Telecom AI Governance: 72% Trust It, 14% Can Prove It

Telecom AI Governance: Closing the Gap Between Trust and Proof

Telecom AI governance has a credibility problem that the industry has now measured. Research from TM Forum and IBM's Institute for Business Value found that 72 percent of communications service providers consider their AI capabilities trustworthy, yet only 14 percent could provide evidence to support that view. In October, TM Forum and Accenture launched a multi-year programme to create a trustworthy AI governance certification for operators, aimed squarely at autonomous and agentic systems that execute decisions rather than suggest them. The timing is not accidental. Operators are putting AI into live voice calls, fault handling, and network operations, and a belief that the AI is trustworthy is no longer an adequate answer when something goes wrong.

Where Operators Use AI and What Evidence Looks Like

Each use carries a different risk, and each needs a different kind of proof that it is working as intended.

Why the Proof Gap Matters Now

For years, most operator AI was advisory. A model predicted a fault or flagged a churn risk, and a person decided what to do. Governance could be light because a human sat between the model and the consequence. That arrangement is changing quickly. Virgin Media O2 has launched an AI voice agent for routine broadband fault calls, with human advisers kept for complex cases. Vodafone and Ericsson have trialled AI voice services running inside the mobile network itself, including live translation between several languages. Network automation is pushing toward systems that detect, diagnose, and fix problems without waiting for an engineer.

Each of those steps removes a human checkpoint, and each raises the question regulators, enterprise customers, and boards will ask after an incident: how did you know it was safe to let the system act?

What Telecom AI Governance Must Cover

The new certification programme frames governance around three areas. Governance maturity covers how oversight is structured and who is accountable. Workforce readiness covers whether the people working alongside autonomous systems understand them and can intervene. Operational accountability covers verifiable records of what AI decided, which data it used, and what happened as a result. That last area is where the 14 percent figure bites hardest, because most operators have policies on paper and very little traceable evidence in practice.

From Recommend to Act

The industry already has a useful vocabulary here. Autonomous network frameworks describe levels of autonomy, from fully manual operation to systems that run themselves across domains. The governance burden rises steeply at each level. A system that recommends a configuration change needs accuracy monitoring. A system that applies the change needs limits on what it can touch, approval rules for higher-impact actions, automatic rollback, and a record a reviewer can follow afterwards.

In our view the most common mistake is treating autonomy as a property of the model rather than of the deployment. The same model can be safely autonomous for low-impact actions in one domain and dangerous if given the same freedom in another. Autonomy should be granted per action type, with explicit boundaries, and reviewed as evidence accumulates.

What Counts as Evidence

Evidence is what lets someone outside the AI team check a claim of trustworthiness. For operators, the useful forms are fairly concrete:

• Decision logs. What the system decided, when, on which inputs, with which model version.
• Performance against thresholds. Agreed accuracy, false positive, and service quality measures tracked over time.
• Intervention records. How often humans overrode the system, and why.
• Rollback history. Every automated change that was reversed and what triggered it.
• Testing results. Pre-deployment and periodic tests, including edge cases and adversarial inputs.
• Customer outcomes. Complaint rates, escalations, and appeal results for customer-facing AI.
• Data lineage. Which data sources fed the system and how their quality was checked.

Seven Steps to Close the Gap

1. Inventory every AI system that acts. List which systems change networks, accounts, or customer outcomes without a human step.
2. Assign an accountable owner to each. A named person who answers for its behaviour, not a committee.
3. Set autonomy boundaries per action type. Define what each system may do alone, what needs approval, and what is off limits.
4. Instrument decisions from day one. Build logging into deployment rather than retrofitting it after an incident.
5. Test fallbacks, not just models. Confirm that rollback and human takeover actually work under pressure.
6. Review evidence on a schedule. Monthly reviews of performance, overrides, and customer outcomes keep claims honest.
7. Prepare for external assessment. Assume a regulator, auditor, or enterprise customer will ask to see the evidence.

Pros and Cons of Autonomous AI for Operators

• Pro: faster fault resolution. Systems that act immediately can restore service before customers notice.
• Pro: energy savings. Dynamic capacity management cuts power use in quiet periods.
• Pro: scalable customer service. Routine queries get resolved at any hour, freeing advisers for complex cases.
• Con: errors spread at machine speed. A wrong automated fix can propagate across a network quickly.
• Con: harder accountability. Without records, it is difficult to explain or defend what a system did.
• Con: regulatory exposure. AI in critical digital infrastructure and customer treatment attracts growing scrutiny.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how telecom AI governance plays out in practice rather than presented as verified case studies.

An operator's energy-saving system switches off capacity overnight across a region. A local event runs late, demand spikes, and service degrades for an hour before thresholds trigger a reversal. Because the system logged every decision and the thresholds were documented, the review takes a day and the fix is a simple calendar input. Without the records, it would have been an argument.

An AI service agent handling broadband faults performs well in testing. A review of real conversations shows it tells some customers an engineer visit is free when their contract says otherwise. Conversation sampling caught the pattern within weeks rather than months, which is the evidence process working as intended.

A large enterprise customer asks its operator for proof that AI used in its managed network is governed. The operator can show an inventory, owners, autonomy boundaries, and six months of decision logs. The contract renews. A competitor with only a policy document does not get the same benefit of the doubt.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

Open Models, Vendors, and the Accountability Chain

Many operators are building on open models, which offer more control and customisation. That flexibility comes with responsibility: an operator that fine-tunes and deploys a model owns its evaluation in a way it does not when buying a finished product. Vendor systems raise the opposite problem. A capable black box with no access to decision logs leaves the operator accountable for outcomes it cannot explain. Contracts should guarantee access to the evidence, not just the service.

The same logic applies whether an operator is consuming AI or selling it, a distinction we explored in telco AI clouds. For the general controls that should sit around any agent before it acts in production, see agentic AI governance.

Why Talkory Wins

Customer-facing AI is where operators most often lack evidence, because answers vary in ways that are hard to see until a complaint arrives. Talkory runs the same question across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 in one pass. A service team can take its most common customer questions, about contracts, outages, roaming, or billing, and see where six independent models give consistent answers and where they diverge. Divergence marks the questions that need tightly controlled content or human handling, and the comparison itself becomes part of the testing record that governance reviews increasingly expect.

Final Verdict

Telecom AI governance is moving from policy to proof. The gap between operators who believe their AI is trustworthy and those who can demonstrate it is large, and it matters more as systems start acting in networks and customer channels on their own. Close it by inventorying systems that act, naming owners, setting autonomy boundaries per action, logging decisions from day one, testing fallbacks, and reviewing evidence regularly. Certification programmes will help set the bar. The operators who already have the records will find clearing it straightforward.