Mobile AI: the next AI assistant lives in your SIM

Why is voice making a comeback?

Because AI finally made talking faster than typing, and people noticed.

For a decade, voice looked like it was losing. Messaging took over personal conversation, and call revenue fell everywhere. MTN reported its South African voice revenue down 10.2% in the half to June 2026, while its CEO said he believes voice is now “bottoming out”.

Then AI changed what voice is for. OpenAI says more than 150 million people a week talk to ChatGPT using voice and dictation, and in July 2026 it launched GPT-Live, a voice model that listens and speaks at the same time, with the small “mhmm” and “yeah” of a real conversation. Speaking is the most natural interface people have. Once the machine on the other end can keep up, voice stops being the old channel and becomes the fastest one.

But almost all of that happens inside an app. You open it, you talk to it, you close it. The phone call – the place where business actually happens – has been left out. That is the gap mobile AI closes.

Why does AI in the SIM beat AI in an app?

Because the network already sits inside every call, on every phone, for both people on the line.

An app-based assistant has to be downloaded, opened and fed. It never hears your calls unless you route them through it. Network-based AI works the other way round: the operator places the AI in the voice core, where every call already passes, and you switch it on from the phone’s own dialer.

That placement gives the network four advantages no app can copy:

The last row cuts both ways, and it matters. An operator-run assistant concentrates a great deal of trust in one company. The operators going first seem to understand that: Deutsche Telekom’s assistant processes nothing until the customer activates it, and tells the other person on the call when it is switched on.

Which operators are already building AI into calls?

The largest operators in Europe, the US and Asia, all in 2026, and the industry body has written the playbook.

Look at what is not on that list: a new app. Every one of these runs in the operator’s network and appears inside the ordinary phone call. That is the bet the industry is placing.

What does mobile AI mean for a business?

Every mobile call becomes a record: summarised, searchable and ready for the CRM. That is conversational intelligence.

Consumers will notice translation and scam screening first. Businesses will feel something bigger. Most customer conversations in a small business never touch a desk phone or a contact centre. They happen on a salesperson’s cellphone in the car, a technician’s phone on site, an owner’s phone at 7pm. None of it is recorded, summarised or logged anywhere.

Conversational intelligence has been the contact centre’s privilege: transcripts, summaries, action items and compliance records for every call. Companies like Dubber already deliver it through operators rather than apps, capturing calls “across telephone networks, unified communications platforms and mobile services” and adding AI summaries in 2026. Put that capability in the mobile network and the contact-centre toolkit reaches every phone a business owns.

This is also where mobile AI and an AI voice agent differ. An AI voice agent answers calls for you. Mobile AI helps you on the calls you take yourself. A business will want both, and the providers that connect them – agent at the front door, intelligence on every mobile call behind it – will own the next decade of business telephony.

Why could South Africa leapfrog with mobile AI?

Because South African business already runs on the mobile network, and the network is being rebuilt for exactly this.

Three things make South Africa a stronger candidate than most markets:

• The phone system is the mobile. ICASA counts 117.3 million mobile subscriptions and 83 million smartphone subscriptions. For many small businesses, the owner’s cellphone is the switchboard. AI in the network reaches them without a single new system.
• The network is being rebuilt. MTN plans to retire 3G by the end of 2026, moving voice onto 4G and 5G. Network AI calling is built for those modern voice cores, which is why the GSMA framed it as part of the “mobile AI era”.
• Eleven official languages. Live translation is a nice-to-have in Germany. In a country where a customer, a supplier and a technician may each prefer a different language, it is a business tool.

The operators have the incentive too. Voice revenue has been shrinking for years. AI is the first thing in a decade that makes a phone call worth more than it was the year before, and South African networks already use machine learning at network level: Vodacom says it deploys “real-time traffic monitoring and machine learning tools to detect and block suspicious call patterns”. The step from blocking bad calls to improving good ones is shorter than it looks.

What has to go right for mobile AI to work?

Consent, language accuracy and trust. Get those wrong and network AI becomes network surveillance.

Accuracy in local languages. Translation and transcription are only as good as the speech models underneath. Independent benchmarks of code-switched African speech still show high error rates, even from the best-known global models. Operators that train for South African accents and languages will have a real advantage over those that switch on a global default.

Trust in the operator. An app vendor that misuses your data loses a customer. An operator that does it loses a country’s trust. That raises the bar, and it should. The operators that publish exactly what their AI hears, stores and deletes will earn the right to put it in every call.

Keep reading