Telus Is Using AI to 'Americanize' Call Center Accents. That's a Problem.

Canadian telecom giant Telus is deploying AI that rewrites the voices of its overseas call center workers — agents in India, the Philippines, and elsewhere — into something closer to a Midwestern American accent, in real time. The official line is operational efficiency. The harder question, the one nobody at Telus seems eager to answer, is what exactly we’re erasing in the name of a smoother customer experience.

What’s actually happening

The technology isn’t new. Startups like Sanas and Krisp have been selling “real-time accent neutralization” to call centers for years. An agent in Bangalore speaks in their natural English. A customer in Ohio hears something that sounds, for all intents and purposes, like a fellow American. The conversion happens with minimal latency, mid-sentence.

Telus is now pushing this stack across its global support operations. The company’s pitch is clean: shorter calls, higher first-call resolution rates, fewer racist tirades aimed at agents. Sanas claims its tech cuts average handle time by more than 20%. The business case writes itself.

The efficiency argument

For a company, the appeal is obvious. You hire in low-wage markets and deliver a customer experience that feels local. Costs stay flat. CSAT goes up. From a P&L standpoint, it’s almost too good to argue with.

There’s even a worker-protection framing. Non-native English speakers in call centers absorb a steady stream of abuse — “go back to your country” is a Tuesday afternoon, not a notable incident. If the accent disappears, so does some of the abuse. That’s not nothing.

The direction matters

But here’s the thing this technology never does: it never works the other way. The Indian agent’s voice gets converted for the American customer. The American customer’s voice never gets converted for the Indian agent. This isn’t a neutral tool. It’s a one-way mirror, and the asymmetry is the entire point.

Critics have started calling this digital whitewashing, and the label fits. Voice and accent are among the first things that signal who a person is. Erasing them — quietly, by default, because someone on the other end finds them inconvenient — doesn’t fix prejudice. It just sands the edges off so the prejudice runs more smoothly.

The “consent” question is also doing a lot of unspoken work here. How many call center workers, with their job on the line, are in a position to say no when management announces an AI will be modifying their voice on every call? Consent under economic duress isn’t really consent.

What about the customer?

There’s another piece nobody’s resolved: disclosure. When you call Telus support, you’re told the call may be recorded for quality assurance. You are not told that an AI is modifying the voice on the other end of the line in real time. That’s a meaningful omission.

Regulators are starting to notice. The EU’s AI Act and Canada’s proposed AIDA both lean on a principle that people have a right to know when they’re interacting with — or being mediated by — an AI system. Real-time voice modification of a human agent walks straight through that gray zone. It’s not a deepfake, exactly. But it’s not nothing, either.

The takeaway

What’s uncomfortable about this story is how much “efficiency” is being asked to justify. Behind the tidy KPIs of cost-per-call and customer satisfaction, someone’s voice is being deleted and someone else’s bias is being insulated by software. Technology likes to call itself neutral. The direction it runs in rarely is.

Two questions worth sitting with: When you call support, do you have a right to hear the actual person on the other end of the line? And does that person have a right to be heard in their own voice? How we answer those will shape the next decade of AI at work — and who pays the price for the convenience.