AI Outbound Follow-Up Calls: what an AI voice agent can and cannot do on a real phone line
A voice agent is not a receptionist. It is a very consistent junior who handles the first thirty seconds perfectly and knows exactly when to pass the call.
- 14:02Call from +61 4** *** 219 rings out. Everyone is on a roof.
- 14:02Number checked: mobile, not a known spam range.
- 14:02SMS out: “Sorry we missed you, it’s Dan at Coastline.”
- 14:03Reply: “hot water system leaking, can someone come today”
- 14:03Dan paged. Job drafted with the transcript attached.
- 14:04Booked. Nobody touched a phone.
What it actually does
Audio streams into transcription, a bounded language model decides what to say next, speech comes back, and the whole loop stays inside the latency a human conversation tolerates.
Voice agents, document extraction, retrieval chat over your own material, and assistants with scoped access to your systems. In practice, AI Outbound Follow-Up Calls is the version of that we deploy when a business needs the result rather than a project. Built on Pipecat or LiveKit Agents depending on the call shape, with answering-machine handling and a strict do-not-call list.
Why businesses ask for this
AI is worth buying when it removes a specific, repeated, checkable task. It is worth avoiding everywhere else, and we will tell you which is which.
The people who get the most out of it: businesses with quotes going cold and no one to chase them.
- The repetitive half of a job handled, the judgement half escalated
- Answers grounded in your documents rather than invented
- Every automated decision logged so you can read it back
How we build it, step by step
The sequence below is the one we follow on every voice agent build. It is deliberately boring, because the interesting version is the one that breaks in month three.
- Define the narrow job the agent may do and everything it must hand off
- Write the prompt, the refusals and the escalation triggers before any code
- Build an evaluation set from real recorded calls and score against it
- Run it on our own line before it ever answers one of yours
What we change before it goes live
A reference implementation is a starting line, not a product. Every one we deploy gets the same treatment:
- Your numbers, your sender identity and your wording, so nothing reads as generic
- Secrets moved out of the code and into managed configuration
- Retries, rate limits and idempotency, so a hiccup never sends twice
- Structured logging and alerting, so a failure is noticed by us and not by a customer
- Consent, opt-out and record-keeping built in rather than bolted on
- Source control, a staging environment and a rollback that takes a minute
Compliance and risk
Two failure modes matter: confident wrong answers, and silent scope creep. We bound what the model is allowed to do, log every turn, and set a hard escalation path to a human.
The technical foundation
Open frameworks running on infrastructure you control, a model provider of your choice, and an evaluation set built from your real cases.
- python — Python, a good fit when the work sits next to data, scheduling or machine learning.
References worth reading before you buy this from anyone:
What it costs
Three ways to buy this, and the honest recommendation is usually the middle one:
- Starter build, from $2,500 — we build it, hand it over and warrant it for 30 days. Suits a business with someone technical in-house.
- Managed, from $390/mo — we build it and then own it: monitoring, changes, compliance upkeep and a monthly report. Suits everyone else.
- Platform, from $2,400/mo — when this is one of several systems and you want them designed as one layer instead of five.
Platform usage is billed at cost on top and itemised on the invoice. There is no margin on it and no minimum spend.
Common questions
How long does AI Outbound Follow-Up Calls take to build?
Two to four weeks. One week to build, one to two weeks running against real calls on our own line and tuning, then a staged rollout on yours.
What does it cost to run each month?
Two lines: our managed plan from from $390/mo, and platform usage billed at cost. Usage for this kind of system usually lands between $30 and $300 a month depending on volume. You see both itemised, and the platform account stays in your name.
Do we own it, or are we locked in?
You own it. The account, the numbers, the phone history and the source code are yours, and the foundation is open source. If you take it in-house, we hand over the repository and the runbook and that is the end of the conversation.
What if it breaks at 6pm on a Friday?
It is monitored. Failures raise an alert, the system degrades to something safe rather than silent, and hello@betr.agency is the inbox that answers. That is what the managed plan buys.
Can it work with the systems we already use?
Usually yes. Open frameworks running on infrastructure you control, a model provider of your choice, and an evaluation set built from your real cases. Where a system has no API, we look at whether an export, a shared inbox or a scheduled sync gets you 90 percent of the value for 10 percent of the cost.
Where to next
The product page for this build lists the specification, the timeline and what is included: AI Outbound Follow-Up Calls. If you want to talk it through against your actual process, a scoping call is 30 minutes and costs nothing.