What AI did to call-centre productivity in support, sales and collections

AI lifted support output 14% across 5,179 agents, yet Klarna reversed its AI-first push. What the evidence says for support, sales and collections.

Call centres were meant to be the first place AI paid for itself. A few years in, the picture is more mixed than the pitch. One careful study of thousands of support agents found a real gain, concentrated among new staff. One company, Klarna, made a very public U-turn after going further than the evidence supported. Sales has only a few controlled studies, and they point in different directions, while collections runs on vendor claims and a new rulebook from the Reserve Bank of India. This case study takes each team in turn, sets out the metrics as they stand, and ends with what is still unproven and what is due next.

The short version

In a study of 5,179 customer support agents, an AI assistant raised issues resolved per hour by about 14% on average, and about 34% for the least experienced agents. Experienced agents gained little. Klarna said in 2024 that its AI assistant did the work of 700 agents. In May 2025 its chief executive said cost had been "a too predominant evaluation factor" and had led to "lower quality", and Klarna began hiring human agents again. For sales, the evidence is thinner and more mixed. In a 429-agent field study, an AI coach's edge over a human coach was largest for middle-ranked agents. A bot that hid its identity sold about as well as proficient humans, but telling customers it was a bot cut purchases by more than 79.7%. For collections, published recovery figures come from the vendors who sell the tools, with no baselines. The rules that matter most are the lender's own conduct rules, which take effect on 1 January 2027. Industry forecasters expect more automation, but also expect many companies to keep human agents. Both are forecasts, not results.

Background

A call centre is judged on a few numbers.

Average handle time (AHT) is how long an agent spends on a contact, including after-call work. First-contact resolution (FCR) is the share of issues solved the first time, with no repeat contact. Customer satisfaction (CSAT) is usually a survey score. Attrition is the share of agents who leave in a year. It is expensive because new staff take weeks to become productive. In sales, the headline number is conversion: the share of calls that end in a purchase. In collections, it is the share of overdue money recovered, and how borrowers are treated along the way.

Generative AI enters in two ways. An assistant sits beside a human agent and suggests replies in real time. A bot talks to the customer directly. The strongest studies cover the first. Klarna's story is mostly about the second. Mixing the two up is the most common mistake in headlines about "AI replacing call centres".

How an AI call centre works

An AI voice agent is software that answers or places a phone call and holds a spoken conversation. Five parts sit behind one call.

A phone number from a licensed operator. Inbound calls reach the company through a number, and outbound calls show a caller ID, both supplied by a telecom operator or a cloud-telephony provider. A SIP trunk. SIP, the Session Initiation Protocol, is the standard that sets up, transfers and ends a call, while the voice travels as small audio packets. The trunk is the link between the operator's network and the AI platform. LiveKit, an open voice platform, documents these pieces in its telephony guide: trunks that bridge the phone provider and the platform, rules for inbound and outbound calls, answering-machine detection, touch-tone input, transfers to a human, and region restrictions to meet local telecom rules. Speech-to-text, a language model and text-to-speech. On every turn, the first converts what the caller said into words, the model decides what to answer and what to look up, and the third speaks the reply, ideally before the whole sentence is written. Business systems. The agent reads loan or order records, sends payment links or books appointments through APIs. A human hand-off. When the agent is unsure, or the caller asks, the call transfers to a person with the history attached.

Delay is the make-or-break measure, because a pause of more than a second feels like a broken line. One telephony vendor puts a typical set-up at roughly 0.6 to 1.7 seconds from the caller finishing a sentence to the reply starting (about 0.1 to 0.3 seconds for recognition, 0.35 to 1 second for the model and 0.09 to 0.2 seconds for speech). That is a vendor estimate, not a test, and ordinary phone audio is narrower than app audio, which makes recognition harder. Cost is quoted per minute: Retell AI's published list price is $0.07 to $0.31 per minute for AI voice agents, depending on the voice and model chosen, before any enterprise discount.

Where things stand today

Benchmarks for call centres are scattered, and the sources disagree, so the numbers below come with their source and date.

