The Hidden Cost of AI Voice Assistants

The Hidden Cost of AI Voice Assistants: Is Traditional IVR Still the Smarter Investment?

For the better part of two decades, I’ve worked in telephone audio. I’ve recorded greetings, directed voice talent, prepared announcements and watched businesses move from one telephone platform to another. And lately, as more companies introduce AI assistants to answer their phones, I keep coming back to a question: haven’t we been here before?

Those of us who worked in this industry in the early 2000s remember the excitement around speech recognition. Instead of pressing 1 for sales or 2 for service, callers would be able to say what they needed. It sounded like a step forward. Sometimes it was. But plenty of customers ended up repeating themselves, shouting “representative,” or trying to find a way around the system.

Now we’re hearing many of the same promises, this time with technology that is undeniably more capable. Today’s AI can understand far more natural language, respond in complete sentences and, when properly connected to business systems, carry out useful tasks. That’s real progress. But it doesn’t settle the most important question: does it make the call easier for the customer, and does it make financial sense for the business?

I don’t think we should assume the answer is yes simply because the technology is new.

We’ve seen the speech recognition pitch before

The first generation of speech-driven telephone menus had a fairly simple ambition. Let customers speak their request and route them accordingly. In practice, those systems often depended on a limited vocabulary and tightly defined phrases. Background noise, accents and unexpected wording could throw them off. A caller who said “billing” might get through, while someone who said “there’s a problem with my invoice” could end up in a loop.

That wasn’t just a recognition problem. It was also a design problem. Callers weren’t always given a clear escape route. Menus asked for too much information. People were made to feel as though they had to learn the system’s language rather than the other way around.

Modern conversational AI is not the same technology. It can interpret many ways of expressing the same request. It can keep track of context and ask follow-up questions. In some cases it can authenticate a customer, check a record or make a change without involving an employee. Dismissing those improvements would be a mistake.

But we should also be careful not to confuse a more natural conversation with a better result. If I call to report a lost card, I don’t necessarily want a conversation. I want the card secured. If I know which department I need, a clear menu option may be quicker than explaining my reason for calling to a machine. If my issue is unusual, on the other hand, a capable assistant might save me from navigating five layers of menus.

The right approach depends on the call. That’s less exciting than saying AI will replace IVR, but it’s a better starting point for a business decision.

What do callers actually prefer?

One of the problems with this discussion is that three different things are often treated as interchangeable: a human customer service representative, a human-recorded IVR, and an AI assistant. They aren’t.

A human-recorded greeting is still automation. It can sound familiar, reassuring and professional, but it cannot solve an account problem on its own. An AI assistant can speak in a very natural voice, yet still fail to complete a request. And a live representative may be the only practical option for a complicated or sensitive issue.

Independent consumer research suggests people are open to automation, but don’t want to lose access to human help. In a 2026 study of 503 adults in Canada and the United States, Metrigy reported that 85% preferred interacting with humans rather than AI agents for customer service. Gartner’s 2026 survey of 3,566 customers found that 87% considered access to a human agent essential when companies use generative AI. At the same time, half of Gartner’s respondents said generative AI made customer service interactions easier.

Those findings aren’t contradictory. A customer can appreciate a fast automated solution and still want a person available when something goes wrong. What the surveys do not establish is that 85% of callers prefer a recorded human voice to a synthetic one. That’s a separate question, and we should be precise about it.

The practical lesson is that callers want control. They want to understand their options, know what the system can do, and get help without starting over. An automated experience can be excellent when it handles a simple task quickly. It can also be infuriating when it insists on continuing a conversation it isn’t equipped to finish.

A Voice That Sounds Intelligent Isn’t the Same as a System That Helps

A lot of attention goes to how human an AI voice sounds. I understand why. Voice is our business, and the improvements in synthetic speech have been remarkable. But the voice is only the surface of the experience.

Consider a caller asking to change an appointment. The assistant understands the request perfectly, asks for a preferred date and speaks naturally throughout. But if it can’t access the booking system, it still has to transfer the call. The customer may then have to repeat everything. The system sounded capable; it just wasn’t capable of finishing the job.

