By Nigel Arkwright, CIO – ATM Group
I write this in early 2026. By the time you read it the technology will have moved on and some of what I describe may already feel dated. That is the nature of this space. But the human truths underneath the technology? Those won’t have changed.
“Working with AI and creating Virtual Voice Agents is easy,” a colleague told me recently. “Just hook up your agentic chatbot to a Speech to Text engine and job done.”
That comment told me everything about how well most people understand human conversation.
It’s like quantum physics. Everyone has heard of Schrödinger’s cat and sort of knows what it means but when you press them on the detail, they go glassy eyed and say “you know… it’s quantum.”
Human conversation is the same. We do it effortlessly, without thinking, thousands of times a day. And that ease completely disguises how extraordinarily complex it actually is.
The journey
As a business we embarked on a project to build a Virtual Sales Agent for one of our largest clients with no integration available from their technology team. We assembled our best technical minds, configured the third-party platform the client had chosen, and attempted to build our first test agent.
The lessons came thick and fast.
Lesson 1 – Colloquial language will break your agent
It is straightforward to build an agent that looks good in testing. It is genuinely hard to build one that performs consistently when the inputs are a human being speaking naturally.
Here is a simple example. In a text chat someone might type “No Problem” clear, unambiguous, there is no problem. But spoken aloud, the same person might say “Yes, No Problem” and suddenly the agent is confused. Is there a problem or isn’t there?
Multiply that ambiguity across thousands of conversations and you understand very quickly why natural language processing is still one of the hardest problems in AI.
Lesson 2 – The thinking pause will frustrate your customers
The more you say to an agent, the longer it needs to process a response. That pause – even half a second longer than expected – is genuinely frustrating for people.
We heard countless calls where customers were repeating their input while waiting for the agent to reply. And here is the cruel irony: the better quality the voice, the longer the processing time. The more human it sounds, the more the pause feels wrong.
You will not spot this in testing. Testers know they are talking to a virtual agent. They are patient. They know what is coming. Real customers expecting a human response are far less forgiving.
Lesson 3 – Without CRM integration your agent is flying blind
If the agent cannot verify who it is talking to, it will struggle with identity capture and names and email addresses are surprisingly difficult for AI to parse accurately.
These are specific, low-frequency combinations. They are not common words the model has seen thousands of times. An agent that mishears a name or mis records an email address does not just create a data problem. It creates a trust problem. And in a sales process, a trust problem is a lost sale.
Lesson 4 – Listen to the calls. All of them.
Third-party transcript and sentiment tools are genuinely useful. But do not let them become a substitute for actually listening.
We listened to as many calls as we could – especially the ones where the customer asked to speak to a human. And what we found changed how we built everything that followed.
In a five or six step process, the customer will sigh approximately two stages before they ask for a human.
You cannot hear that sigh in a transcript. The sentiment tool will not flag it. But it is there a small, audible signal that the customer has already decided the conversation is not going the way they need it to go.
Listen to the calls. Trust your ears more than your analytics platform.
What actually works
This is not an AI-bashing piece. With the right toolkits, the support of our technology partners and – critically – a realistic client, we launched a Virtual Agent assisted sales process within three months.
That word realistic matters more than any technology decision we made.
Here is what we learned about building something that actually works:
Keep it quick and to the point. Function over friendliness. People will know it is a virtual agent whatever you do. The conversational warmth you spend weeks optimising for will not make them forget they are talking to a machine. A fast, accurate, frictionless process will.
Listen to everything. Do not trust transcripts. People process data differently through their eyes and their ears. What reads as resolved in a transcript can sound unresolved in the call. Your ears will catch what your data misses.
Make your testing blind. People who know what they are testing behave differently to real customers. If your testers know they are talking to a virtual agent, they will show patience your real customers will not. Build blind testing into your process from day one.
In early life, make sure you can listen in and take over. Virtual agents can be difficult to interrupt. Having a human available to step in during the first weeks of live operation is not a failure – it is good engineering.
Use your best agents to build the conversations. They are the people who understand how trust is built and how customers are guided through a decision. That knowledge is irreplaceable. Train them in the technology and let them shape how the agent behaves.
Reframe what success looks like. If a customer hangs up as soon as they hear the virtual agent, that is not necessarily a failure. Realistically, fifty percent of those people would have hung up on a human agent too. The number of people who refuse to engage with virtual agents is diminishing. But it will never reach zero – and that is fine.
The thing I would tell anyone starting this journey
You will eventually find a process that works for your customers. But you will not get there by chasing the most sophisticated technology.
The trap is trading customer experience for a pure technological solution – optimising for what the AI can do rather than what the customer actually needs.
The answer, especially if you are selling, is to blend people and technology deliberately. Customers are sensitive to how they are treated. They will tolerate an imperfect virtual agent. They will not tolerate feeling like they do not matter.
Build the agent around the customer, not around the capability.
That is the lesson six months of building this thing taught us.
Nigel Arkwright is CIO at ATM Group. ATM Group builds AI-enabled customer operations for brands across telecoms, financial services, retail and utilities.
