Your AI Agent is Failing Before It Ever Speaks
Broken telephony infrastructure is undermining AI customer service agents — and the calls most at risk are often the ones that matter most.
A test call went to an adverse event reporting line in Germany, run for a global pharmaceutical company. It rang for 52 seconds, then said the call couldn’t be connected. An outsourcing partner had been contracted to answer it. Nobody answered in the end, and no internal system raised a flag.
In another instance, a patient safety line wasn’t working for some time in a mature European market, and the business only found out about it from an external test.
These are just some of the standout examples I’ve seen over the years of underlying telephony infrastructure — carrier routing, in-country provisioning, and the like — that is working poorly or is outright broken. Put an autonomous AI agent on top of that, and you’re adding further complexity and challenges.
Customers dial when the stakes are highest
ContactBabel’s 2026 US Customer Experience Decision-Maker’s Guide surveyed 1,000 US consumers and 190 organisations and found that telephony is dominant wherever urgency, emotion and complexity run high, across all age groups and income brackets. Contact Center Pipeline puts it simply: People choose the phone for the hard problems, even when they expect to wait.
So, in essence, the calls most exposed to infrastructure faults carry the most value for the business and the greatest urgency for the caller. Things like a mortgage application, an incorrect online shopping order, or a safety report. On the caller’s side, it could be a health emergency, a flight change, or a stolen credit card.
The budget lands at the top of the stack
We recently audited one financial services client’s published numbers and found some interesting results. Of more than 3,000 numbers on their books, over 280 rang out or returned an invalid number message. Every one of them was printed where a customer could find it: on a website, in a statement, on a credit card. That’s the baseline infrastructure most AI agent programmes sit on.
AssemblyAI’s 2026 Voice Agent Report shows that 75% of voice agent builders were encountering reliability barriers related to accuracy, integration, and cost. Confidence in the AI model itself and confidence in its delivery are two separate measures, and the visibility gap between them is where customers can end up falling through the cracks and where things break down.
Machines are less forgiving than people
Here are three interesting faults we’ve seen lately, all of which a customer would simply report as “bad audio.” In Spain, mean opinion scores — also known as audio quality scores — were optimal on the carrier leg into the IVR, then fell once the call was answered. In Colombia, they dropped low enough that audio matching failed. In Australia, we saw prompt recordings peak beyond acceptable levels, a fault in the client’s audio rather than the network.
A human absorbs all of this without comment. You and I can have a conversation on a crackly line and understand each other well enough to make sense of what’s being said. An AI agent, however, will take the degraded transcript from that crackly line and misread the intent. It then acts on that misread intent, proceeding down the wrong path — leading to a frustrated caller and a poor customer satisfaction score.
Accents sit on top of all of this. Travel 20 miles down the road in parts of the US, and both accent and transcription accuracy change. An AI agent signed off at headquarters in Dallas hasn’t really been tested until it’s taken a call from Boston. If you don’t have a good-quality line to layer your AI agent on top of, the quality will simply fail.
Trust runs through the whole stack
Governance starts with ownership. The hard part, understandably, is trusting an AI agent to act on your behalf, especially when it hasn’t been tested continuously in a live environment. For large domestic or multinational enterprises, there is a lot at stake, both financially and reputationally.
In February 2024, a tribunal ruled in Moffatt v. Air Canada that the airline was liable for advice its chatbot gave a bereaved customer: book now, and claim the bereavement fare back later. Company policy said otherwise, and the airline argued that the bot was a separate entity responsible for itself. The tribunal saw it differently.
The common-sense response would have been to accept that the bot went outside policy and make the customer whole — a much cheaper solution. Contesting it cost far more in the end than the fare would have, and it has left a precedent for everyone else with a chatbot. It also caused reputational damage to Air Canada among many existing and potential customers.
The lesson here is to put the agent in the best position you can: good audio, low latency. Ask it what it does with the intent it receives — does it solve the problem without hallucinating or breaking out of its guardrails? Where appropriate, give it a workflow with no capacity to deviate, for example, when cancelling a credit card.
A human agent reads the person in front of them and adjusts naturally. A bot needs that range built in on purpose to deal with the many personas of callers who are frustrated, elderly, in a rush, or all three.
Nobody sounds like your test room
Recently at Genesys Xperience, most of my conversations weren’t about building and deploying bots, but about ensuring they have the data, context and ability to solve customer problems out in the real world, where conditions are always changing.
How can they be tested for that? The answer is to test the telephony lines through to the bots from the outside in — the way the call actually arrives in the real world: a mobile in the customer’s hand, in their country, on the move on a busy street, in their accent, dialling your number.