Now imagine the same assistant is securely connected to the booking calendar, checks availability and confirms the change. That’s a genuine improvement. It saves the caller time and can free an employee for more demanding work.

That distinction should drive purchasing decisions. What can the system actually complete? What happens when it can’t? And does it pass the information it gathered to the next person or make the caller begin again?

A field experiment published in Production and Operations Management in 2023 offers a useful reminder that results can be mixed. The researchers found that introducing voice-based AI initially increased automated service duration and demand for human support, while also reducing customer complaints over time. Recognition failures increased complaints and requests for human help. In other words, the technology produced benefits, but not without friction.

We should expect both sides of that story. AI can make some interactions easier. It can also create new failure points. The job of a contact-centre team is to determine which outcome is happening in its own environment, not rely on a demonstration or a promise.

The cost question is bigger than the monthly subscription

Cost reduction is one of the main reasons companies look at conversational AI. For a large contact centre, even a modest reduction in calls requiring employees could be worth a great deal. Labour, scheduling, training and turnover are expensive. If an assistant resolves thousands of routine issues accurately, there may be a strong business case.

But the comparison needs to be fair. Replacing a call that previously required five minutes with an employee is not the same as replacing a recorded menu that already routed the caller in 15 seconds.

Traditional IVR has costs. The telephone platform needs to be maintained, call flows need to be designed, recordings need to be updated and integrations need support. But once a professionally recorded prompt is delivered, it can be played again and again without generating a new voice response for every caller.

Live conversational AI may involve speech recognition, language-model processing, text-to-speech generation and connections to business applications. Depending on the architecture and contract, those services can create ongoing usage charges. More calls, longer conversations and more complicated tasks may mean higher bills.

Then there’s implementation. A genuinely useful assistant often needs access to appointment systems, customer records, payment tools, account information or authentication processes. That work requires development, security review, testing and maintenance. A demonstration that answers general questions is one thing. A production system that can safely change a customer’s banking details is another.

Businesses also need people to monitor the system. Questions change, policies change, integrations break and unusual requests appear. Someone has to review failed interactions, update workflows and test changes before they reach customers. Those costs may be carried by an internal team rather than shown on the AI invoice, but they’re still real.

And we can’t ignore the unsuccessful calls. If an AI assistant spends three minutes with a customer before transferring them to an employee, the company may have paid for both parts of the interaction. Perhaps the assistant collected information that saved the employee time. Perhaps it didn’t. Without measuring the full journey, there’s no way to know.

This is why the most useful financial metric is cost per successfully resolved issue, not cost per automated minute or the percentage of calls that never reach a person. A call that stays inside automation is not necessarily a call that has been resolved.

In January 2026, Gartner forecast that generative AI customer-service resolutions could cost more than US$3 each by 2030, potentially exceeding some offshore human-agent costs. That’s a forecast, not a universal price tag, but it should make buyers cautious about assuming AI processing will always become cheaper.

The numbers can work very well for the right use case. They can also look very different once integration, ongoing supervision, failed calls and repeat contacts are included.

Deprecation: The Cost Nobody Wants to Discuss After Launch

There’s another issue that deserves far more attention in enterprise telephone planning: deprecation.

In plain language, deprecation means a product, model, feature or service is being phased out. It may continue operating for a while. Support may be reduced. Eventually, customers may be required to migrate or stop using it altogether.

This is normal in technology. Traditional PBX equipment, operating systems and software also reach end of life. AI isn’t unique in that respect. What makes service-dependent AI worth examining is the number of moving parts a business may rely on, and how much control it has over their lifecycles.

A conversational assistant may depend on a speech recognition service, a language model, a voice generation service, authentication tools and several integrations. Each component may have its own updates, pricing changes and retirement schedule. Replacing one part can mean retesting the whole customer experience.

AWS provides concrete examples. Amazon Lex V1, an earlier conversational bot service, reached the end of support and access in September 2025. Organizations using it needed to plan for Lex V2 or another approach. Lex V2 wasn’t simply a renamed version of the same service; differences in architecture and interfaces could require migration work.

