TL;DR

  • Keep your IVR as the front door. Send chosen calls to an AI voice agent.
  • Ground the agent in a knowledge base. Let it say "I don't know".
  • Build and test a handover to a real person.

List your call flow, then pick a routing option

List your current call flow, then choose the pattern that changes the least on your phone system. Every company with a working call centre first asks whether a voice agent replaces its stack or plugs into it. Four patterns let it sit alongside what they already run.

a) Transfer from an IVR branch. Add a menu option, or a catch-all branch, that sends the call to the agent's phone number. The PBX usually does this with a SIP REFER message, which asks the other side to start a new call leg and hand the caller over.

b) Agent first, IVR as fallback. The agent answers every call. Callers who press keys, or who need a legacy menu, drop back to the IVR.

c) Overflow and after hours. The agent takes only calls that queue too long or arrive when the office is closed.

d) API or SIP trunk integration. Caller ID and IVR context, such as the menu path already chosen, pass to the agent. It can skip questions the caller already answered.

Pattern Effort Risk PBX change
a) IVR branch Low Low One menu option or transfer rule
b) Agent first High High Reroute the main number
c) Overflow Medium Low Queue-timeout or hours rule
d) API or trunk High Medium SIP trunk and context hand-off

Check three things on any pattern:

  • Caller ID. Test which caller ID your agent receives after a transfer.
  • Latency. Each extra hop delays a spoken reply, so keep the media path short.
  • Recording consent. A new system may record a transferred call, so confirm your announcement still covers it.

A consumer lender told us the agent had to connect to both its IVR and its internal database. That is pattern (d), the heaviest lift.

Insight7's inbound voice lines are in beta. A number people ring is answered by the AI agent, which can look the caller up. The agents also route calls intelligently.

Our advice is to start with outbound. Much inbound volume exists because companies never send proactive updates.

Map your IVR menu and calls first

Start with the call tree you already run. Pull a few weeks of IVR logs and list the top intents by volume. Then find the menu options where callers press 0. Those branches confuse people or fail them.

Sort each intent into two piles. The safe pile holds calls with a clear answer, such as status checks, FAQs, and appointment or payment questions. The person pile holds disputes, complaints and anything with many variables. Pick two or three safe intents for the first release.

Then check the telephony facts:

  • Concurrency. How many simultaneous calls can your trunk carry in a busy hour?
  • Number formats. Confirm the agent's number dials cleanly from your PBX.
  • CRM access. See what your CRM exposes through its API, because that limits what the agent can look up.

IVR logs show what callers pressed, not why they called. Call scoring on your existing recordings shows the real reasons. Insight7's call scoring evaluates 100% of conversations instead of a small manual sample. Scorecards are weighted and owned by your team. Voicemail, hold and too-short calls can be left out, so they do not skew your counts. It also connects to telephony systems such as RingCentral, Talkdesk and Five9.

Build the knowledge base the agent uses

Load the agent with approved facts and nothing else. A voice agent that answers from a small, clean source is safer than one that answers from memory. Buyers fear AI that sounds scripted. They also ask what happens when the customer goes off script. The knowledge base is your answer to both.

What to load:

  • Policies, such as refunds, fees and opening hours.
  • Product and pricing facts.
  • Approved wording for sensitive topics.
  • Your FAQ list.

Rewrite documents meant to be read. A policy PDF full of clause numbers sounds bad out loud. Turn each answer into one or two plain sentences a person could say in a breath. Callers cannot scroll back.

Ground the answers. Grounding means the agent first looks up passages in your knowledge base that match the question. It then builds its reply from those passages only. If nothing matches, the right behaviour is to say it does not know and offer a person. Insight7's voice and chat agents are grounded in the company knowledge base. Our knowledge base product ingests internal docs, product guides and playbooks as one source of truth.

Keep it alive. Name one owner for the content. Give every price and date a review date. Each week, read the transcripts and list the questions the agent could not answer. Write an approved answer for each gap, then add it.

Test before any caller hears it. In Insight7's voice campaigns, which are in beta, you can talk to the agent in your browser with no phone number needed. Ask it your ten hardest questions, plus three it should refuse.

