AI voice agent or IVR? Choose, replace in stages and keep a human handoff.

Both designs answer the call and route the caller. They differ in how the caller expresses the request, what happens when the system does not understand, and how much has to be tested before launch.

On Twilio, for example, a classic IVR collects input with <Gather>, which accepts dtmf, speech or dtmf speech. A conversational agent can run through ConversationRelay, which converts speech to text and streams it to your own application over a WebSocket. Amazon Connect calls an Amazon Lex bot from its Get customer input block, and Google offers a Dialogflow CX phone gateway. Platform documentation was checked on 13 October 2026; features and availability change.

Keep the existing route working while the agent earns each part of it. A staged replacement lets you compare like with like and switch back quickly.

1. Baseline the current IVR. For a fixed date range, record calls per menu option, callers who press for an operator or ask for a person, transfers between teams after the menu (misroutes), abandoned calls and after-hours volume. Use the same definitions after the change.
2. Keep keypad input working. Some callers prefer keys, and noisy lines defeat speech recognition. Gather’s dtmf speech mode accepts either, ConversationRelay can forward keypresses with dtmfDetection, and Amazon Connect’s Lex integration accepts “press or say” input.
3. Start with one intent or a defined share of calls. Choose a frequent, low-risk request such as opening hours or a new enquiry. Leave the rest of the menu unchanged so the remaining callers are unaffected.
4. Write rollback criteria before go-live. Decide in advance which signals pause the agent, for example rising requests for a person, failed transfers or records that do not match what callers confirmed, and who can restore the old route.
5. Test the handoff, not only the conversation. With ConversationRelay, your application ends the session with handoff data and Twilio calls your action URL, which can return a <Dial> to a person. Twilio’s warm transfer tutorial uses a conference so the caller is not dropped. Test busy, unanswered and abandoned transfers.
6. Retire menu options one at a time. Remove an option only when the replaced intent meets the targets you set at baseline. Keep a simple menu or operator route as the fallback layer.

Several vendor pages publish containment or first-call-resolution ranges for IVR versus AI agents without naming a study or defining the terms. Treat those numbers as marketing claims. Define containment for your own pilot, for example calls resolved without a transfer and without a repeat call from the same number within a period you choose, and measure both routes with that same definition.

A business with three departments and predictable requests may need a short IVR menu. A business receiving open-ended enquiries may benefit from an agent that asks a relevant follow-up question. More conversational freedom also creates more cases to evaluate: misunderstood intent, unsupported answers, interruptions and incomplete records.

Speech input alone does not make a system a conversational AI agent. For example, Twilio’s Gather documentation supports speech and keypad input for collecting caller responses. The distinction is what the workflow does with that input and which actions it is allowed to take.

A hybrid can retain familiar routing while adding a conversational intake for one enquiry type. Use the decision table above to choose the first workflow to test.

Rules depend on where your callers are, whether calls are inbound or outbound and what you record. Confirm them with qualified counsel for your jurisdictions; this section is general information, not legal advice.

• Outbound AI calls in the United States. The FCC’s February 2024 declaratory ruling treats AI-generated voices as “artificial” voices under the Telephone Consumer Protection Act, so the consent rules for artificial-voice calls apply. It concerns calls placed to consumers, not callers who phone you.
• Call recording. US federal law (18 U.S.C. § 2511(2)(d)) allows one-party consent, but several states require every party to consent. Announcing the recording at the start of the call and offering an alternative is the safer default.
• Telling callers they are speaking to AI. Article 50(1) of the EU AI Act requires providers to ensure people are informed that they are interacting with an AI system, unless that is obvious, and applies from 2 August 2026. Introducing the assistant as automated, as in the example below, also helps callers decide whether to ask for a person.

Imagine a services company receiving a call about connecting website enquiries to a CRM. The caller wants to discuss the current process with someone. This is a design example, not a live-call test, customer deployment or measured result.

