AI voice agent vs a human call centre: what the evidence actually says
We have no client case study to sell you, so this page cites other people's — a peer-reviewed study of 5,172 agents, a national tax agency's own containment rate, an Indian airline's published deflection figures, and a government audit of what fixing a helpline the human way costs. Every number is linked to the organisation that published it.
- 9 sources
- Each opened and read, not taken from a snippet
- 0 of ours
- AiSewak has no published client outcomes. None are implied.
- Named orgs
- Peer-reviewed, government and first-party enterprise only
- Caveats kept
- Forecasts labelled as forecasts, surveys as surveys
Almost every claim you will read about AI voice agents replacing call centres traces back to a vendor. We checked. The widely-quoted containment rates of 60 to 90 percent, the "cut cost per call by half" headlines, the Indian per-minute cost comparisons — all of them lead to marketing pages with no primary source behind them, and several lead back to this very website's blog, which is not evidence of anything.
So this page does something less flattering and more useful. It collects what named organisations have actually published about conversational AI against human contact centres, links each one, and states plainly what each figure can and cannot be used to argue. Where a number is a forecast, it says so. Where a study measured AI helping human agents rather than replacing them, it says that too.
None of it is an AiSewak result. We are an early company with live demo agents and no client deployment whose outcomes have been published. Any vendor showing you a glossy case study should be asked which client, which period and who measured it — including us, and the honest answer for us today is that we do not have one.
What this evidence does and does not support
Supported: containment around 40% is achievable
A national tax agency published it on 4.8 million calls. That is a defensible planning assumption for a well-built government voice deployment, and it is far below what the industry advertises.
Supported: volume can grow without headcount
Air India's published position is that passengers doubled while contact-centre volume stayed flat. That is the shape of the argument — absorbing growth, not firing people.
Supported: the biggest gains go to the least experienced
The QJE study found +30% for novice agents and almost nothing for experts. In a sector with high attrition and constant retraining, that is where the value concentrates.
Not supported: a specific rupee saving per call in India
We could not find a single government, analyst or industry-body source for Indian cost-per-call or per-minute figures. Everything circulating traces to vendor pages. We will not publish a number we cannot source.
Not supported: 60-90% containment
Repeated constantly across vendor marketing with no primary behind it. The only government-published containment rate we found was 40%.
Not supported: that any of this will happen to you
Different callers, languages, systems and expectations. Published results from other organisations establish a category is real; they do not forecast your deployment.
Where a human call centre is still the right answer
An agent that cannot say "I do not know" is worse than a queue. Anything where a wrong answer carries real cost to the person on the line — a medical instruction, a legal deadline, an entitlement amount that a household will act on — needs a human in the path, and the honest design is an agent that recognises the boundary and routes across it rather than one that guesses confidently.
Distress calls are the clearest case. Tele-MANAS, India's mental-health helpline, has handled more than 32 lakh calls across 53 cells in 20 languages, staffed by people. Nothing in the evidence above suggests that work should be automated, and we would decline to build it.
Low volume is the other case. Below a few thousand conversations a month, the setup, language and compliance work does not amortise and a small human team is simply cheaper. The economics of an agent improve with scale, which is the same statement as: they are poor without it.
How to read any vendor's case study, including ours
- 1
Ask which organisation, by name
"A leading telecom operator" is not a citation. If the client cannot be named, the result cannot be checked, and an unverifiable result is marketing.
- 2
Ask who measured it and over what period
A vendor measuring its own deployment over a chosen window is the weakest form of evidence that still counts as evidence. A published company statement is stronger. A peer-reviewed study is stronger again.
- 3
Separate forecasts from outcomes
A great deal of what circulates as "AI saved X" is an analyst projection for a future year. Both are useful; conflating them is not.
- 4
Follow the number to its primary source
We traced the most-repeated statistics in this category and most ended at a vendor blog citing another vendor blog. Two ended at our own site, which is why they are not on this page.
- 5
Ask what the comparison baseline was
Against an unstaffed helpline, almost anything looks transformative. Against a well-run one, the honest gains are narrower and more specific.