Typical operating metrics. I could not verify public benchmarks for handle time, first-contact resolution or attrition: the figures I found came from vendor blogs and disagreed widely, so none is quoted here. Check how a benchmark defines each measure, and who compiled it, before using it as a target.

How far AI has actually spread. Gartner's own releases give the clearest recent picture. In a survey of 321 customer service leaders run in September and October 2025 and published on 28 April 2026, 31% had implemented or were planning AI-driven frontline layoffs through the first quarter of 2027. Another 63% were cutting frontline headcount gradually through attrition, 85% were adding tasks to agents' roles and 75% were moving agents into new roles. An earlier poll of 163 leaders, run in March 2025 and published on 10 June 2025, found 95% planned to keep human agents. In a survey of 199 leaders in April and May 2026 (published 26 August 2026), AI spending rose 38% while overall service budgets rose 2%. These are surveys and forecasts, not counts of what happened.

What customers say. In a Gartner survey of 3,566 customers run in February and March 2026, 49% said they would use a company chatbot if offered, but only 7% had used one in their latest service contact. Only 27% would try a chatbot again after a bad experience. 87% said companies using generative AI must offer a way to reach a human, and 50% said interactions were easier with it. Customers were about three times more likely to use a third-party generative AI tool than the company's own chatbot (Gartner, July and September 2026). Switch between the two surveys below.

Who is selling AI call centres. Three groups compete, and every figure below is company-reported unless stated.

AI-native platforms. Sierra, co-founded by Bret Taylor, raised $950 million in May 2026 at a valuation above $15 billion. It says more than 40% of the Fortune 50 are customers and reported $150 million in annual recurring revenue in February. Vapi, a voice infrastructure platform, raised a $50 million Series B at about a $500 million valuation after Amazon Ring chose it from more than 40 vendors. Vapi says it has handled over 1 billion calls and now processes 1 million to 5 million a day, and Ring says its satisfaction scores improved. TechCrunch lists Decagon, PolyAI, Bland, Retell and ElevenLabs in the same wave. India-built voice AI. Gnani.ai of Bengaluru reported FY25 revenue of ₹53.87 crore, up from ₹23.09 crore, and a net profit of ₹3.19 crore, according to Entrackr. It raised a Series B of about $10 million led by Aavishkaar Capital in March 2026 and says it handles over 30 million voice interactions a day in 12-plus languages for 200-plus enterprises. Exotel, a cloud-telephony company, says it facilitates more than 25 billion interactions a year and sells AI voice agents for support, sales, EMI reminders and collections. Ringg AI sells voice and chat agents for lead qualification and support. Intalk (intalk.io, from Agami Tech) sits underneath: a cloud contact-centre platform with a predictive dialler, voice and messaging APIs, a helpdesk and an AI voice bot, which can be hosted in the cloud, on premises or both. A predictive dialler places calls automatically and passes only answered ones to a free agent, which is the plumbing a human collections team already uses and an AI agent can plug into. Fyno works on messaging: it makes no calls, but routes and audits a lender's messages across its existing telecom providers, checking consent, DLT registration and Do Not Disturb status at the moment of sending, and claims 15 to 30% lower messaging costs. I found no independent result for any of them. The incumbent outsourcers. Concentrix and Teleperformance, the two largest call-centre operators, sell AI as an add-on to people. On 30 June 2026 Concentrix cut its 2026 revenue outlook to $9.93 to $10.03 billion from $10.04 to $10.18 billion, its operating income fell to $95.4 million from $148.3 million, its shares fell more than 21% in pre-market trading and Teleperformance's fell about 13%, according to The Next Web. The companies say AI also makes their own agents more productive.

The telephone rules in India. Calls ride on licensed operators' networks, so the telecom regulator TRAI decides who may call from which numbers. Promotional calls use the 140 series. Service and transactional calls from banks, financial firms and insurers, which include payment reminders, are meant to use a separate 1600 series assigned by the Department of Telecommunications. TRAI's direction of 19 November 2025 set the deadlines: commercial banks by 1 January 2026, large NBFCs (assets above ₹5,000 crore), payments banks and small finance banks by 1 February 2026, and remaining NBFCs, cooperative banks and regional rural banks by 1 March 2026. About 485 entities had already adopted the series, with over 2,800 numbers.