Design the fallback to humans

Decide when the agent hands over before launch, and make every handover reach a person. Buyers raise this on their own. They ask whether the handover is polite, what happens with a sensitive issue or a disputed amount, and who chases a handover nobody picked up.

Callers expect an exit too. Gartner surveyed 3,566 customers, and 87% said a company using GenAI for service must offer a way to reach a human. Gartner also advises against making AI a mandatory first step. AI should try to resolve a call only when confidence is high, with a clear path to a person.

Set the triggers. Hand over when:

  • The caller asks for a person, in any wording.
  • The knowledge base has no match, or confidence is low.
  • The agent misunderstands the caller twice in a row.
  • The topic is sensitive or an amount is disputed.
  • The caller sounds angry.
  • A critical-error word appears, such as "fraud" or "legal".

Make the transfer warm. A warm transfer sends the caller ID, a short summary and the transcript to the human's screen before they say hello. The caller never repeats themselves. Keep cold transfers as a last resort.

Plan for nobody free. Offer a callback, voicemail or a ticket. Insight7's inbound voice lines are in beta, and callback requests there go to a queue. Name one person who clears that queue every morning.

Say it is AI. The first sentence of every call should disclose it. Rules differ by region:

  • EU: Article 50 has applied since 2 Aug 2026. Any AI that talks to people, including automated phone systems, must say it is AI from the first interaction.
  • US: AI-voice calls already need consent under the TCPA, following the FCC's 2024 ruling. A proposed rule to announce AI at the start of a call was not final as of mid-2026. Colorado's Chatbot Safety Act, effective 1 Jan 2027, also requires public-facing conversational AI to disclose it is AI.

Test, pilot, then widen the routing

Run scripted calls first. Then send a small live slice, such as 5 to 10% of one intent or one time window. Put a rollback switch in the IVR so one change sends everything back. Banks ask for this kind of short, low-risk pilot.

Track four numbers: containment, handover rate, transfer success, and QA score on AI calls against human calls. Score every AI call on the same weighted scorecard you use for people. Insight7's beta voice dashboard shows calls in progress and lets a reviewer read each transcript.

Measure from day one. Gartner's review of 432 AI service use cases found only a quarter with positive ROI. In 42%, leaders could not say.

Related reading: How to Use AI Voice Agents for First Contact and Pass Interested Customers to Sales Reps, Outbound IVR vs Conversational AI Voice Agents Compared for Customer Notifications and How to Automate Network Outage Alert Calls and Escalations for NOC Teams.

Questions people also ask

What is the difference between an AI voice agent and an IVR?

An IVR follows a fixed tree. Callers press keys or say set phrases, and every path is written in advance. A voice agent understands free speech, answers from a knowledge base, and asks follow-up questions.

What is a voice agent?

A voice agent is software that talks with callers in natural speech. It works out what they want, then answers or acts. Insight7's agents answer inquiries, qualify inbound leads, and route calls.

Can an AI voice agent work with a traditional IVR system?

Yes. Consider keeping the IVR for keypad entry of sensitive digits, such as account or card numbers. Send open-ended calls to the agent.

Do I need to replace my phone system to use an AI voice agent?

No. The agent gets its own number or SIP endpoint, and your PBX, handsets and carrier contract stay as they are. Start with one transfer rule for one intent. Deleting it is your undo.

How does an AI voice agent hand off to a human agent?

The agent spots a trigger and transfers the call to a human queue with the summary and caller details. Test with every agent logged out, because the no-answer path fails most often. In Insight7's beta, callback requests go to a queue.

How do you stop an AI voice agent from making up answers?

Restrict it to approved passages, and have it decline when nothing matches. Then try to break it. Ask questions your documents do not cover, and check that it declines and offers a person.

Does an AI voice agent have to say it is an AI on calls?

Often yes. Under EU AI Act Article 50, automated phone systems must say they are AI from the first interaction. Make the disclosure a fixed opening line the agent cannot skip.

Keep the IVR as your safety net

Give the agent more of the call tree only as the data proves it safe. Before you start, confirm that a real person answers a triggered handover.

Book a demo of Insight7