AI voice agents are the most over-promised category in marketing software right now. The demos are genuinely impressive and the failure modes are genuinely bad, and almost nothing written about them distinguishes the two.

This is a practical account of what they do well in lead qualification, where they fail, and how to deploy one without damaging the relationships you are trying to build.

Key takeaways

What a voice agent actually does

Stripped of the marketing, a voice agent is three things stitched together: speech recognition, a language model working from a script and your CRM data, and speech synthesis. It answers or places a call, follows a conversation, writes what it learns to the contact record, and either books time or hands off.

What makes it useful is not the conversation quality. It is availability. An agent answers at 11pm on a Sunday, on the first ring, every time — which no human team does.

Where they work well

Speed-to-lead on inbound enquiries

The single strongest use case. Someone fills in a form; the agent calls within sixty seconds while intent is at its peak, confirms what they need, and books a slot. The alternative is a rep calling back four hours later, by which point the prospect has contacted two competitors.

Qualification before a human's time is spent

Confirming budget range, timeline, team size or use case is structured work with a small number of branches — exactly what these systems handle reliably. The rep arrives at the call already knowing whether it is worth having.

Rebooking no-shows and reviving stalled leads

Low-stakes, high-volume, and nobody enjoys doing it. A no-show costs nothing extra if an AI attempts the rebooking, and the leads are already in a relationship with you.

Out-of-hours coverage

For businesses whose enquiries arrive in the evening — home services, clinics, hospitality — the comparison is not agent versus human. It is agent versus voicemail.

Where they fail

Cold outbound

Calling people who never asked to hear from you is a bad idea with a human and a worse one with a machine. Regulations on automated calling are tightening across jurisdictions, consent requirements vary, and the reputational cost of being the company that robocalls is not recoverable with a better script. If you take one thing from this article, take this.

Anything emotionally loaded

Complaints, cancellations, billing disputes, bad news. A caller who is already frustrated and realises they are talking to a machine becomes considerably more frustrated. Route these to a human immediately.

Complex or high-value discovery

A $50,000 deal involves reading hesitation, following an unexpected thread, and building trust. Voice agents follow scripts with branches; they do not do that.

Anything requiring genuine judgement

If the right answer is "it depends", the agent will confidently pick one. Confident wrong answers are worse than no answer, because the caller believes them.

How to deploy one without damage

• Start with inbound only. People who contacted you first. Prove the pattern before considering anything else.
• Write the handoff before the script. Decide exactly what triggers a transfer — frustration, an off-script question, a direct request — and make it work before going live.
• Give it one job. Qualify and book. Agents that try to answer pricing questions, handle objections and close are the ones that embarrass you.
• Cap the retries. Two attempts. An AI that calls five times is a harassment complaint.
• Listen to the recordings weekly. Not the metrics — the actual calls. Containment rate hides bad conversations that technically completed.
• Make the escape obvious. "Say 'representative' at any time." Never trap a caller in a loop.
• Write everything to the CRM. If the transcript and outcome do not reach the contact record, you have automated a call and lost the information.

What to measure

Containment rate — the share of calls handled without a human — is the metric vendors lead with and the one that misleads most. A call can be contained and still have annoyed the customer.

• Booked meetings per 100 calls, against your human baseline. This is the number that matters.
• Transfer rate and reason. Rising transfers for off-script questions means the script has a gap.
• Show rate for AI-booked meetings. If it books meetings nobody attends, it is qualifying badly.
• Call duration distribution. Very short calls usually mean people hanging up.
• Complaints, tracked deliberately. Low volume, high signal.

Questions to ask before you commit

Whichever way you are leaning, these are the questions that change the answer. Work through them against your own numbers rather than anyone's feature matrix — including this one.

• What does the handoff to a human look like, and how fast is it?
• What disclosure does it give, and when in the call?
• Where does the transcript go? If not the CRM record, it is lost.
• How is it priced — per minute, per call, or bundled?
• What happens when it does not understand? Guess, loop, or transfer?
• Can I hear real customer calls, not a demo reel?

If you cannot answer the first three, you are not ready to choose between Scalry and a voice agent vendor yet — you are still deciding what you need, which is a cheaper problem to solve first.

A realistic first thirty days

Most voice agent deployments fail in the first month, and usually for the same reason: they went live on too much, too early. This sequence avoids that.

Week one: shadow mode

Configure the agent but do not let it call anyone. Feed it historical enquiries and read what it would have said. This surfaces the hallucinated pricing and the confidently wrong answers while nobody is listening.

Week two: one narrow segment

Point it at a single inbound source — one form, one campaign — during business hours only, with a human available to take transfers immediately. Listen to every call. There will not be many, and the detail matters more than the volume.

Week three: extend the hours, not the scope

Add evenings and weekends, where the comparison is against voicemail rather than a person. Keep the same narrow script. Resist adding objection handling; that is the change that breaks deployments.

Week four: measure against the human baseline

Compare booked meetings per hundred enquiries, and show rate, against what your team achieved on the same source before. If booked meetings are up and show rate is flat, extend. If show rate dropped, the agent is booking people who should not have been booked — tighten qualification before doing anything else.

Frequently asked questions

Are AI voice agents worth it for lead qualification?

For inbound speed-to-lead, usually yes — calling a fresh enquiry within a minute at any hour is something human teams rarely manage, and it is where the measurable gains are. For cold outbound, complaints or high-value discovery, no.

Do I have to tell people they are talking to an AI?

Increasingly yes as a legal matter, depending on where you and the recipient are, and rules are changing quickly — check what applies to you. Practically, always disclose. People work it out, and discovering it mid-call is far worse than being told at the start.

What is a realistic containment rate?

For simple qualification, roughly 60 to 80 percent of calls completing without a human is a reasonable expectation. Treat higher claims with suspicion, and remember containment measures completion, not satisfaction — a contained call can still have gone badly.

Can a voice agent replace an SDR?

No, and vendors who imply otherwise are overselling. It replaces the queue an SDR cannot get to — after-hours enquiries, instant callbacks, no-show rebooking. The SDR still handles the conversations that need judgement.

Should I use voice agents for cold calling?

We would advise against it. Automated calling regulations are tightening, consent requirements vary by jurisdiction, and the reputational cost of robocalling is not fixable with a better script. Voice agents belong on people who contacted you first.