On 18 September 2026, TRAI's third amendment to its anti-spam regulations defined an application-to-person (A2P) call as a voice call started by software without a person dialling, including autodialling, robocalls and pre-recorded or artificial voice. Any entity using such calls must declare this to its operator with the numbers it will use, or the calls count as unsolicited commercial communication. A termination charge of up to ₹0.05 per minute applies, except for calls on the designated regulated series. Operators must flag suspected spam senders using AI tools, but the 140, 1600 and 1601 series will not be flagged as spam. An outbound AI call starts without a person dialling, so it plausibly falls inside this definition. Ringg AI's compliance guide reads it the same way, saying AI voice agents must be declared in advance or the calls count as spam, but that is a vendor's interpretation. I found no TRAI statement on AI voice agents specifically, and the start date should be checked in the regulation's text: a policy-law summary says most provisions start 30 days after Gazette publication, and some at 60 or 90 days.

India. NASSCOM's FY26 review, as reported by ETV Bharat, puts the technology and business-services industry at about $315 billion, up 6.1%, with direct employees up 2.3% from 58.15 lakh to 59.5 lakh, a gap its chairman called a "non-linearity". Business process management, which includes call centres, is projected to grow from $55 billion to $59 billion. That is the whole industry, not call centres, and the gap fits more than one explanation. For the call-centre part, Gulf News, citing a Reuters report (15 October 2025), puts India's BPO segment at roughly 1.65 million people and says annual headcount growth in business process management fell from tens of thousands to under 20,000. NASSCOM's full report is paid and I could not open it, so these figures are second-hand, and I found no official figure for voice work or the effect of AI on it.

Regulation. On 6 August 2026 the Reserve Bank of India issued amendment directions on loan recovery conduct for commercial banks (reference RBI/2026-27/223) and for NBFCs (RBI/2026-27/230), with a parallel amendment for housing finance companies. All take effect on 1 January 2027.

Disclosure rules abroad. Article 50 of the EU AI Act requires people to be told when they are talking to an AI system, and law-firm summaries say it became enforceable on 2 August 2026, with fines up to €15 million or 3% of worldwide turnover. Regulation (EU) 2026/1744, signed on 8 July 2026, delayed the high-risk rules to 2 December 2027 and 2 August 2028 and gave existing generative systems until 2 December 2026 for the content-marking duty in Article 50(2). I found no provision in its text delaying the duty to disclose AI interaction. One outlet reported a postponement of the transparency rules to December 2027, which looks like a confusion with the high-risk delay. India has no equivalent disclosure rule that I found for sales or collections calls.

What happened

Here is the evidence for each team side by side, before the detail. Team · Best evidence · Sample · What it found · How strongSupport, assistant · Brynjolfsson, Li and Raymond, Quarterly Journal of Economics · 5,179 chat-support agents at one software company · About 14% more issues resolved per hour, about 34% for novices · Strong: staged rollout, peer reviewed Support, bot · Klarna's own figures and later statements · Two-thirds of chats in its first month · Under 2 minutes to resolve, then the CEO said a cost-first approach led to lower quality · Company-reported Sales, coaching · Luo and colleagues, Journal of Marketing · 429 agents, then 100 and 451 in follow-ups · The AI coach's edge over a human coach was largest for middle-ranked agents · Randomized field tests, about a month Sales, bot calls · Luo, Tong, Fang and Qu, Marketing Science, 2019 · More than 6,200 customers · A bot nobody knew about matched proficient humans, and disclosure cut purchases by more than 79.7% · Randomized field experiment Sales, phone dialogue · Ma, Han, Lee and Kwon, ICIS 2024 · 90,000 participants · One-way messages had an 18.43% higher response rate than AI dialogue · Randomized, conference paper Collections · Gnani.ai and Edesy customer pages · One unnamed NBFC, one unnamed lender · 63% of targeted accounts repaying, 21% higher first-call recovery · Company-reported, no baseline ### Support: a measured gain for the newest staff

Erik Brynjolfsson, Danielle Li and Lindsey Raymond studied a staged rollout of a generative AI assistant to 5,179 customer support agents. Their paper, first an NBER working paper and later published in the Quarterly Journal of Economics, measured issues resolved per hour. Productivity rose about 14% on average. Novice and low-skilled agents improved by about 34%, while experienced and highly skilled agents saw minimal impact. The authors also report better customer sentiment and higher employee retention, and suggest the tool spreads the habits of the best agents to newer ones.

