Industries / Finance and business services / Call centers and BPO
Agents and coaching
This is the work of the people who answer customers' calls and chats: how they are hired, trained and coached, where they work, why they leave, and the software that now prompts them as they talk.
7
Milestones achieved, from 2010 to 2026.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2010
The Philippines passes India
About 350,000 Filipinos worked in call centers in 2010, against 330,000 in India, by the count of the Contact Center Association of the Philippines, which also put expected Philippine call center revenue at $5.7 billion against $5.5 billion for India. The sociologist Jeffrey Sallaz dates the change earlier: in his 2019 book he says the Philippines became the world's "voice capital" around 2005.
Credit
Philippine call center operators; the count is the Contact Center Association of the Philippines'
2020
Paying to take calls
In October 2020 ProPublica and NPR's Planet Money reported on Arise Virtual Solutions, which had about 70,000 work-from-home agents in April 2020 answering for clients such as Disney, Airbnb and AT&T. Arise classed them as independent contractors who bought their own equipment, paid for client training (one paid $199 for an AT&T course) and paid Arise $19.75 twice a month. In March 2024 Arise agreed, without admitting wrongdoing, to pay $2 million to more than 250 workers in the District of Columbia plus $940,000 in penalties and to stop operating there; the US Department of Labor had sued in 2023, alleging more than 22,000 workers were misclassified.
LessonWhen home agents are treated as contractors who pay to train, the cost of the job falls on the worker, and regulators can challenge the label.
Credit
Ken Armstrong, Justin Elliott and Ariana Tobin (ProPublica); D.C. Attorney General Brian Schwalb
2020
India frees home agents
On 5 November 2020 India's Department of Telecommunications rewrote its Other Service Provider (OSP) guidelines, the rules covering call centers and other outsourcing firms. Registration was "done away with altogether", BPO firms doing only data work were taken out of the rules, and bank guarantees, static IP addresses and other requirements that had kept firms from "Work from Home" and "Work from Anywhere" were removed.
LessonTelecom rules, not technology, had kept India's voice agents in the office.
Credit
Department of Telecommunications, Government of India
2023
AI help lifts newest agents
Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked an AI chat assistant rolled out to 5,172 support agents at a Fortune 500 business-software firm, 89% of them outside the US, mainly in the Philippines. In the published version (Quarterly Journal of Economics, February 2025) the assistant raised issues resolved per hour by 15% on average and by about 30% for less-skilled and newer agents; agents with two months' tenure and the tool matched agents with more than six months' tenure without it, while the most skilled agents saw a small drop in quality. Customers were more polite and asked for a manager less often; the 2023 working paper had given 14% and 34% from 5,179 agents.
LessonThe assistant passes the best agents' know-how to new ones, so it does most where experience is thin.
Credit
Erik Brynjolfsson, Danielle Li, Lindsey Raymond
2024
Half may work from home
The Philippines' CREATE MORE Act (Republic Act 12066, approved 8 November 2024) lets IT and business process firms in economic zones, which get tax incentives for operating inside the zone, run work-from-home programs for up to 50% of their workforce. It followed a reversal: by Grant Thornton's account, ecozone firms were allowed up to 90% home work in late 2021, the Fiscal Incentives Review Board refused to extend it beyond 31 March 2022, the Philippine Economic Zone Authority then allowed only 30% for the rest of 2022, and some firms moved their registration to the Board of Investments, which permitted 100%.
LessonWhere tax breaks are tied to a building, tax law decides how many agents may work from home.
Credit
Congress of the Philippines; Fiscal Incentives Review Board; Philippine Economic Zone Authority
2025
No reading workers' emotions
Since 2 February 2025, Article 5(1)(f) of the EU AI Act has banned AI systems that infer the emotions of people at work, with narrow exceptions. The European Commission's guidelines of 4 February 2025 say the ban covers employees, not customers, so a call center may use voice analysis to read callers' emotions but may not infer its own agents'.
LessonIn the EU, coaching software may read the caller's mood but not the agent's.
Credit
European Parliament and Council (AI Act); European Commission (guidelines)
2026
Smaller gains in other firms
Shunyuan Zhang and Das Narayandas of Harvard Business School ran a year-long randomized field trial, published in Management Science, at a meal-delivery company with 138 agents: agents given real-time AI reply suggestions answered 22% faster, worth nearly five months of experience for new agents. A larger randomized trial at Alibaba, posted in 2026 by researchers from Fudan, Zhejiang and Dartmouth with Alibaba staff, gave a drafting assistant to 2,895 of 5,940 chat agents for four weeks in 2024; time to identify the issue fell 8.2%, chat length fell 1.1%, customer ratings rose 1.2%, repeat contacts did not change, and the top performers' service quality fell.
LessonThe gain depends on the firm, the task and the agent, and suggestions can slow down the best agents.
