BRENNEN

Industries / Finance and business services / Call centers and BPO

Self-service

Self-service is letting customers get answers or finish tasks without a person, through phone menus (IVR, interactive voice response), speech recognition, help pages, chatbots and AI voice agents, under rules on telling people they are dealing with a machine and letting them reach a human.

7

Milestones achieved, from 1981 to 2025.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1981

    Speech answers bank calls

    Japan's NTT combined speech recognition and speech synthesis in Anser, a phone information service for Japanese banks in which callers spoke from a 16-word vocabulary of ten digits and six control words. By 2000 it ran in more than 70 cities for over 600 banks and handled more than 360 million calls a year, by the account of AT&T researchers who call it "the first deployed voice-enabled service".

    eecs.yorku.ca

    Credit

    NTT

  2. 1992

    Machines take collect calls

    AT&T's Voice Recognition Call Processing (VRCP), built and trialed from 1986 to 1992, asked callers to say what kind of operator call they wanted (collect, calling card, person-to-person, operator-assisted or bill-to-third-number) from a five-word vocabulary. By 2000 it handled more than 1.3 billion recognition attempts a year with a 0.3% word error rate and saved AT&T over $300 million a year, by its own researchers' count; it introduced word-spotting (finding the key word inside a longer reply) and barge-in (letting the caller talk over the prompt).

    LessonThe early voice services that worked were simple, extended the touch-tone systems already in place and were built to "fail soft", recovering gracefully when the recognizer got a word wrong.

    eecs.yorku.ca

    Credit

    AT&T

  3. 1997

    Callers say what they want

    AT&T Labs' "How May I Help You?" system replaced the touch-tone menu with an open question and routed each call by understanding the caller's own words; a prototype was tested on over 20,000 live customers in 1997. AT&T researchers reported that a touch-tone menu routed only 84% of callers correctly, after 40% had given up and gone to an operator, while natural speech routed about 90% correctly. A fully automatic version was later deployed in the AT&T network for customer care.

    LessonMenus fail when the words a customer would use do not match the menu's labels.

    eecs.yorku.camirror.aclweb.org

    Credit

    Allen Gorin, Giuseppe Riccardi and J. H. Wright, AT&T Labs

  4. 2023

    A regulator flags chatbot loops

    The US Consumer Financial Protection Bureau (CFPB) estimated that about 37% of the US population used a bank's chatbot in 2022 and found that all of the top ten US commercial banks use them. Its issue spotlight of 6 June 2023 reported "numerous complaints" from customers stuck in "repetitive loops of unhelpful jargon" who could not reach a human representative.

    LessonThe route from a bot to a person is now something regulators examine, not only a design choice.

    consumerfinance.gov

    Credit

    Consumer Financial Protection Bureau

  5. 2024

    Airline liable for its chatbot

    In Moffatt v. Air Canada (February 2024), the chatbot on Air Canada's website told Jake Moffatt, booking flights after his grandmother died, that he could apply for a bereavement fare after travelling, while another page on the site said otherwise. British Columbia's Civil Resolution Tribunal rejected the airline's position that "the chatbot is a separate legal entity that is responsible for its own actions", held that "it makes no difference whether the information comes from a static page or a chatbot", and ordered damages of just over C$650.

    LessonA company answers for what its bot tells customers, so the bot's answers must match the company's published policy.

    barrysookman.comglobalnews.ca

    Credit

    BC Civil Resolution Tribunal

  6. 2025

    Klarna brings humans back

    In February 2024 the payments company Klarna said its OpenAI-powered assistant had handled 2.3 million chats in its first month, about two-thirds of its customer service chats and the work of about 700 full-time agents by its own figures. In May 2025 chief executive Sebastian Siemiatkowski told Bloomberg that the AI-first approach had given "lower quality" service, that Klarna was hiring people again, and that it is "critical that you are clear to your customer that there will always be a human if you want".

    LessonCheaper automated answers did not hold up on quality, and a guaranteed way to reach a person became part of the service.

    klarna.comentrepreneur.com

    Credit

    Klarna

  7. 2025

    Spain guarantees a human

    Spain's customer service law, Ley 10/2025, published on 27 December 2025, bars automatic answering systems from being the only channel, requires chatbots to let customers ask for a person "at any moment of the interaction, from its start", and requires 95% of those requests and of phone calls to be answered within three minutes on average. It covers providers of basic services (water, energy, transport, post, telecoms, finance) of any size and other consumer businesses with at least 250 staff or more than €50 million in turnover, which have twelve months to comply. Separately, Article 50(1) of the EU AI Act has required since 2 August 2026 that a chatbot tell people they are dealing with an AI unless that is obvious.

    boe.esartificialintelligenceact.eu

    Credit

    Spain's Cortes Generales; European Union

Watch for

  1. AI agents replace support staff

    Salesforce's chief executive said the company cut its support staff from about 9,000 to about 5,000 because AI agents now handle part of the work, with about 1.5 million conversations through AI agents and about 1.5 million through people in the same period and satisfaction "about the same".