The paper's abstract adds that gains were largest for rare problems, where agents have the least experience, that customers were more polite and less likely to ask for a manager, and that the most experienced agents saw small gains in speed and small declines in quality. The company is a Fortune 500 business-software firm. The arXiv version counts 5,172 agents and about 15%, and the NBER version 5,179 and 14%, so quote the figure with its source. Because the rollout was staggered, the authors could compare like with like, which is why the study carries more weight than a typical vendor case study.

The gain depended on who was on the team. Move the sliders below, especially the share of novices, to see how the same tool gives a different team-wide result.

Support: the bot that went too far

Klarna, the Swedish payments company, launched an AI assistant built with OpenAI. According to Customer Experience Dive, in its first month the assistant took over two-thirds of customer service chats, 2.3 million in total, with average resolution under two minutes. Klarna said it did the work of 700 agents. Fortune adds that Klarna said in December 2024 it had made no human hires in the previous year, and had cut its headcount by over 1,000.

Then came the reversal. In a Bloomberg interview on 8 May 2025, chief executive Sebastian Siemiatkowski said cost had been "a too predominant evaluation factor" and led to "lower quality", and that it was critical customers know "there will be always a human if you want". Klarna began recruiting for an on-demand, remote support pilot. It did not abandon AI: a spokesperson told Fortune the company was "very much still AI-first", and Customer Experience Dive reports the assistant still handles two-thirds of inquiries, with response times improved 82% since launch and repeat issues down 25%. The lesson is not that the bot failed. It is that cutting people first, and watching quality second, did not hold up.

Sales: help for the already skilled

Sales has three controlled studies worth knowing, and they say different things from the support result.

An AI coach, not a bot. In a Journal of Marketing study by Xueming Luo and colleagues, 429 agents at fintech firms were randomly given an AI coach or a human coach for on-the-job training. The AI coach's advantage over a human coach was largest for middle-ranked agents and smaller at both ends, an inverted-U pattern. The authors say bottom-ranked agents were overloaded by the volume of feedback, and top-ranked agents resisted advice from software. In a second test with 100 bottom-ranked agents, limiting how much feedback the AI gave produced a significant improvement, because rookies seemed to be overwhelmed by full feedback. In a third test with 451 agents, an AI coach and a human manager working together beat either alone. The field tests ran about a month, so long-run effects are unknown.

A bot that makes the call. In a Marketing Science field experiment (Luo, Tong, Fang and Qu, 2019) with more than 6,200 customers, people were randomly given structured outbound sales calls from chatbots or human workers. When the chatbot did not reveal itself, it performed about as well as proficient human workers and roughly four times better than inexperienced ones. When customers were told up front they were talking to a bot, purchases fell by more than 79.7%. Customers rated the disclosed bot as less knowledgeable and less empathetic, even though it was objectively competent. To see what that drop could mean for a sales team, try the calculator.

Dialogue is not always better. A 2024 randomized experiment with 90,000 participants (Ma, Han, Lee and Kwon, an ICIS conference paper) compared AI-enabled two-way phone calls with one-way recorded messages. The one-way "monologue" calls had an 18.43% higher response rate. Within the AI dialogue calls, those stressing practical benefits did 16.59% better than those stressing enjoyment.

Cresta, a vendor, publishes conversion figures tied to particular call behaviours (for example 46% conversion when an agent "assumes the sale" against 34% when they do not), but these are correlations from the vendor's customers, not controlled results. I found no controlled trial of a live AI assistant on outbound sales calls of the kind the support study ran.

