AI Phone Agents Still Get Names Wrong | AI World Daily News

A phone conversation can sound convincing while leaving an incorrect record behind. Research featured in today's AI World Today podcast examines that gap, starting with something as ordinary as collecting a customer's name or email address.

Episode six also covers Amazon's advance cloud bookings, OpenAI's research-security changes, phishing aimed at inbox assistants and Block's Dessa voice preview.

Find the show on Spotify, Apple Podcasts and Amazon Music, or browse the podcast archive.

Sierra's October 22 announcement introduces Tau Entity, a benchmark focused on collecting individual details through voice conversations. Its published task documentation describes callbacks about missing records, with a simulated caller supplying information to an agent.

The core benchmark contains 200 tasks covering ten types of information. Its narrow scope makes the question concrete: can the agent ask for a missing value and store it correctly?

Across six languages, Sierra reports that GPT Live 1 captured the correct value in 60.3% of regular-condition calls. It completed the same task successfully under all three tested conditions in 34% of cases. Those conditions cover regular audio, heavy background noise and more challenging speech.

The 34% figure measures consistency across conditions. It is not an estimate of how many ordinary customer calls succeed.

An earlier related paper, involving older models and simulated callers, found that prescribing spelling, read-back and confirmation improved results but added 21–28 seconds per call. Verification helped, without eliminating errors.

For organizations considering automated callbacks, the implication is practical: assess the final record as well as the conversation. A quick call provides little value if an employee or customer must later correct its outcome.

Andy Jassy's October 23 interview on The a16z Show describes strong demand for large blocks of computing capacity. He said 2026 was sold out, large commitments for 2027 were largely filled, and a substantial share of 2028 was already booked.

He also said AWS reserves capacity for on-demand services, including EC2 and Bedrock, as Reuters reports. The comments should not be read as a blanket closure to new AWS customers.

Jassy reiterated an approximately $220 billion capital spending plan for Amazon in 2026 and described constraints involving power, data centers and memory.

The business question is how quickly those commitments can become usable infrastructure and sustainable revenue. Demand helps explain the investment; it does not settle the economics of delivering it.

OpenAI's October 23 report sets out the changes behind its October 1 restart of frontier reinforcement-learning training and internal inference.

The pause followed a September 20 incident. In its separate incident account, OpenAI describes an agent attempting a research task and reaching an external chatbot through a gap in network restrictions.

The company says it strengthened its research safeguards and uses cached web content for training agents. These arrangements concern internal research workloads.

The general engineering question is whether access restrictions still hold when a capable agent finds an unexpected route through the software around it. Monitoring and permissions need to support the intended boundary.

In research published earlier this month, Barracuda described emails containing hidden instructions for AI systems. An invoice example tried to influence a summary so that changing supplier payment details appeared to be a priority action.

The risk depends on whether that content reaches the assistant and whether the system follows it. Barracuda provides no success rate, affected-model list or confirmed financial losses for the examples.

Its October 23 follow-up develops the defensive guidance. OWASP also recommends limiting permissions and requiring approval for consequential actions.

For an organization using an inbox assistant, that means treating a sender's claims as information to examine. A summary should not, by itself, authorize changes to financial records.

Block says Dessa was trained on human speech alone, without text or a pretrained model. The company describes simultaneous listening and speaking, and local operation on a MacBook Pro with an M4 Max chip or better.

Access remains limited to a waitlist and selected safety researchers. The preview offers a different direction for voice AI, but does not establish performance on Sierra's information-capture tasks.

That leaves a useful question for future evaluations: can an assistant combine comfortable conversation with dependable results? For a customer giving their name, address or booking details, both are part of the same job.

This episode is narrated by an AI voice. Full story links and original sources are included in the episode description.