Credit
Shunyuan Zhang, Das Narayandas; Xiao Ni, Yiwei Wang, Tianjun Feng, Lauren Xiaoyuan Lu and Alibaba colleagues
Watch for
Accent translation for agents
Sanas sells a tool that changes an agent's accent as they speak. It says its AT4.0 model cut word errors by 27.0% on average for Indian-accented English, measured with an open speech recognizer (parakeet-ctc-1.1b) on its own unpublished set of contact center calls, and that blind American listeners preferred it to the previous model in 3,528 ratings of 168 clips; in October 2025 it added Kenyan, Nigerian and South African models with no error figures. CX Today reported that Teleperformance began rolling the tool out in its Indian centers in early 2025, and quoted critics who called it "playing into racism"; Sanas says it reduces accent-based discrimination.
Achieved whenan outside party tests it on held-out calls with consenting agents and publishes how well listeners understood and how the calls ended, not only machine word-error rates.
Claimed by
Sanas, September and October 2025
AI role-play in new-hire training
Transcom says an AI simulator that plays customers in a logistics client's new-hire classes raised training-class retention from 45% to 68% and the customer experience score from 3.83 to 4.41, and cut average handle time from 596 to 580 seconds. These are before-and-after figures with no control group or sample size.
Achieved whena randomized or matched-cohort comparison, run or checked by an outside party, shows new agents reaching proficiency sooner with equal or better quality.
Claimed by
Transcom, September 2026
Agents as supervisors of AI
A randomized trial covering 647 Taobao service workers and 680,676 chats in August 2024 found that chats an AI agent handled under human supervision were 16.8% shorter but rated lower. When the AI handed over a frustrated customer, the chat ran 40.8% longer than with human-only handling, repeat contacts rose 6 percentage points, ratings fell 0.928 points on a five-point scale, and workers put less effort into these handovers.
Achieved whenthe pattern holds in a second firm or the paper passes peer review.
Claimed by
Yiwei Wang, Chuan Zhu, Tianjun Feng, Lauren Xiaoyuan Lu and Bingxin Jia (Zhejiang, Fudan, Dartmouth, Alibaba), May 2026 preprint
Open
01
An open lab replication
A preregistered online experiment in which paid participants play support agents for a fictional shop, answering AI-simulated customers drawn from a fixed set of cases with known correct outcomes. Half get an AI reply assistant and half do not; the study measures correct resolutions, time per case and policy errors, split by each participant's skill at the start.
The briefCredit · unclaimed02
Accent fairness on phone audio
Recruit consenting speakers of Philippine, Indian, Kenyan, Nigerian and South African English to read and role-play scripted agent calls that use made-up names and account numbers, convert the recordings to 8 kHz telephone-band audio, and measure word errors, and errors on names, numbers and addresses, for the speech recognizers used in agent-assist tools.
The briefCredit · unclaimed
Open 01 · the brief
An open lab replication
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A preregistered online experiment in which paid participants play support agents for a fictional shop, answering AI-simulated customers drawn from a fixed set of cases with known correct outcomes. Half get an AI reply assistant and half do not; the study measures correct resolutions, time per case and policy errors, split by each participant's skill at the start.
Why it is still open
Brynjolfsson and colleagues write that their findings "apply for a particular AI tool, used in a single firm, within a single occupation", and the later trials (Zhang and Narayandas; Alibaba) are also single-firm field studies on private data. τ-bench supplies simulated customers, policy rules and a checkable end state, but it tests AI agents, not people. This research found no preregistered experiment with people handling simulated support chats with and without an assistant.
Built with what exists
τ-bench retail tasks (open source); an open-weights model for both the customer and the assistant so others can rerun it, such as Mistral Large 4 (see the Note "A trillion open parameters"); Prolific for paid participants; OSF for preregistration; oTree or a small web app.
First step
Pick 30 τ-bench retail tasks, write the preregistration and run a 20-person pilot.
Done looks like
A filed preregistration, at least 400 randomized participants, and published effects on correct resolutions per hour, minutes per case and policy violations by starting skill, with data and code released.
Open 02 · the brief
Accent fairness on phone audio
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Recruit consenting speakers of Philippine, Indian, Kenyan, Nigerian and South African English to read and role-play scripted agent calls that use made-up names and account numbers, convert the recordings to 8 kHz telephone-band audio, and measure word errors, and errors on names, numbers and addresses, for the speech recognizers used in agent-assist tools.
Why it is still open
Svarah (117 Indian speakers, 9.6 hours) and AfriSpeech-200 (200 hours from 13 African countries) are general or clinical speech, not call center tasks. AppTek's Call-Center Dialogues (April 2026, 14 accent groups, 128.6 hours of role-played calls) covers Indian and South African English but not Filipino, Kenyan or Nigerian, and its pages give neither the sample rate nor per-accent results. Sanas tests its accent tool only on its own unpublished calls.
Built with what exists
ffmpeg or sox to resample to 8 kHz and apply telephone encoding; open recognizers AppTek benchmarked (Whisper Large v3, Parakeet v3, Qwen3-ASR); ElevenLabs Scribe v3 (see the Note "Fewer word errors"); jiwer and AppTek's scoring script.
First step
Write 10 scripts and a consent form, then record five pilot speakers.
Done looks like
100 consenting speakers (20 per country) recording 10 scripted calls each, with per-accent word-error and entity-error rates and 95% confidence intervals published at full quality and at 8 kHz for at least four recognizers, and the audio released under an open license with consent.