Collections: vendor results and a new rulebook

Vendors report impressive recovery numbers. Gnani.ai's customer story, for an unnamed auto-finance NBFC founded in 2010, says its voice agents handle routine dialling in the borrower's own language while human collectors keep disputes, settlements and hardship cases. It reports ₹20.03 crore recovered each month, ₹1.38 crore saved in monthly collections spend, 63% of targeted accounts repaying and a cost of 0.07 times manual calling. Edesy's marketing page carries an unnamed lender's quote: 84% pickup for Hindi voice AI against 58% for English-only agents, and 21% higher first-call recovery. Neither gives a baseline or comparison group that would let an outsider check the improvement, and both are company-reported.

On the rules, the RBI's text is specific for human conduct. For NBFCs, staff and recovery agents may contact borrowers only between 8:00 am and 7:00 pm unless the borrower asks otherwise. NBFCs must log the time and number of recovery calls, record call content, keep recordings for six months and tell the borrower a call is recorded. Every recovery communication must carry the grievance officer's contact details, and prohibited conduct includes abusive language, threats and excessive calls. The NBFC must supervise its recovery agencies and compensate borrowers for non-compliant recovery, and agents need an IIBF certificate, with a year's grace for those already engaged. The commercial-bank directions carry the same core rules. The RBI text I read does not mention automated calls, robocalls or AI, but TRAI's September 2026 amendment above does regulate automated voice calls, so a lender using an AI collections caller has to satisfy both. Test a call against three of the stated RBI rules below.

Why it worked (or did not)

It worked where AI taught. In the 5,179-agent study, the assistant gave the weakest agents the answers the best agents already used. That is why the average gain hid a large gain at the bottom and almost none at the top. Experienced agents had less to learn.

It failed where AI replaced and quality was not tracked. Klarna's early numbers were speed and volume. The chief executive's own words say that cost drove the decision and that quality suffered. A metric such as handle time can fall while resolution gets worse, because rushing a call is not the same as solving it. Gartner's Eric Keller makes a related point about chatbots in a September 2026 Q&A: leaders should "prioritize reliability over reach", with a visible path to a human, instead of a bot that tries every issue and often fails.

Sales is different, because the job is persuasion and trust. The controlled studies point in three directions at once. An AI coach helped the middle of the team most, and rookies needed less feedback, not more. A bot could sell as well as a proficient person, but only while customers did not know it was a bot. And, in a 90,000-person test, a simple recorded message beat an AI conversation on response rate. The common thread is that the customer's reaction to the machine, not the machine's skill, often decides the result. Any firm claiming a sales uplift should be asked for its control group.

Collections is bounded by conduct, not tooling. The risk is complaints and penalties, not speed. The rules apply to the lender whoever, or whatever, makes the call: the directions say the lender is responsible for its agencies. An AI caller has to clear three separate layers: the telecom rules on number series and automated calls, the data rules on recordings and customer data, and the lender's conduct rules. Whether the RBI's conduct rules apply to AI calls word for word is not answered by the text I read, so lenders should ask their counsel and not rely on a vendor's reading.

Experience explains part of the pattern, not all of it. In support, AI helped most where the human was least experienced and the issue was repeatable or rare enough to need a prompt. In sales coaching, the biggest gain was in the middle, not at the bottom. It helped least where the work is judgement, persuasion or an upset customer. A manager should therefore test by team and by task, not assume one result carries across.

