AI voice agents
Make the introduction count.
Distinct voice agents answer inbound calls, work qualification questions into a conversation, and collect structured answers. They execute the conversational segment; your configured call flow determines the next destination. Nell is the intelligence and orchestration layer around that work—not the agent on the line.
The execution boundary
A conversation within your call flow.
- 01 Route an inbound call to the agent
- 02 Converse and capture the required answers
- 03 Return control to configured routing
- 04 Review the call and its disposition
Workflow overview—not a recording, timing guarantee, or customer result.
The problem
The expensive calls are the ones nobody handled properly.
A caller who reaches a buyer without being qualified costs you the buyer's patience and, eventually, the buyer. A caller who reaches voicemail at 7pm cost you the media spend and returned nothing. And a call that should never have been paid for at all — a duplicate, a fraud attempt, a consent question that never got asked — costs more than either, because it comes back later wearing a different name.
Both problems are the same problem: nobody was on the call at the moment it mattered.
Qualify
What happens on the call.
The agent answers when the call is routed to it.
Configure a voice agent as a destination in your inbound call flow: overflow, after hours, or a campaign’s qualification step. Review capacity, failure handling, and fallback behavior for your deployment.
She works your questions into a conversation.
Not a menu. She asks what you need asked, in an order that survives the caller answering out of turn, and follows up when an answer is incomplete. Callers can interrupt her mid-sentence and she stops and listens; if they'd rather press a number, the keypad works too.
Answers become fields, and fields become tags.
Each required answer is captured as a field on the call and, where you want it, as a qualifier tag routing can act on. The same fields map through webhooks into your CRM as the call ends, not in a nightly export.
She hands the call over warm — or ends it honestly.
A qualified caller is transferred to a buyer with the collected answers attached, so the human on the other end starts the conversation already knowing. A caller who doesn't qualify gets a proper ending and a disposition, and your buyer never pays for the call.
What a script looks like
Your questions, your required answers, your tags.
Two configuration examples, so the shape of the work is concrete. These are illustrations of how a script is set up — not defaults shipped in the product, and not anybody's live campaign.
Insurance — Medicare
| Question the agent asks | Required answer | Qualifier tag |
|---|---|---|
| Are you currently enrolled in Medicare Parts A and B? | Yes | medicare_ab |
| What's your ZIP code? | Valid ZIP in the buyer's licensed states | zip_verified |
| What's your date of birth? | Age within the buyer's accepted range | age_qualified |
| Are you the person on the policy? | Yes | decision_maker |
Home services — roofing
| Question the agent asks | Required answer | Qualifier tag |
|---|---|---|
| Is this for a home you own? | Yes | homeowner |
| What's the property address? | Address inside the buyer's service radius | in_service_area |
| Is there active water coming in right now? | Either — routes differently | emergency |
| Are you filing an insurance claim on this? | Either — captured for the buyer | insurance_claim |
Protect
One bad call can cost more than a good week earns.
Qualification is the half of this everybody sells. The other half is the calls that should never have been paid for — and those are only catchable while somebody is listening.
Fraud signals on the call, not on the invoice
Repeat callers arriving under fresh source IDs, scripts that don't match the traffic they were sold as, callers working a payout rather than buying — the agent is on the line while it happens, and every call gets the same post-call analysis afterwards.
Consent and compliance checked on analyzed calls
The questions you're required to ask are questions the agent actually asks, every time, in the same words. Analysis flags the calls where something required didn't happen.
PII redacted before a transcript is stored
Sensitive values are removed from the stored transcript rather than left in it with access controls in front.
Litigator scrubbing at a published rate
Available as a metered line item on the published rate table — $0.05 per pull, the same price on every plan.
One customer's voice agent flagged a fraudulent call within a day of being switched on.
A second call it flagged the same week turned out to be arguable, and we report that too. The false-positive rate is what decides how much of this you should let an agent do without a person checking, and it is a number we are still earning rather than one we are quoting. The customer is unnamed because we don't publish a customer's name until they have agreed to it in writing.
How it's built
The agent is a node in the call, not a number we forward to.
Moja runs the agent in-band, inside the active call session, using an external conversational runtime invoked from within our own call flow. That removes the separately dialled agent leg our earlier integration used, and it can avoid extra carrier legs entirely in the common case where the alternative forwards an already-anchored call back out over the PSTN to a hosted agent number.
Fewer call boundaries mean fewer opportunities for setup delay, jitter buffering, transcoding, duplicated transport charges, and fragmented failure recovery. That is a structural argument, and it is the honest size of the argument.
What we are not going to tell you: how many milliseconds faster this is, or what percentage cheaper. Actual latency and cost depend on the route, the codec, the region, transfer behaviour, the provider, and the contract — and we have not published an apples-to-apples measurement against another vendor. When we have one, it will be on this page with its method attached. Until then, anybody quoting you a number for this is quoting you a target.
The specifics
What the agents do today.
End-to-end call handling
Answer, converse, qualify, transfer or disposition.
In-band, not forwarded
The agent runs inside the active call session as a routing destination.
Barge-in
Callers interrupt; the agent yields and listens.
DTMF input
Keypad entry handled alongside voice.
Warm transfer
Hand-off to a human buyer with collected fields attached.
Webhook + CRM field mapping
Collected fields land in your systems as the call ends.
Real-time transcripts
Every agent call transcribed as it happens.
Script A/B testing
Run variants and compare outcomes in the reporting you already use.
Post-call analysis
Agent calls get the same scoring and review as human-handled calls.
Where it fits
She's a destination, so she goes wherever you route.
Visual Call-Flow Builder
Route to an agent like any other destination.
Nell
Configuration, orchestration, evaluation, and improvement around the job.
Pulse QAI
The review layer every agent call lands in.
RTB & Real-Time Bidding
Fallback and callback capture for traded calls.
Migrating to Moja
How your calls get here in the first place.
Pricing
The published platform fee, and what's quoted.