AI voice agents in 2026: how they fill your funnel and collect your receivables

For two years, voice agents sounded like a polite robot: they waited for your silence, interrupted at the wrong moment, and went mute while "thinking." That changed in 2026 — and it changed enough that putting a voice agent on the phone with your customers is now a reasonable business decision.

It's worth understanding what actually shipped, because there are two different things and they get confused constantly.

The two voice models you need to tell apart

GPT-Live-1 and GPT-Live-1 mini (OpenAI announcement, July 8, 2026) are the conversational models in ChatGPT. They're full-duplex: they speak and listen at the same time, so you can interrupt without them talking over you. They're the ones that say "hmmm," "mm hmm," "interesting" while you talk — what linguists call backchanneling — which is why it feels like someone is genuinely listening rather than waiting for their turn.

GPT-Realtime-2 (documentation, available in the API since May 7, 2026) is what you use to build business voice agents. It's OpenAI's first voice model with GPT-5-class reasoning: a 128k-token context window (four times the previous generation), parallel tool calling while it keeps talking, configurable reasoning effort, and native SIP support — meaning you can point Twilio, Telnyx or your own carrier straight at the endpoint and have the agent answer or place calls, with no intermediate transcription layer.

One practical detail that surprises people coming from the ChatGPT demo: OpenAI's own prompting guide recommends avoiding fillers like "Hmm..." or "Let me think..." before looking something up, and using a short, concrete line instead: "I'm checking that now." Natural backchanneling is lovely for conversation; on a collections call, what builds trust is the agent saying what it's doing.

Why this is business now, not a demo

The year's numbers explain it better than any argument. Zillow reported a 95% call success rate with voice agents, up from 69% with the previous model generation. Latencies run under 500 ms, the threshold below which a conversation stops feeling awkward.

And cost stopped being the obstacle: at current API rates — 32 USD per million audio input tokens and 64 per million output — a conversation runs about 0.05 USD per minute. A three-minute collections call costs cents. Compare that to the loaded cost of a phone agent making that same call a hundred and fifty times a day.

Case 1: collections, where the return is easiest to measure

Collections is the best first use case because the result is measured in recovered cash, not satisfaction. The 2026 data from callbot operations is consistent:

Preventive reminders before the due date: reduce accounts entering delinquency by 20-30%.
Early delinquency: 40-60% of accounts resolve with no human intervention.
Recovery on contacted receivables in the first 30 days: 15-25%, versus 8-12% for a traditional IVR.
Full cycle: 30-50% lower collection cost with 15-25% higher recovery.

The key isn't the script, it's the data. An agent that calls without knowing the exact balance, the outstanding invoice and the last payment applied only creates friction. One connected to your receivables can do this:

Agent: Good afternoon, I'm calling on behalf of Distribuidora del Norte regarding invoice A-4821, for 38,400 pesos, which came due on August 22. Can I help you schedule the payment?

Customer: I already paid that last week, I think.

Agent: I'm checking that now... I see a 20,000-peso transfer applied on September 1, which leaves a balance of 18,400. Would you like me to send the payment link over WhatsApp to settle the difference?

That second reply — looking up the applied payment mid-call and answering with the right figure — is exactly what parallel tool calling enables. The agent doesn't hang up to "verify and call you back."

Case 2: the sales funnel, where you lose more money than you think

In an SMB, the lead that arrives at 8pm on Friday gets handled Monday at 11. That delay is where the sale dies. A voice agent holds the funnel at four points:

Immediate contact. Call the lead within a minute of the form submission, at any hour, and qualify with three or four questions: what they need, by when, what volume, who decides.
A quote on the same call. With access to the catalog and inventory, the agent confirms stock and current pricing, and leaves the quote generated before hanging up.
Follow-up on cold quotes. The "let me look into it" nobody ever chases. A call on day three and another on day seven recover a surprising share of proposals that were already dead.
Reactivating dormant customers. Calling the account that used to buy monthly and hasn't in four months. It's the most profitable list in your database and almost nobody works it.

Agent: Hi, I'm calling from Aceros del Bajío about your quote request for PTR tubing. Are you still looking for material this month?

Customer: Yes, but I need 40 lengths and I was told there weren't any.

Agent: Let me confirm availability... We have 62 lengths available at the León warehouse. Shall I reserve the 40 and email you the quote right now?

What separates an agent that sells from one that annoys

Real data, not memory. Balance, invoice, stock and price are queried live; never improvised.
Permissions and limits. The agent shouldn't offer a discount or an extension a human rep couldn't authorize.
Logs and recordings. Every call transcribed and linked to the customer. In collections this isn't optional.
A clean handoff to a human. When the customer gets upset or the case gets complicated, transfer immediately with the context already loaded.
Respect for hours and frequency. Mexican collection regulation is clear, and reputation is more expensive than the receivable.

In Centella, the assistant operates on the same base where your customers, quotes, inventory, invoices and receivables live, under the same roles and permissions as the rest of the system. That's why it can query and act by chat or voice without inventing figures: we're not asking it to remember, we're giving it access.

Checklist before launching your first voice agent

Do I have my overdue receivables — balance, aging and last payment — in one place?
Will the agent be able to check inventory and current pricing during the call?
What can it promise without human approval: extensions, discounts, reservations?
Where does it transfer when the case gets complicated, and who's available then?
How do I measure results: effective contacts, payment promises, pesos recovered?

Frequently asked questions

Will customers notice it's an AI? Often yes, and that's fine as long as the agent is useful, resolves quickly and doesn't lie about what it is. What annoys people isn't the robot: it's the robot that knows nothing about their account.

Does it work for collections in Mexico, where people rarely pick up? That's precisely where it pays off most. The agent can retry at different hours and on different days with no meaningful marginal cost, and log every attempt.

Do I need to change my phone system? Not necessarily. With native SIP support it connects to providers like Twilio or Telnyx, or to your current PBX if it speaks SIP.

How long until it's running? A collections agent scoped to one flow — reminder, balance verification, payment link — goes to production in weeks, provided the receivables data is in order.

If your overdue receivables are growing while your leads go cold, you have the same problem twice: nobody is calling. Request a Centella demo, review our AI agent and integration services, or tell us about your operation and we'll tell you what to automate first.