, decides on a course of action, and executes it independently, with human involvement limited to setting the policy boundaries in advance and reviewing outcomes after the fact rather than approving each action.
| Tier | What the System Does | Where the Human Sits |
| Assistance | Analyses data and presents a recommendation | Decides whether and how to act on it |
| Execution under supervision | Carries out a predefined action within bounded conditions | Approves the specific instance in advance |
| Autonomy | Senses, decides, and executes independently | Sets policy boundaries upfront; reviews outcomes after the fact |
Why the Distinction Matters More Than the Label
An operator or vendor describing a capability as an ‘AI agent’ is telling you almost nothing about where it sits on this spectrum, which is precisely why the label alone shouldn’t be taken as evidence of autonomy. A system that drafts a suggested customer response for an agent to review and send is doing something valuable, but it’s assistance, not autonomous execution, even if it’s marketed with agentic language. The practical question worth asking of any specific claim is which of the three tiers it actually occupies for the specific task in question, since the same vendor’s platform may span multiple tiers depending on the use case, the environment maturity, and how much the operator has configured it to act independently versus flag for review.
How This Plays Out Differently Across Network, Back-Office, and Customer-Facing Domains
The maturity curve toward higher autonomy isn’t uniform across a telecom operator’s business.
| Domain | Typical Tier Today | Why |
| Network operations | Furthest toward genuine autonomy | Fault detection and remediation within known parameters is more deterministic and easier to bound safely |
| Back-office workflows | Often execution under supervision | Structured tasks, but still enough variability that full autonomy remains rare |
| Customer-facing | Frequently the least autonomous in practice | Reputational and regulatory cost of an autonomous error is high |
Network operations, assurance, configuration, and troubleshooting, has arguably progressed furthest toward genuine autonomy in specific, bounded domains. Back-office workflows, service activation, billing exception handling, RFP response drafting, sit in a middle zone. Customer-facing AI agents, despite being the most visible and heavily marketed category, are frequently the least autonomous in practice, since most operators keep a human genuinely in the approval loop for anything beyond routine, low-stakes interactions.
Where to Verify Autonomy Claims, Not Just Read Them
TeckNexus’s own research independently verifying AI agent deployment against named, corroborated sources — rather than press release claims — across dozens of global operators found substantial variation in how much of the industry’s ‘AI agent’ activity actually reaches genuine autonomy versus staying at assistance or supervised execution. The detailed findings, scored against the TM Forum framework with evidence strength assessed for each claim, are available in full through TeckNexus’s Intelligence Platform. The takeaway relevant to any buyer or industry observer evaluating a specific vendor or operator claim is the same regardless: ask which tier the specific capability actually occupies, ask what evidence supports that claim beyond the vendor’s own description, and treat the word ‘agent’ as the start of a question rather than an answer.
A Simple Internal Checklist for Evaluating Any Autonomy Claim
For an operator or enterprise fielding a vendor pitch that uses agentic or autonomous language, a short internal checklist does most of the work of separating substance from marketing:
- Does the vendor specify which TM Forum level, or an equivalent structured framework, the claimed capability actually maps to, rather than using ‘autonomous’ as an unqualified adjective?
- Is there named, checkable evidence behind the claim, a specific deployment, a specific outcome, rather than an aggregate or anonymised statistic?
- Does the vendor’s own description distinguish between what the system does today in production versus what it’s designed to eventually do on a future roadmap?
A vendor that answers all three specifically and confidently is offering a materially more credible claim than one that answers only in general terms, and building this checklist into procurement conversations costs little while meaningfully improving the quality of vendor comparison.
Why Overstated Autonomy Claims Are So Common
Understanding why the gap between claimed and actual autonomy tends to be wide is useful context for evaluating any specific claim fairly. Marketing incentives reward the more impressive-sounding claim, and ‘autonomous’ carries more weight in a press release than ‘assists a human reviewer,’ even when the latter is the more accurate description of what’s actually shipping. Genuine autonomy is also simply harder to achieve safely than assistance, which means the honest gap between a vendor’s product roadmap and its current production capability is often wider than marketing timelines suggest, not because of dishonesty but because the industry as a whole is still working through how to validate higher autonomy levels safely enough to deploy them widely. Neither of these explanations is a reason to distrust every claim by default, but they’re a reason to verify specifically rather than accept the label at face value.
See the full operator-by-operator autonomy scoring in TeckNexus’s AI Agent Intelligence tracker — https://tecknexus.com/intelligence/ai-agent-for-telecom-operators/