    Achieved whenan outside party confirms both the headcount change and that resolution and satisfaction held steady on AI-handled cases, for example through company filings or an independent customer survey.

    itpro.com

    Claimed by

    Marc Benioff, Salesforce, September 2025

  2. Speech-to-speech phone agents

    OpenAI released gpt-realtime, calling it its "most advanced speech-to-speech model yet", declared its Realtime API ready for production voice agents and added SIP calling so an agent can answer ordinary phone lines. On Sierra's τ-voice benchmark of simulated phone calls, gpt-realtime-1.0 completed 30% of tasks; by April 2026 the best voice model, xAI's grok-voice-think-fast-1.0, reached 67%, against about 85% for reasoning text agents on the same tasks.

    Achieved whenan independent benchmark shows voice agents matching text agents on realistic phone audio, and a company running one on real phone lines publishes resolution and transfer figures that an outside party has checked.

    community.openai.comsierra.ai

    Claimed by

    OpenAI, August 2025

  3. US bill on AI disclosure

    The Keep Call Centers in America Act of 2025 (S. 2495) would make businesses that use AI in customer service say at the start of each contact that "a nonhuman, artificial intelligence or machine" is being used, and transfer the customer on request to a human operator physically located in the United States, with breaches treated as unfair or deceptive practices under the FTC Act. It was referred to the Senate Commerce Committee with no later action recorded.

    Achieved whenit passes both chambers of Congress and becomes law.

    congress.govcongress.wiki

    Claimed by

    Senators Ruben Gallego and Jim Justice, July 2025

Open

  1. 01

    Real callers for voice benchmarks

    Run the open τ-voice customer service tasks (retail, airline, telecom) with consenting human volunteers calling the agent over a real phone line, in place of synthetic callers, and publish how the scores differ from the simulator's on the same tasks.

    The brief
    Credit · unclaimed
  2. 02

    A reach-a-human test

    An open test suite that plays scripted customers against chat and voice bots and records whether the bot says it is an AI before its first real answer, whether a request for a person at any turn, including the very first, leads to a handoff, and how many turns and seconds the handoff takes. It publishes a tally for each bot and runs only on bots the tester controls or on companies that agree in writing.

    The brief
    Credit · unclaimed

Open 01 · the brief

Real callers for voice benchmarks

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Run the open τ-voice customer service tasks (retail, airline, telecom) with consenting human volunteers calling the agent over a real phone line, in place of synthetic callers, and publish how the scores differ from the simulator's on the same tasks.

Why it is still open

τ-voice (Sierra and Princeton, March 2026) voices every caller with text-to-speech and a language-model script; its authors write "We evaluate English only using TTS rather than recorded speech", say their simulator "is more patient than real users" and propose human user studies as future work. The September 2026 extension τ-Multilingual adds five languages but still uses simulated callers, and notes that "Simulated callers may not reproduce human timing or repair."

arxiv.orgarxiv.org

Built with what exists

Sierra's open-source τ-voice framework and tasks; a voice agent on a phone line through SIP, such as OpenAI's Realtime API; standard 8 kHz G.711 phone audio; ElevenLabs Scribe v3 to transcribe calls for review (see the Note "Fewer word errors"); a one-page written consent form for every caller.

  1. First step

    Run 20 τ-voice tasks with the built-in simulator against one voice agent, then pilot the same tasks with five volunteer callers.

  2. Done looks like

    30 consenting volunteers covering at least five accents or languages make 10 task calls each, 300 calls in all, each scored by the benchmark's own database check, with a published table that puts the human-caller pass rate beside the simulator's for the same tasks and transcripts released where callers agreed.

Open 02 · the brief

A reach-a-human test

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open test suite that plays scripted customers against chat and voice bots and records whether the bot says it is an AI before its first real answer, whether a request for a person at any turn, including the very first, leads to a handoff, and how many turns and seconds the handoff takes. It publishes a tally for each bot and runs only on bots the tester controls or on companies that agree in writing.

Why it is still open

Spain's Ley 10/2025 requires bots to accept a request for a person "from the start", and Article 50(1) of the EU AI Act requires disclosure, but the open agent benchmarks score neither rule. τ-bench's airline policy tells the agent to transfer "if and only if the request cannot be handled within the scope of your actions", the opposite of a right to a human on request; VAmoS Bench (July 2026) counts containment, the share of calls finished without a person, but does not describe testing requests for one. The published reach-a-human timings are one-off calls to live company lines, such as Call4me's 49-call test (October 2026), which its author calls a small, non-random sample.

raw.githubusercontent.comarxiv.orgruntimewire.com

Built with what exists

The τ²-bench user simulator to play the scripted customer; test bots on accounts the tester controls, such as a Realtime API voice agent on a SIP line; Browser Use 1.0 to drive web chat windows (see the Note "Deleting 80% to start over"); Gemini Decision 1 to score each transcript against the list of yes/no checks (see the Note "A thousand questions, one pass").

  1. First step

    Turn Article 8 of Ley 10/2025 and Article 50(1) of the AI Act into a written pass/fail checklist, and run it on two bots you build yourself.

  2. Done looks like

    40 scripted scenarios in Spanish, English and one more EU language, each run five times on at least ten bot setups, with a published table of the share of bots that disclose AI at the first turn, the share that hand off when asked at turn one, and the median turns and seconds to reach a person, all transcripts attached.