What the published evidence actually shows
Nine findings, each confirmed by opening the publisher's own page. Ordered by how much weight the evidence can carry, not by how impressive the number looks.
Read this first. Every figure on this page was published by the organisation named beside it, about that organisation's own deployment or research. None of them is an AiSewak client, none of them used AiSewak, and none of these results is ours. They are here to show what has been demonstrated in the category — the honest basis for what a deployment might achieve, not a promise of what yours will.
- +15% / +30%Measured result
Access to a generative-AI conversational assistant increased customer-support issues resolved per hour by 15% on average across 5,172 agents — and by 30% for the least experienced. Novice agents with two months' tenure performed as well as untreated agents with more than six months'.
The strongest evidence on this page: peer-reviewed, 5,172 agents, staggered rollout. Read it carefully though — it measures AI assisting human agents, not replacing them, in an offshore contact centre where most agents were based in the Philippines. The paper also found little effect on already-skilled agents and a small quality dip for the most skilled.
Brynjolfsson, Li & Raymond · "Generative AI at Work", The Quarterly Journal of Economics 140(2), 889-942, doi:10.1093/qje/qjae044 (author-hosted copy of the published paper; the journal itself is paywalled) · 2025
- 40%Official figure
The US Internal Revenue Service reported that its voice bots handled more than 4.8 million calls, of which 40% were contained within the voice bot without escalating to a live assistor. Its chat bot resolved 42% of 450,000+ interactions without escalation.
The most credible government voice-bot containment figure available, because the agency published its own. Note how far below the industry's marketing it sits: vendors routinely claim 60–90% containment, a national tax authority measured 40%. Treat 40% as the realistic planning number and anything higher as a claim requiring proof.
Internal Revenue Service (US) · "Using voice and chat bots to improve the collection taxpayer experience" · 2022–2026
- 97%Measured result
Air India reported that its AI agent has resolved more than 13 million conversations at a 97% success rate, handling around 40,000 customer queries daily across more than 1,300 distinct question types. Its Chief Digital & Technology Officer stated that passenger numbers doubled since early 2022 while contact-centre call volume stayed flat at about 9,000 queries a day, saving "several million dollars a year".
The most India-relevant deployment with published numbers: a named Indian enterprise, a named executive, and both a deflection rate and a cost outcome. The caveat matters — "several million dollars a year" is the company's own estimate, not an audited figure, and the deployment is chat-led rather than pure voice.
Air India, via Microsoft Customer Stories · Air India customer story · 2024–2026
- 93 lakhOfficial figure
The Government of India stated that Kisan e-Mitra, a voice-based AI chatbot for PM-KISAN queries built on the Bhashini stack, had answered more than 93 lakh queries and was handling over 8,000 farmer queries daily in 11 regional languages.
India's flagship government AI voice deployment, with figures placed on the public record. One inconsistency worth knowing: an April 2025 Lok Sabha reply gave "over 20,000" queries daily against 92 lakh cumulative, while this December 2025 figure gives "over 8,000" daily against 93 lakh. The two government statements do not reconcile, so we cite the later one and date it rather than presenting either as a trend.
Press Information Bureau, Government of India · "Artificial Intelligence Transforming Indian Agriculture" backgrounder (figures as of December 2025) · 2026
- 500 → 1,600Official figure
The US Government Accountability Office documented that the National Passport Information Center tripled its call-agent staffing from roughly 500 in FY2020 to about 1,600 in FY2024, adding two sites and around 7,000 phone lines, to bring average caller wait times down from about 45 minutes to under one minute.
Included as the counterfactual rather than as evidence for AI: this is an audit body documenting what it costs to fix a government helpline by hiring. The report does not mention AI and GAO endorses nothing here. It is the honest baseline any automation business case should be measured against.
US Government Accountability Office · GAO-25-107409 · 2025
- 50%Measured result
Bank of America reported that its internal virtual assistant reduced calls into the IT service desk by 50%, is used by over 90% of employees, and that more than 98% of users of its customer-facing assistant find what they need — which the bank states significantly decreases call-centre volume.
