Agentic AI on your inbound calls: what actually works

What an agentic AI actually does on a call

The word “agentic” gets stapled onto everything with a voice. The test is simple: can it complete a task in another system without a human, and does it know when to stop trying?

The old automation ladder went auto attendant, then IVR menus, then chatbots with a phone number. All of them route or recite. An agentic AI is different in kind: it listens in natural language, decides what the caller needs, and uses tools (a calendar API, a CRM, a helpdesk) to finish the job on the call. The underlying capability went mainstream when OpenAI’s Realtime API hit general availability in August 2025 with direct SIP calling into phone networks and PBX systems, which is why production deployments at firms like T-Mobile and Oscar Health exist today rather than in a slide deck.

In production, four jobs hold up well on inbound calls:

Bookings. The agent checks a live diary, offers slots and confirms, with the invite sent before the caller hangs up. Routing with context. Instead of “press 2”, the caller says what they need and arrives at the right person with a summary attached. CRM and ticket actions. New lead captured, contact record updated, or a level-1 ticket opened with the caller’s details already filled in. Escalation. A well-built agent transfers to a human with a warm handover the moment it is out of its depth, and that handover is a feature you should test, not assume.

How does that differ from the auto attendant you already have? Bluntly:

What still fails in production

The demo is always flawless. Production numbers are not, and the honest ones are published if you know where to look.

Start with the model itself. OpenAI’s own benchmark for gpt-realtime, the engine behind many commercial voice agents, scores 66.5% on complex function calling. That is a big jump from the previous generation’s 49.7%, and it still means roughly one in three complex tool calls goes wrong without careful guardrails. This is why serious deployments constrain the agent to a narrow task with hard rules rather than letting it freelance.

Then the deployment picture. Voice AI vendor IrisAgent, reporting across 500+ customer deployments, puts the industry-average resolution rate for routine, tier-1-eligible calls at 45–60%. Order-status queries resolve at 70–85%; complex complaints manage only 10–25%. Note the metric: resolution means the problem was solved, while containment only means the caller never reached a human. A vendor quoting containment may be counting callers who gave up.

Gartner supplies the caveat at portfolio level: it expects over 40% of agentic AI projects to be cancelled by end-2027 on cost, unclear value and weak risk controls, and estimates that of the thousands of vendors claiming agentic AI, only around 130 sell the genuine article. Its phrase for the rest is “agent washing”.

POPIA: the part the sales deck skips

An AI agent on your inbound line records callers, processes their personal information and usually ships the audio overseas. Three parts of POPIA apply before your first live call.

Tell callers they are talking to a machine

Call audio is personal information, and POPIA’s processing conditions require openness about how it is collected and used. The clean, defensible pattern is a short disclosure up front: the caller hears that an AI assistant is answering, that the call is processed and recorded, and how to reach a human. Burying it is a complaint waiting to happen.

Section 71: automated decisions need a human path

POPIA restricts decisions based solely on automated processing where they have legal consequences or substantially affect the person, and it gives the caller the right to make representations and to be told the logic behind the processing. An AI booking a boardroom is not a section 71 problem. An AI declining a credit application or triaging a medical query on its own could be. Keep consequential decisions with humans and give callers a working escalation route.

Section 72: your call audio is probably leaving the country

Most agentic platforms process audio on servers outside South Africa. Section 72 only permits that transfer on specific grounds: adequate protection under the recipient’s law or a binding agreement, the data subject’s consent, or necessity for performing a contract. Ask the vendor where audio is processed and stored, get the answer in the contract, and record which section 72 ground you are relying on. Your POPIA information officer, not the vendor, carries the accountability.

Will it work with your SA cloud PBX?

In 2026 there are two real routes into a South African phone system, and the right one depends on how much you want the agent to do.

Route 1: the PBX-native AI receptionist

The PBX vendors are building it in. Yeastar, widely resold in South Africa, launched an AI Receptionist inside its P-Series cloud and virtual PBX in April 2026: 24/7 answering in 34 languages, intent-based routing with conversational context, and a knowledge base built from your documents and website. The catch sits in the fine print: workflow execution into CRMs and appointment systems was flagged as a future capability at launch. Native receptionists answer and route brilliantly; check exactly what they can complete before calling them agentic.

Route 2: a SIP trunk bridge to a specialist platform

Dedicated voice-agent platforms connect to whatever PBX you already run. ElevenLabs’ agent platform, for example, documents plain SIP trunking into existing phone infrastructure: TLS 1.2+ signalling, G711 or G722 audio and optional SRTP encryption. Practically, your SA provider points a number or a trunk at the platform, and the agent sits in front of (or behind) your existing call flow. This route gets you real tool-calling and CRM actions today, at the cost of an extra vendor, an integration project and a harder POPIA conversation about where the audio goes.

How to roll one out without regretting it

Every deployment that works starts small, measures honestly and keeps a human one transfer away.

Keep reading