In May 2026, AWS discontinued Amazon Connect Voice ID, a voice biometric authentication feature. This wasn’t the synthetic voice that speaks to callers, and it wasn’t the retirement of Amazon Connect. It was a specialized authentication service. But its closure still matters because businesses relying on it had to find another way to perform that function. AWS also noted limitations around transferring existing voice enrolments to a suggested alternative.

Those examples do not prove that AI voice technology is failing. Providers retire services for all kinds of reasons, including consolidation and newer architectures. They do, however, demonstrate why a business cannot evaluate a cloud voice service solely on what it costs to deploy today.

Think about the work involved when a bank or insurer changes a customer-facing system. There may be new integrations, privacy assessments, security testing, updated documentation, staff training and approval processes. Even if the replacement is better, the organization is paying to preserve a capability it had already built.

Over five or ten years, those costs can matter. A procurement team should ask how long the service will be supported, how much notice it will receive before retirement, whether it can export its data and voice assets, and who pays for the migration. It should also ask what happens if the service is unavailable on a busy Monday morning.

That last question is especially important. A customer should still be able to reach the business if an AI component fails. Good fallback routing is not an optional extra. It is part of a responsible enterprise design.

The value of a recording you can actually keep

One reason professionally recorded IVR has lasted is its simplicity. A recorded greeting is an audio file. Once approved, it can be stored, backed up and used across compatible systems. If the company changes telephone platforms, the file can often be converted into the required format and carried forward.

That doesn’t mean the recording lasts forever without attention. Operating hours change. Department names change. Promotions expire. A voice recording is only useful while its content is accurate. But the underlying audio doesn’t need a cloud model to remain available.

To be fair, a synthetic recording exported as a WAV file can offer the same technical portability. The important distinction isn’t always human versus AI voice. It’s a fixed, owned audio asset versus a live service that must keep running, processing and billing every time someone calls.

Where human voice production continues to matter is in performance and consistency. A professional voice talent can be directed to handle names, terminology, sensitive announcements, pacing and tone. The company can approve the delivery and know exactly what the caller will hear. For established brands, that degree of control is valuable.

A familiar voice can also become part of a company’s identity. It carries across greetings, transfers, after-hours messages and service announcements. That continuity shouldn’t be dismissed simply because a new platform can generate speech on demand.

The right question isn’t whether a synthetic voice can sound pleasant. Many can. It’s whether a business gains enough from generating speech dynamically to justify the added technology, especially when most of its messages are known in advance.

Simplicity is not a step backward

There’s a tendency in technology discussions to treat fewer features as a disadvantage. In telephone systems, that can be exactly the wrong way to think.

A caller contacting a veterinary clinic may need an appointment, a prescription refill or emergency instructions. A hotel caller may want reservations or the front desk. A banking customer may need to report a lost card or reach a specific department. Those are often clear, repeatable paths.

A short, well-written menu can handle them quickly. A professional recording can deliver the instructions clearly. And a properly configured telephone platform can transfer the call without adding a new conversation.

Of course, some menus are terrible. They are too long, bury important options, repeat legal language and make it difficult to reach a person. But those are design failures, not proof that every menu needs replacing with AI. Sometimes the most effective upgrade is to rewrite the prompts, reduce the number of options and fix the routing.

Before a business invests in a conversational assistant, it should audit what callers actually do. Which options are selected most often? Where do people abandon calls? Which calls are misrouted? What do employees hear from frustrated customers? What tasks genuinely require access to data or complex decisions?

If the problem is a confusing menu, fix the menu. If the problem is that thousands of employees spend time rescheduling straightforward appointments, consider transaction-capable automation. Different problems call for different investments.

Where AI Earns Its Place

None of this is an argument to reject conversational AI. There are real opportunities, particularly when a system can finish work that previously required an employee.

Appointment changes are one example. So are order status inquiries, routine account questions, certain service requests and after-hours support. When an assistant is securely connected to the right systems and has clear limits on what it can do, it may give customers faster service while reducing repetitive work for staff.

