It answers the phone
The AI gets its own leg on the real call, bridged by SIP — not a bot on a web page. It speaks to the caller on the line they dialled.
Engagement Orchestrator
Building an AI that can hold a conversation is the easy part. Getting it onto a real phone line — with consent checked, capacity handled, a person ready to catch what it can't, and a bill at the end — is the platform. We built the platform.
Good morning, you're through to Acme. Am I speaking to Thabo?
Yes. I want to sort out my arrears before it goes to legal.
I can help with that. Your balance is R 4 218, and I can set up an arrangement over three months.
Can you take off the fees? I lost my job in July.
Fee waivers are outside what the AI may agree. Handing to Collections with the transcript, the intent and a suggested outcome — on the same call.
Elapsed
01:48
Intent
Arrears
Suggested
Arrangement
AI at the centre
It answers the call, holds the conversation, does the work it is allowed to do, and decides for itself when a person is needed. Everything else on this page exists to make that safe to do at scale.
The AI gets its own leg on the real call, bridged by SIP — not a bot on a web page. It speaks to the caller on the line they dialled.
The same AI, the same session, on WhatsApp, SMS, email, Messenger, Instagram and web chat. One conversation per customer, whatever they used to start it.
A context policy decides which customer fields reach the AI on a given campaign. Nothing outside that policy is ever sent, and what was sent is on the record.
The conversations are run by Untapped AI. The platform gives it the real phone line, the customer context it may use, and people to hand over to — within the rules you set.
The part everyone gets wrong
The hand-over is where most AI deployments fall apart: the customer is dropped into a queue and asked everything a second time. Here it is one call, all the way through.
STEP 1
Because the customer asked, or because what they want is outside what the AI may agree to.
STEP 2
Into a queue by skill, priority and channel licence — with a service-level clock already running.
STEP 3
Their device rings, a whisper tone plays to them alone, and the brief is already on their screen.
STEP 4
Onto the same call they were already on. The AI leg is released. The customer hears a person, not a new dial tone.
STEP 5
The AI's summary and suggested disposition are filled in; the agent corrects rather than composes.
To someone outside
A broker, a branch, a specialist on their mobile. They are rung while the customer waits on the line, hear a spoken brief of who is calling and why, and press 1 to take the call. No answer, or a different key, and the customer goes to the queue. Inbound or outbound, the same way.
Guard rails
An AI that transfers by itself can name a queue, a particular agent or an outside number. You decide which queues, whether it may pick an agent, and which numbers or prefixes it may use. Anything else is refused and logged, and the call takes its usual route. An AI never dials a number nobody listed.
When the AI can't
Nothing here assumes the AI is up, in capacity, or right. Every one of those has a defined answer, so a bad afternoon for the AI is not a bad afternoon for your customers.
After repeated failures the platform stops trying and sends the work to people instead of stacking up retries.
The dialler paces against AI capacity and agent capacity together, so it never creates a call nothing can answer.
The no-agent rule is yours to set: hold, voicemail, callback or a graceful goodbye — not an accident.
Every AI session, event, transcript and hand-over is traceable end to end, and audited.
Everything else
Not modules bought separately and bolted together. One data model, one set of permissions, one bill — built as a single platform.
Integrations
Connected from the marketplace in a few minutes, with credentials sealed at rest, a test before anything is saved, retries when a vendor is down, and a step-by-step guide for the other side.
Every finished call logged on the customer's record, with the outcome and notes.
A row per call in your warehouse; live queues and reports in your own tools.
Payment links during a call, and the channels customers already use.
Other contact centres, other AI, recordings to your QA tools, and everything else through a webhook.
Credit bureaus (TransUnion, Experian, XDS) and DebiCheck mandates are connected on request, under your own contract with them.
Built here, for here
Local numbering, local carriers, local messaging, and a privacy posture that starts from POPIA rather than retro-fitting it.
E.164 throughout, routing between the carriers you chose, and failover when one stops answering.
SMS through SMS Portal or Clickatell; WhatsApp, Messenger and Instagram straight from Meta; payments through PayFast, Peach Payments or Ozow.
Suppression, registries and quiet hours are enforced by the platform, not by a habit.
One tenant cannot read another's rows even if the application asks. Enforced by the database.
We will walk through a live call, the hand-over to a person, a payment taken on the call, a WhatsApp conversation and the bill that comes out the other end — using your queues, your carriers and your script.