Current limitations

The best study is narrow. The 5,179-agent study covers text-chat support at one company, measured by issues resolved per hour. It does not show that voice calls, sales or collections behave the same way. Speed is not the same as quality. The study's main measure is volume. Customer sentiment and retention are reported as improved, but long-run effects on loyalty are unmeasured in what I read. Customers do not yet trust bots. Only 7% of surveyed customers used a company chatbot in their latest service contact, and only 27% would return after a bad experience (Gartner, 2026). Gartner's advice is to prioritise reliability over reach and to give a visible path to a human. Delay, accents and line quality. A one-second pause feels like a broken line, phone audio is narrow, and vendor latency figures are not independent tests. I found no independent accuracy test of AI voice agents across Indian languages and accents. The telecom layer can stop a call. A call placed on the wrong number series, or an undeclared automated call, can be treated as spam under TRAI's rules, and operators can bar the numbers of repeat senders. Sales evidence is thin and dated. Three randomized studies, from 2019, 2020 and 2024, none on today's generative assistants running live calls. The coaching field tests ran about a month. The rest is vendor correlation. Disclosure is a trade-off. Hiding that a caller is a bot may raise sales, but disclosure is now legally required in the EU, and the 2019 study shows telling customers can cut purchases by more than 79.7%. No source I found tests what happens in Indian languages or collections calls. Collections evidence is mostly company-reported. No independent before-and-after study of AI collections calls was found, and no regulator statement on AI calls. Public benchmarks are weak. Handle-time, resolution and attrition figures come mostly from vendor blogs and are defined differently, so comparisons across firms are unsafe. India headcount is second-hand. The roughly 1.65 million BPO workers and the slowdown in hiring come from a Reuters report as relayed by Gulf News. Startup claims of handling 70 to 80% of queries are vendor claims. No official source gives the effect of AI on call-centre jobs. Vendor numbers are company-reported. Valuations, call counts and customer shares for Sierra, Vapi, Gnani.ai and Exotel come from the companies or their investors, and none comes from an audited source.

What is coming next

Dated items first, then forecasts.

Dated.

1 January 2027: the RBI's recovery-conduct directions take effect for commercial banks, NBFCs and housing finance companies. Lenders running AI voice systems will need call recording, time limits, identification and grievance details in place by then. By 1 January 2028: existing NBFC recovery agents without the IIBF certificate have one year from the effective date to obtain it. Mid-May 2027: the main duties on organisations that handle personal data fall due under India's Digital Personal Data Protection Rules. The government said they were notified on 14 November 2025 with an eighteen-month phased period. Call recordings and customer data used by voice AI will be in scope. Date to confirm: TRAI's third amendment of 18 September 2026 is notified, and its start dates for A2P pre-declaration and the termination charge should be checked in the regulation's text. The 1600-series deadlines for banks, NBFCs and other lenders have already passed (1 January to 1 March 2026). 2 December 2026: the EU grace period ends for marking AI-generated content in generative systems that were on the market before 2 August 2026.

Forecasts (not results).

Gartner's March 2025 prediction is that agentic AI, meaning systems that act on their own, will resolve 80% of common customer service issues without human intervention by 2029, with a 30% reduction in operating costs. Gartner's June 2025 prediction is that by 2027, 50% of organisations that expected to significantly cut customer service staff will abandon those plans. Gartner's 28 October 2026 prediction is that by 2029, 30% of employees who were terminated and replaced by AI will be rehired, often at a higher cost, because firms decided headcount before understanding the work. That one is about workforce planning in general, not call centres. In its August 2026 survey, service leaders expected generative AI chatbots to become their most valuable channel within two years, followed by live chat and generative AI voice bots.

These forecasts pull in different directions: more automation of routine issues, but humans kept for the rest. That is consistent with the evidence so far.

Open. Whether the RBI will say anything specific about AI or automated recovery calls, and whether independent studies of voice AI in collections will appear, are both unknown. A reader should watch the regulator's notices and any new peer-reviewed field studies.

What other companies can learn

Measure resolution and satisfaction next to speed. Do not cut on handle time alone. Use AI first to coach new agents. That is where the strongest evidence is. Expect smaller gains from veterans. Keep a human route. Klarna's reversal and Gartner's analysts say the same: customers will still need people. Ask vendors for a baseline. A recovery rate with no before-and-after comparison is a sales claim. Treat headcount stories as anecdotes until an official source gives numbers. Design for the rules before the tool. For collections, build recording, time limits and grievance paths first, then add automation. Run a controlled pilot. Give some agents the tool, others not, and compare the same metrics, as the large study did.