AI may also help behind the scenes. It can assist with identifying caller intent, preparing a summary for the employee who takes over or organizing information without making the customer listen to a lengthy conversational exchange.

I suspect the most practical approach for many enterprises will be a hybrid. Keep professionally recorded greetings and direct menu options where they work. Use conversational AI for tasks where it can deliver a measurable benefit. Give callers a clear way to reach a person when necessary.

That approach may not make for the flashiest product demonstration, but it respects the reality of enterprise operations. A large organization cannot simply replace its telephone experience overnight and hope customers adapt. Reliability, security and consistency matter just as much as innovation.

How to tell whether the investment worked

If I were evaluating an AI voice project, I would want more than a report showing how many calls the system handled.

I’d want to know how many customers actually resolved their issue on the first contact. How many called back? How long did they spend in automation before reaching a person? Did the system pass along the information it collected? Were customers more satisfied? And what did each successful resolution cost after accounting for technology, employees, integration and maintenance?

I’d also compare performance by type of call. An assistant might be excellent at checking an order status and poor at handling a disputed transaction. Combining those results into one impressive automation percentage can hide important problems.

Customer effort matters. So does accessibility. Callers have different accents, speech patterns, disabilities and levels of comfort with technology. Some will welcome a conversational interface. Others will want a keypad option or a person. A telephone system should not make people work harder simply because they interact with it differently.

And the measurement shouldn’t stop after launch. AI models, business policies and customer behaviour change. An assistant that performs well in testing can struggle with real-world exceptions. Ongoing review is part of the cost of operating it responsibly.

Five Questions Every Business Should Ask

Before replacing a working IVR, I would put five questions on the table.

1. What problem are we trying to solve, and how often does it occur?
2. Can the new system complete the task, or will it simply create another step before a transfer?
3. What is the full cost per successful resolution, including implementation, supervision, usage and failed interactions?
4. What happens if a model, voice service or integration changes, becomes unavailable or is retired?
5. Are customers genuinely better served, or have we simply introduced a newer way to answer the phone?

If the answers are clear and the results support the investment, AI may be the right choice. If not, a simpler solution deserves serious consideration.

Frequently Asked Questions

Where Does IVR Go From Here?

I don’t believe traditional IVR will disappear simply because conversational AI exists. I also don’t believe every company should keep the system it has out of habit. Telephone communications should evolve when there’s a reason to improve them.

Some businesses will use AI to resolve substantial volumes of customer inquiries. Others will continue relying on straightforward menus because they are effective, affordable and easy to maintain. Many will combine both approaches.

The bigger change may be in how businesses evaluate their telephone systems. For years, the discussion has focused heavily on what new technology can do. It needs to focus just as much on what customers need, what systems actually deliver and what organizations will pay to keep them working over time.

At COHM, we’ve spent decades producing professional telephone audio for businesses that need their communications to be clear, consistent and dependable. That work has taught us something worth remembering in the current rush toward AI: callers aren’t impressed by complexity for its own sake. They want to get where they’re going.

Sometimes the best solution is a capable AI assistant. Sometimes it’s a person. And sometimes it’s a clear, professionally recorded message followed by the right transfer.

The goal should never be to put more AI into a telephone system just because we can. The goal should be to make the experience better, without creating costs and dependencies that outweigh the benefit.

Innovation is valuable when it solves a problem. Simplicity is valuable when it already does.

Sources and further reading

Gartner (August 2026), “Gartner Survey Finds 87% of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent.”

Metrigy (February 2026), “85% of Consumers Prefer Interacting with Humans vs. AI Agents for Customer Service.”

Gartner (January 2026), “Gartner Predicts GenAI Cost Per Resolution for Customer Service Will Exceed Offshore Human Agent Costs by 2030.”

Wang et al. (2023), “Voice-Based AI in Call Center Customer Service: A Natural Field Experiment,” Production and Operations Management.

Amazon Web Services, Amazon Lex documentation and migration guidance.

Amazon Web Services, Amazon Connect Voice ID end-of-support documentation.