What an AI voice agent can and cannot do in a contact center today

A voice bot used to mean a rigid menu and an annoyed caller shouting "representative" into the phone. Large language models and good speech recognition changed that: today an AI agent sounds almost like a person. But expectations run ahead of reality, and companies that start without a plan end up disappointed. Here is what already works and how to launch a project.

What an AI voice agent does well

• Understands natural speech: callers say what they need in their own words, with no menu options.
• Answers frequent questions from the knowledge base: business hours, shipping terms, documents, prices.
• Looks up and updates data in your systems: order status, booking details, account balance, appointment time.
• Collects details – name, account number, reason for calling, address – and passes them to an agent or the CRM.
• Makes outbound calls: reminders, confirmations, short surveys, and past-due notices that follow your rules.
• Works 24/7 and handles the hundredth simultaneous call the same way as the first.
• Writes a summary of every conversation and tags the call reason automatically.

What it still cannot do reliably

• Handle complex complaints that call for a decision on compensation or an exception to policy.
• Calm down a very upset customer the way an experienced agent can.
• Work without integrations: without access to your systems it can talk, but it cannot solve anything.
• Answer accurately on topics it has no knowledge about – which is why it is limited to approved sources.
• Always understand speech on a very noisy line as well as a person does.

A simple rule: an AI agent is strong at frequent, clear-cut requests backed by data and weak at rare, ambiguous, and emotional ones. In a typical contact center it cannot replace agents entirely: it takes routine calls and peaks, and people handle the complex cases.

Pick your first call reasons

Do not try to automate everything at once. Start with data.

1. Take a month of calls and group them by call reason. Speech analytics does this automatically.
2. Find the two or three reasons with the most calls and the simplest logic.
3. Define what data the agent needs for each reason and which system holds it.
4. Estimate the impact before any work begins.
5. Agree on how you will measure success: share of calls resolved without a person, customer ratings, repeat calls.

Here is an example from the AI agent savings calculator. A contact center receives 20,000 inquiries a month, and the AI agent fully resolves 30% – 6,000 inquiries at 10 minutes each. Since agents spend 60% of their working time on inquiries, that frees up about 10 full-time positions. Per inquiry, 10 minutes of agent time, counting breaks and idle time, costs about $13.90 – nearly 28 times more than the AI at $0.50, and that gap repeats on each of the 6,000 inquiries every month. The calculator counts only agent time: sales from calls that used to go unanswered at night are not included. Plug in your own agent hourly cost and AI cost per inquiry, and run the AI readiness assessment to find weak spots before launch.

Connect your systems

An AI agent without integrations is an answering machine with good diction. The value comes when it takes action.

• CRM, such as Bitrix24 or Kommo (amoCRM): identify the caller by phone number and see their history.
• Order, booking, or billing systems: check status and make simple changes through an API.
• Knowledge base: the single source of answers, so the agent does not make things up. The agent finds the relevant passage and answers from it – an approach called RAG (retrieval-augmented generation).
• Calendar or schedule: book a customer and reschedule a visit.
• Telephony: transfer the call with its context to the right queue.

Time to launch depends mostly on integrations. A pilot on one or two call reasons with ready APIs goes live much faster than a full implementation.

How a conversation with an AI agent works. Every reply takes three steps. Speech recognition turns the caller's words into text. The language model reads that text along with its instructions, looks up the answer in the knowledge base or pulls data from the CRM if needed, and decides what to say. Speech synthesis voices the reply. That is also where the weak spots are. If speech recognition on a noisy line hears "fifteen" instead of "fifty," the model works faithfully with the wrong number – which is why the agent reads numbers and dates back to the caller. And if the answer is in neither the knowledge base nor your systems, there is nothing useful to say, and the right move is to transfer the call to a person.

Plan the handoff to a live agent

The handoff is where a project either earns customer trust or loses it. Callers should never feel trapped.

• Say at the start of the call that an AI is answering. Callers are fine with it when the agent solves the issue quickly and transfers to a person on request.
• Transfer right away whenever the caller asks for a person, at any point in the call.
• Transfer when the agent is unsure, the topic is outside its scope, or the caller is upset.
• Pass the summary, the collected details, and the reason for the transfer to the agent so the customer does not have to repeat anything.
• After hours, offer a callback instead of a dead end and create a task for the morning shift.
• Review transferred calls every week: they show what to teach the agent next.

Launch, measure, expand

1. Start with a share of calls or a single line, and listen to real conversations every day.
2. Update the knowledge base and the agent's instructions based on real mistakes, not guesses. Review its errors the way you would a new agent's.
3. Track the share of calls resolved without a person, transfers, average call duration, and customer ratings.
4. Add the next call reason only when the first one works well.
5. Keep people on quality assurance: review a sample of AI conversations the same way you review agent calls.

Where projects usually start

Successful projects most often start with requests that are frequent, repetitive, and rely on data stored in one of your systems. Examples that fit many industries:

• Order, delivery, or service request status: the agent identifies the caller and reads the status from the system.
• Booking and rescheduling appointments: the agent checks open slots and confirms the new date.
• Payment and contract reminders: outbound calls with a short conversation.
• Lead qualification: the agent finds out what the customer needs, collects details, and transfers the call with a summary.
• After-hours calls: the agent answers frequent questions and logs callback requests for the morning shift.

Outbound calls from an AI agent follow the same rules as calls from people and dialers. In the US, the FCC has ruled that AI-generated voices count as artificial voices under the TCPA, so these calls need the same prior consent as prerecorded calls. Respect the National Do Not Call Registry and calling hours, and do not use an AI agent for cold sales calls without consent.

Where the agent runs and what it costs

A voice agent processes personal data and call recordings, so decide early where the platform runs and where data is stored. US privacy laws such as the CCPA/CPRA in California and similar state laws set rules for collecting and using this data, and some states require every party's consent to record a call, so play a recording notice. CONTACT.WORK runs AI voice agents in our data centers in the US and the UK or on the customer's own servers. That matters for companies that must keep data on their own infrastructure.

The CONTACT.WORK platform is free: you pay for technical support under your support plan. Speech recognition, speech synthesis, and the language model for the AI agent are paid separately – or you run them on your own GPU or a Mac mini, and the data never leaves your network.

What to do this week

1. Export a month of calls and find two or three frequent call reasons with simple logic.
2. Check which systems hold the data for the answers and whether you can connect to them through an API.
3. Run the savings calculator with your own numbers: inquiries per month, handle time, agent hourly cost.
4. Decide where the agent will run – in the cloud or on your servers – and who will review transferred calls every week.
5. Pilot it on one phone line: within a few weeks, you will have enough data to decide whether to expand.

AI voice agents run on one platform with telephony, the IVR, the knowledge base, and reporting. A call moves from the AI agent to a live agent with its full history, and both show up in the same reports.