1. Introduce the assistant and the exit. “You’re speaking with the automated assistant. I can collect your enquiry or help you reach the team.” State the available way to request a person, and actually implement it.
2. Identify the request. Ask what the caller needs. Keep the original meaning alongside any proposed category; “website leads are getting lost” should not become a confirmed software requirement.
3. Ask one useful clarification. For this request, ask which CRM the team uses. Leave an unknown answer blank rather than guessing.
4. Confirm the summary. “You want help getting website enquiries into your current CRM. Is that right?” Let the caller correct it.
5. Offer an available next step. Attempt the approved team handoff during covered hours. Otherwise, offer a callback request and confirm the contact details the caller chooses to provide.
6. Verify and close. Confirm that the handoff connected or the callback request was saved before describing it as complete. Do not promise a callback time the team has not agreed to cover.

The agent does not quote a project, promise compatibility or approve a contract. Those decisions stay with the team. The human approval workflow guide explains how to connect a proposed action to an explicit decision.

Use a stable call reference to join the intake, transfer attempt and eventual business outcome. A short structured record is easier to act on than an unreviewed transcript.

• Request: caller’s stated need and any correction.
• Known context: current system, requested next step and facts the caller confirmed.
• Unknowns: fields still unanswered; do not substitute model inference.
• Routing: destination, attempt time and confirmed result.
• Follow-up: named owner, agreed action and caller’s preferred contact method.
• Outcome: pending review, reviewed fit, not a fit or no further contact requested.

Keep operational contact details in the authorised customer system. Aggregate reporting needs references and outcome categories, not raw audio or transcripts. Decide separately whether recording is needed, who can access it and when it is removed before enabling it.

A human request should interrupt qualification rather than trigger another sales question. If nobody is available, say so and offer the defined alternative. A request to stop should end the automated exchange; it should not silently enrol the caller in a follow-up sequence.

• Silence or poor recognition: repeat once in simpler language, then offer the implemented keypad or human route.
• Unsupported question: state the limit and transfer the question instead of improvising pricing or policy.
• Busy or unanswered transfer: return to the caller with an honest status and callback option.
• Caller disconnects: preserve the partial state; do not mark a transfer successful or automatically call back without an agreed basis.
• CRM write times out: reconcile using the call reference before retrying, so one enquiry does not create multiple tasks.
• Repeated provider event: deduplicate it before updating the record or notifying an owner.

Provider events need explicit interpretation. Twilio’s Dial reference documents connecting an active caller to another party and returning the attempt’s outcome through an action URL. Starting that attempt is not evidence that a teammate received the enquiry.

Keep connected calls, completed intake, connected handoffs, reviewed qualified enquiries and confirmed bookings separate. Twilio’s Call resource documentation defines telephony lifecycle statuses; those statuses do not establish business fit or meeting attendance.

Report the handoff connection rate as connected transfers divided by attempted transfers. Report qualification as team-confirmed fit divided by reviewed valid enquiries. Keep pending reviews visible. Count a booking from its authoritative booking record and attendance from the team’s confirmation.

Also inspect recognition failures, callers requesting a person, abandoned transfers and duplicate records. Review representative successful calls as well as failures: a fluent summary can still contain a wrong detail. Exclude test calls and known spam from the business funnel, while retaining their operational counts.

Compare equivalent intake sources and staffed hours before and after a limited pilot. This guide supplies no latency, conversion or cost benchmark because no live telephony experiment was performed.

• Choose one inbound purpose, supported language and human destination.
• Approve the introduction, allowed knowledge and caller exit behaviour.
• Confirm coverage hours and the fallback owner.
• Test correction, silence, interruption, human request, busy transfer and failed record creation.
• Verify that saved records match what callers confirmed.
• Agree on recording, access and retention settings for this implementation.
• Define the conditions that pause automation and restore the existing routing.

Marsen Agents implementations are scoped around your tools, available access and review requirements. Bring the current call route and one representative enquiry to a workflow walkthrough so the first design has a clear boundary and a measurable next step.