Key takeaways

AI raised support productivity by about 14% in a large study, mostly by lifting new agents. Klarna's public reversal shows that cutting cost first can cost service quality. For sales, controlled studies show AI coaching helps the middle of a team most, and that customers' reaction to a bot can decide the result: disclosure cut purchases by more than 79.7%. Collections results are vendor-reported, and the RBI's recovery conduct rules, effective 1 January 2027, apply whoever, or whatever, makes the call. Industry forecasts expect more automation but also more humans kept in the loop. Judge any rollout on resolution and satisfaction, not handle time alone.

Sources

- [Brynjolfsson, Li and Raymond: Generative AI at Work, NBER Working Paper 31161](https://www.nber.org/papers/w31161) - [Stanford HAI: first real-world study of generative AI and worker productivity](https://hai.stanford.edu/news/first-real-world-study-showed-generative-ai-boosted-worker-productivity-14) - [Fortune: Klarna, AI and human customer service, 9 May 2025](https://www.fortune.com/2025/05/09/klarna-ai-humans-return-on-investment) - [Customer Experience Dive: Klarna reinvests in human talent](https://www.customerexperiencedive.com/news/klarna-reinvests-human-talent-customer-service-AI-chatbot/) - [Ringg AI: voice AI compliance in India, 20 October 2026](https://www.ringg.ai/voice-ai-compliance-india) - [Intalk (Agami Tech): cloud contact centre, predictive dialler and AI voice bot](https://www.intalk.io/) - [Fyno: communication orchestration platform for BFSI](https://fyno.io/) - [TRAI: Telecom Commercial Communications Customer Preference (Third Amendment) Regulations, 2026](https://trai.gov.in/telecom-commercial-communications-customer-preference-third-amendment-regulations-2026) - [Tech Policy Law: TRAI notifies final spam rules, 20 September 2026](https://www.techpolicylaw.org/updates/trai-notifies-final-spam-rules-with-ai-based-detection-charges-on-automated-call) - [The Register: Gartner on AI and customer service staffing, 12 September 2025](https://www.theregister.com/2025/09/12/gartnersupportai/) - [Gartner: 50% of organizations will abandon plans to reduce customer service workforce, press release, 10 June 2025](https://www.gartner.com/en/newsroom/press-releases/2025-06-10-gartner-predicts-50-percent-of-organizations-will-abandon-plans-to-reduce-customer-service-workforce-due-to-ai) - [Gartner: 85% of service leaders expanding human agent responsibilities, press release, 28 April 2026](https://www.gartner.com/en/newsroom/press-releases/2026-04-28-gartner-survey-finds-eighty-five-percent-of-service-and-support-leaders-are-expanding-human-agent-responsibilities-despite-expectations-of-mass-ai-layoffs) - [Gartner: AI spending by customer service leaders up 38%, press release, 26 August 2026](https://www.gartner.com/en/newsroom/press-releases/2026-08-26-gartner-survey-finds-ai-spending-by-customer-service-leaders-has-surged-by-38-percent-despite-overall-service-and-support-function-budgets-rising-by-just-2-percent) - [Gartner Q&A: 87% of customers say a human agent must be available, 4 August 2026](https://www.gartner.com/en/newsroom/press-releases/2026-08-04-gartner-survey-finds-87-percent-of-customers-say-companies-using-genai-for-customer-service-must-provide-access-to-a-human-agent0) - [Gartner Q&A: only 27% of customers would try a chatbot again, 2 September 2026](https://www.gartner.com/en/newsroom/press-releases/2026-09-02-gartner-finds-only-27-percent-of-customers-would-try-a-chatbot-again-after-a-negative-experience) - [Gartner: 30% of employees replaced by AI will be rehired by 2029, press release, 28 October 2026](https://www.gartner.com/en/newsroom/press-releases/2026-10-28-gartner-predicts-30-percent-of-employees-replaced-by-ai-will-be-rehired-by-2029-if-organizations-dont-overhaul-workforce-planning) - [PIB: