AI IVR in 2026: what the record shows when companies replace phone menus

Every AI IVR pitch leads with one number: the share of calls the agent handles without a person. Buyers compare those numbers across vendors as if they meant the same thing.

They rarely do. The public record on voice and automation claims says the containment figure is the single place a vendor can hide the most work.

This post reads that record. It sets one candid operator’s adoption curve against two federal enforcement files where the automation rate was the deception, and one regulator fine over what a bank’s AI did with recorded calls.

What is AI IVR?

AI IVR is a phone system that replaces the press-1 menu with a voice agent that understands what the caller says, resolves the request or routes it, and hands the rest to a person.

A conversational IVR is the same idea described from the caller’s side: speech in, intent out, no menu tree. The engineering is not the hard part any more. The measurement is.

What share of calls can an AI IVR system handle?

An AI IVR system handles a share that starts low and climbs slowly, on the most candid account a large operator has given in public.

Deutsche Telekom is replacing phone menus with AI agents across its European operations. Its chief AI officer, Kartik Sheth, said a newly switched-on use case handles “no more than 20% of those calls”, then 24% or 25% a week later, then 28%, reaching 40% “within a couple of months” [source].

Past that point, he said, 40% to 60% is harder, “60% to 80% is really hard, and anything above 80% is really, really hard” [source]. Company-wide, the share of interactions that never reach a human grows by one or two percentage points a month ([source]).

These are self-reported conference remarks, not an audited metric, and TIN holds no case file on them. They matter for a different reason: they are the shape of an honest curve, and they make a 95% pitch look like what it usually is.

Is a high containment rate a good sign?

No, a high containment rate is not a good sign on its own. The two clearest public records on automation rates are federal enforcement files, and in both the headline rate was produced by people the metric did not count.

Presto: 95% to 99% claimed, a human on every order. Presto Automation reported “automated order completion” rates of 95% to 99% for its drive-thru voice AI [source]. The SEC found the original proprietary version “required human agent intervention, including entering the order, in all instances”, using agents off site in the Philippines and India ([source]).

The trick was the denominator. Per the order, the rate counted only orders completed without restaurant staff, so the contracted agents typing the orders never showed up in it [source]. A drive-thru lane is not a phone menu, but the architecture is the same one an AI IVR buyer is purchasing.

Nate: “effectively zero percent”, by the charging papers. Federal prosecutors allege that the nate shopping app’s actual automation rate was “effectively zero percent” while its founder told investors it worked without human intervention [source]. The work was allegedly done by hundreds of contractors in a Philippines call center ([source]). These are allegations, not findings, and the founder is presumed innocent.

The detail that transfers is the dashboard. The DOJ alleges staff were told to keep the automation rate secret and access to the “automation rate dashboard” was restricted [source]. A vendor that will not show you how its rate is computed is telling you something.

What can an AI IVR system do with recorded calls?

An AI IVR system can analyse every recorded call, and that analysis carries its own legal exposure, separate from whether the agent answered well.

Hungary’s data protection authority fined Budapest Bank HUF 250,000,000 for running AI over its recorded customer service calls that guessed “the emotional state of the client” and ranked callers by likely anger for call-backs [source]. The authority ordered the bank to stop analysing emotions ([source]).

The analysis was “not in itself unlawful”, one law firm noted; the failures were a missing legal basis, no transparency and no effective right to object [source]. Every AI IVR that transcribes calls produces the raw material for exactly this, so the analytics roadmap belongs in the privacy review, not after it.

The proof

The three records this argument rests on are TIN case files, each checked against the primary document before it went live. Read them for the limits, which are stated on each page.

• SEC’s Presto order: a signed cease-and-desist order; Presto neither admitted nor denied the findings.
• DOJ and SEC charges against Nate’s founder: charging papers, so every figure is an allegation.
• Hungary’s NAIH and the bank’s emotion AI: the authority’s own decision; whether the fine was paid is not on the public record.

For the broader voice-agent picture, see what the record shows on AI voice agents.

How the records compare

Bottom line

The obvious way to buy an AI IVR is to pick the vendor with the highest containment rate. The record says that is the worst selection rule available, because the rate is the one figure a vendor controls end to end and the two worst public cases inflated exactly that figure. The operator who sounded least impressive, quoting 40% after two months, is the one whose number a buyer can plan around. The principle carries beyond phones: when an automation metric is the product, ask who the denominator leaves out before you ask how high it is.