AI & Automation

Voice AI in a Contact Centre: Which Calls to Automate First and How to Keep Callers From Hating It

November 11, 2026 · 6 min read · Etzal Group Team

A utility's contact centre takes nine thousand calls a week. A third ask the same four questions: is there an outage in my area, when will it be fixed, what is my bill, and how do I report a fault. Callers wait eleven minutes for an agent to read an answer from a screen. A voice agent could answer in seconds, at three in the morning, in the caller's language. The same technology deployed on the wrong calls, such as a customer disputing a disconnection, would make a bad day worse. Voice AI is a question of which calls, and how the handover works.

What has changed

Older phone menus forced callers through "press one, press two". Speech systems that followed understood a few fixed phrases. Current voice agents combine speech recognition, a language model and speech synthesis, so a caller can speak naturally, interrupt and change topic.

That flexibility is the opportunity and the risk. The agent can do more, and it can also say things nobody approved. Design decides which.

Choosing the first call types

Analyse a few weeks of calls. For each reason for calling, note:

  • Volume.
  • Average handling time.
  • Whether the answer comes from a system lookup.
  • Whether it needs judgement or empathy.
  • What a mistake would cost.

Good first candidates:

  • Status enquiries: order, delivery, repair, application, outage.
  • Appointments: book, change, cancel, remind.
  • Simple account information after verification.
  • Fault or incident reporting that collects structured details.
  • Opening hours, locations, required documents.
  • Outbound reminders and confirmations.

Leave with people:

  • Complaints and disputes.
  • Cancellations where retention matters.
  • Vulnerable or distressed callers.
  • Anything involving a significant financial decision.
  • Cases with no clear rule.

Start with one or two call types that together cover a meaningful share of volume.

Design rules for a phone line

Voice is less forgiving than chat. Callers cannot scroll back, and silence feels broken.

Say what it is

"This is the automated assistant for the company. I can help with outages and bills, or connect you to a colleague." Disclosure, scope and the way out in one breath.

Keep turns short

One or two sentences at a time. A paragraph read aloud loses the caller.

Respond quickly

A pause of more than a second or two feels like a dropped line. If a lookup takes time, say so: "One moment while I check."

Let callers interrupt

People cut in. The agent must stop talking and listen.

Confirm what matters

Read back critical details: account number, date, address, amount. "I have Thursday the twelfth at two in the afternoon. Is that right?"

Handle not understanding

After two failed attempts, do not ask a third time. Offer another way: keypad entry, a text message link, or a person.

Always offer a person

A caller who says "agent" or "human" is transferred, without argument. Hiding the exit is the fastest way to be hated.

Hand over with context

When transferring, pass the caller's identity, the reason and what has been done, so the human agent does not start again with "how can I help you?".

Mind languages and accents

Test with real callers from your customer base, including regional accents, code-switching between languages, noisy lines and older callers. Performance in a quiet demo says little.

Keep it to approved facts and actions

  • Answers come from your systems and approved content, not from the model's general knowledge.
  • Actions are limited to a defined list: look up status, book a slot, log a report, send a message.
  • No promises outside policy: no compensation, no exceptions, no deadlines the system cannot see.
  • When the information is not available, it says so and transfers.
  • Sensitive actions need verification proportionate to the risk.

Security and privacy

  • Verification. Decide what proves identity for each action. Knowing a name and address is weak. Be aware that voices can be imitated, so voice alone should not unlock anything important.
  • Payments. Do not have callers read card numbers to the agent unless the solution is built and certified for it. Use a secure payment step.
  • Recording. Tell callers that calls are recorded and why, as the law where you operate requires.
  • Data. Know where recordings and transcripts are stored, for how long, and whether the provider uses them for training.
  • Manipulation. Callers may try to talk the agent into things. Limits must be enforced by the system, not by the wording of instructions.

Measures

  • Containment: calls completed without transfer, by call type.
  • Resolution: did the caller's need get met? Check by repeat calls within a few days and by sampling.
  • Transfer rate and reasons.
  • Abandonment: callers who hang up mid-conversation, and where.
  • Satisfaction: a one-question survey after the call.
  • Accuracy: from weekly review of a sample of recordings.
  • Waiting time for human agents, which should fall.

High containment with rising repeat calls means the agent is ending conversations without solving them.

A pilot

Weeks 1 to 3: prepare.

  • Call analysis and choice of call types.
  • Conversation design and approved content.
  • Integration with the systems that hold the answers.
  • Test with staff, then with recorded real scenarios.

Weeks 4 to 6: limited live.

  • A share of calls, in limited hours, with human agents ready.
  • Daily review of recordings.
  • Fix misunderstandings and gaps.

Weeks 7 to 12: expand.

  • More volume and hours.
  • Weekly review.
  • Decide the next call type from the data.

Involve contact centre agents from the start. They know what callers really ask, and their work changes: fewer routine calls, more difficult ones. Plan training and staffing for that.

Costs to count

  • Platform and usage, usually by minute or by call.
  • Telephony.
  • Integration with your systems, often the largest initial item.
  • Design and testing.
  • Ongoing review and improvement.

Compare against the cost of the calls it handles and the value of shorter waits, and count the calls it handles badly.

Common mistakes

  • Automating everything at once.
  • Hiding the route to a person.
  • Pretending to be human.
  • Long spoken answers.
  • No integration, so the agent can only talk and cannot do anything.
  • Testing only with staff in a quiet room.
  • Measuring containment alone.
  • No recording review.
  • Transferring without context.
  • Leaving agents out of the design.

We design and build AI agents for customer contact, including integration with business systems and review processes, as part of AI Consulting. If a few call types fill your queues, contact us.

Summary

Voice AI works when it handles frequent, patterned calls that a system lookup can resolve, and hands everything else to people with context. Disclose that it is automated, keep turns short, confirm critical details, always offer a person, and restrict it to approved facts and actions. Measure resolution and repeat calls, not only containment, review recordings weekly, and expand one call type at a time.

Frequently asked questions

Which calls are suitable for voice AI?

Calls that are frequent, short, follow a clear pattern and can be completed with information the system can look up: order or delivery status, appointment booking and rescheduling, opening hours and location, balance or bill enquiries after proper verification, and simple reminders. Calls involving complaints, disputes, distress or complex judgement should go to a person.

Should callers be told they are talking to an AI?

Yes. Say so at the start in plain words. Callers adjust how they speak, trust is preserved and some jurisdictions require disclosure. A voice agent that pretends to be human and is found out damages the brand more than any efficiency gain is worth.

How is the success of a voice AI measured?

By whether callers got what they called for. Track the share of calls resolved without transfer, the share transferred and why, repeat calls within a few days, caller satisfaction, and the error rate from reviewing a sample of recordings. Call duration alone is misleading, since a short call can simply mean the caller gave up.

Author

Etzal Group Team

Etzal Group Team, helping businesses grow with technology. Contact us.

Related service

WhatsApp AI Chatbot

We build, train, and run a WhatsApp AI agent that answers customers 24/7, qualifies leads, and follows up.