TRAI mandates the 1600 series for BFSI entities, 19 November 2025](https://pib.gov.in/PressReleasePage.aspx?PRID=2191647) - [PIB: TRAI amends anti-spam regulations including A2P calls, 18 September 2026](https://pib.gov.in/PressReleseDetailm.aspx?PRID=2311989&lang=1&reg=3) - [PIB: DPDP Rules, 2025 notified, 17 November 2025](https://pib.gov.in/PressReleasePage.aspx?PRID=2190655) - [EUR-Lex: Regulation (EU) 2026/1744 (Digital Omnibus on AI)](https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng) - [TechCrunch: Sierra raises $950M, 4 May 2026](https://techcrunch.com/2026/05/04/sierra-raises-950m-as-the-race-to-own-enterprise-ai-gets-serious/) - [TechCrunch: Vapi hits $500M valuation after winning Amazon Ring, 12 May 2026](https://techcrunch.com/2026/05/12/vapi-hits-500m-valuation-as-amazon-ring-chose-its-ai-platform-over-40-rivals/) - [Entrackr: Gnani.ai Series B led by Aavishkaar Capital, 30 March 2026](https://entrackr.com/exclusive/exclusive-gnaniai-kicks-off-series-b-round-led-by-aavishkaar-capital-11433897) - [Exotel: AI voice agents](https://exotel.com/products/gen-ai-voice-agents/) - [The Next Web: call-centre stocks slide as investors worry AI makes them uninvestable, 30 June 2026](https://thenextweb.com/news/call-centre-stocks-fall-ai-fears) - [LiveKit documentation: telephony](https://docs.livekit.io/telephony/) - [Retell AI: pricing](https://www.retellai.com/pricing) - [Gartner: agentic AI will resolve 80% of common customer service issues by 2029, press release, 5 March 2025](https://www.gartner.com/en/newsroom/press-releases/2025-03-05-gartner-predicts-agentic-ai-will-autonomously-resolve-80-percent-of-common-customer-service-issues-without-human-intervention-by-20290) - [RBI: NBFC (Responsible Business Conduct) Third Amendment Directions, 2026, RBI/2026-27/230](https://rbi.org.in/scripts/NotificationUser.aspx?Id=13672&Mode=0) - [TaxGuru: RBI recovery agent and loan recovery conduct directions for commercial banks](https://taxguru.in/rbi/rbi-issues-comprehensive-commercial-bank-recovery-agent-loan-recovery-conduct-directions-january-2027.html) - [Brynjolfsson, Li and Raymond: Generative AI at Work, arXiv version](https://arxiv.org/abs/2304.11771) - [Luo, Tong, Fang and Qu: Machines vs. humans, chatbot disclosure and customer purchases, Marketing Science 2019](https://ideas.repec.org/a/inm/ormksc/v38y2019i6p937-947.html) - [Luo, Qin, Fang and Qu: AI coaches for sales agents, Journal of Marketing (American Marketing Association summary)](https://www.ama.org/2020/11/19/coaching-sales-agents-use-ai-and-human-coaches-jointly/) - [Ma, Han, Lee and Kwon: Monologue vs. dialogue, a field experiment of AI-enabled telemarketing, ICIS 2024](https://aisel.aisnet.org/icis2024/humtechinter/humtechinter/4) - [ETV Bharat: NASSCOM FY26 technology industry figures](https://www.etvbharat.com/en/business/india-fy26-tech-revenues-hit-usd-315-billion-enn26022404879) - [Goodwin: EU AI Act transparency obligations now in force](https://www.goodwinlaw.com/en/insights/publications/2026/08/alerts-technology-dpc-eu-ai-act-transparency-obligations-now-in-force) - [Gnani.ai: auto-finance NBFC collections customer story](https://www.gnani.ai/customer-stories/auto-finance-nbfc-collections) - [Edesy: AI voice agent for payment collection, vendor page](https://edesy.in/ai-voice-agent/solutions/payment-collection-fintech-india) - [Kotak Neo: NASSCOM forecast for FY26 technology industry revenue](https://www.kotakneo.com/news/market-news/indian-tech-industry-revenue-fy26-nasscom-forecast/) - [Gulf News: AI agents take over India's BPOs](https://gulfnews.com/technology/ai-agents-take-over-indias-bpos-what-it-means-for-millions-of-jobs-1.500308112)

Tags: Call Centres Customer Service AI and jobs Productivity Collections