For the past two decades, press-1-2-3 Interactive Voice Response (IVR) has been the standard front door of every mid-sized and large organisation's phone system. In 2026 that landscape is shifting faster than most people expected. Zendesk's 2025 survey found that 72% of consumers say IVR menus are the most frustrating part of contacting a company, and NICE's 2026 Customer Experience report found that 67% abandon the call while working through the menu. These are not merely uncomfortable numbers — they translate directly into lost sales and a falling NPS.
The numbers that make a CFO listen
Recent reports from Fini Labs and Retell AI give a clear picture of how far apart the cost per call now is:
- Legacy IVR (Avaya, Genesys PureConnect, on-prem Cisco UCCX): an average of USD 7.50 per call once escalation to a human agent is included
- Conversational AI voice agent: USD 0.30-1.20 per call, answering in under 600 milliseconds
- Containment rate (the share of calls that never have to be passed to a person): IVR tops out at roughly 30-40%, while an AI agent reaches 60-80% on a well-scoped use case
Gartner expects that by 2028 AI will handle 40% of customer communication with no human involved at all, while the global voice AI market grows from USD 2.4 billion in 2024 to USD 47.5 billion in 2034.
The problem with IVR is not that it is slow. It is that it does not understand why the person is calling.
What an AI voice agent does that IVR cannot
In 2026, platforms such as PolyAI, Cognigy, Replicant, Five9 IVA, Genesys Cloud AI and Amazon Connect with Lex do much more than replace a menu:
- Understand natural speech, in English and in Thai, including code-switching accents
- Qualify leads and book appointments end to end
- Update CRM and ticketing in real time, rather than transferring the call and letting a person retype everything
- Escalate intelligently — handing the call to a human agent together with a summary of the conversation, which cuts AHT (average handle time) by 30-50%
- Integrate with STIR/SHAKEN so robocalls are filtered before they reach the system at all
So why hasn't everyone moved?
From the vantage point of people who have seen these projects both succeed and fail, a few things deserve care:
1. Hallucination in a high-stakes context
For banking, healthcare and insurance, one wrong answer can mean litigation. Modern platforms therefore use constrained generation (the NVIDIA NeMo guardrails, for example) to force the model to answer only within a verified scope.
2. Voice cloning fraud
AI voice in 2026 is good enough to be turned against you in "grandparent fraud" — cloning the voice of a family member to trick someone into transferring money. Voice biometrics on the receiving side has to be upgraded in step.
3. Integration debt
Many customers still run 7-10 year-old on-prem PBXs (Avaya CM, Cisco UCM) with no REST API rich enough for an AI agent to read and write CRM data in real time. Moving has to be thought through across the whole stack, not as a front-end swap.
4. Measuring the wrong thing
Containment rate on its own is misleading. It has to be read together with post-interaction CSAT, resolution rate and the complaint rate within 7 days — because an AI that "closes the call" while leaving the customer unhappy has only pushed the problem into next week.
A practical path for organisations getting started
- Start with a narrow, measurable use case — order status, appointment booking, password reset. Do not try to make the AI take every kind of call from day one.
- Run it in parallel with the existing IVR for at least 3 months, capturing baseline cost per call and CSAT on both paths.
- Audit audio and transcripts weekly in the early phase — QA should listen to 100-200 calls a week to catch hallucinations.
- Choose a vendor that offers on-prem or private deployment if the business sits in a regulated industry; several platforms can now deploy into the customer's own Azure or AWS region.
- Do not forget STIR/SHAKEN and KYUP compliance — the FCC is tightening its stance on enterprise voice providers in 2026.
This is the year AI voice agents leave the "experimental" phase and become production-ready. For Thai organisations handling hundreds to thousands of customer calls a day, the question is no longer "should we change?" but "where do we start so the ROI shows up in the first six months?"