First-party, from a named bank, and the 50% internal service-desk reduction is a hard number. The customer-facing claim is deliberately quoted rather than quantified: "significantly decreasing call center volume" is the bank's wording, and no percentage is attached to it.
Bank of America · "A decade of AI innovation" newsroom release · 2025
- up to 95%Analyst forecast
Gartner stated that labour expenses can represent up to 95% of contact-centre costs, and forecast that conversational AI deployments in contact centres would reduce agent labour costs by $80 billion by 2026, with roughly one in ten agent interactions automated.
The $80 billion is a 2022 projection, not an observed outcome, and should never be quoted as though it happened. The genuinely useful part is the structural point: if labour is up to 95% of what a contact centre costs, then anything that changes how many humans are needed dominates every other line item — which is also why telephony, not the AI, tends to become the biggest remaining cost.
Gartner · Press release, Stamford · 2022 (forecast for 2026)
- 10 languagesOfficial figure
MeitY announced that UIDAI is deploying Sarvam AI's models for voice-based resident interactions and fraud detection across 10 Indian languages, hosted on-premise inside an air-gapped UIDAI environment, with no data leaving that environment.
Cited for one purpose only: it establishes that a central Indian identity authority has accepted AI voice under strict data-sovereignty conditions. No outcome, savings or accuracy figure is published on that page, and none should be inferred from it.
Press Information Bureau / MeitY · PIB release on UIDAI–Sarvam AI · 2025
- 234m vs 175mMeasured result
A KPMG in India and Google study found Indian-language internet users had reached 234 million against 175 million English internet users, having grown at a 41% CAGR — and projected Indian-language users would be roughly 75% of India's internet base by 2021.
Nine years old, so treat it as establishing direction rather than current scale. The 234m versus 175m is measured to 2016; the 75%-by-2021 figure is a projection made in 2017 and is not a measurement of today. It remains the best-sourced statement of why vernacular reach is the deciding variable in India.
KPMG in India & Google · "Indian Languages — Defining India's Internet" · 2017 (data to 2016)
Call a live agent before you decide
These are running agents, not recordings. Open one, press call and speak to it in Hindi or English — the same stack that runs the deployments described above.
Yojana Didi — Hindi
We cannot show you a client outcome. We can let you judge the agent yourself, which is the more direct test.
Open the demo →Siya — open-domain questions
Try asking something the script cannot have anticipated, and watch what it does when it does not know.
Open the demo →Van Dhan — Santhali
A genuinely non-mainstream Indian language, which is where most platforms stop working.
Open the demo →Go deeper
AI vs traditional government call centres
Where the seated helpline still wins, and where it structurally cannot.
Why government helplines fail
The structural reasons a seated helpline loses citizens, and which of them automation actually fixes.
Human in the loop, done properly
Designing the escalation path rather than bolting it on after a failure.
Voice AI for government: ROI and cost-benefit
Building the business case, and which numbers to distrust while doing it.
Related pages: What it costs, and why we quote · Best AI voice agent in India · For government
Frequently asked questions
Does AiSewak have client case studies?
No, and we will not manufacture one. We have live demo agents you can call and deployments in progress, but no client deployment whose outcomes have been published and can be checked. This page cites other organisations' published results instead, clearly labelled as theirs. If that is a problem for your procurement process, it is a fair one — say so and we will tell you what we can substantiate.
Is an AI voice agent cheaper than a human call centre?
The published evidence points that way at scale, and the structural reason is that Gartner puts labour at up to 95% of contact-centre costs. But we could not find a single credible source for an Indian cost-per-call comparison — every figure circulating traces to a vendor page — so we do not publish one. What is verifiable is that the US IRS contained 40% of 4.8 million voice-bot calls without a human, and that Air India absorbed a doubling of passengers without growing contact-centre volume.
What containment rate should we plan for?
Around 40% is the only containment figure we found published by a government body measuring its own deployment, on 4.8 million calls. Vendor material routinely claims 60 to 90 percent with no primary source. Plan against 40%, treat anything higher as a claim that needs evidence, and remember containment depends far more on how narrow the use case is than on whose AI is behind it.
Will AI replace our call centre staff?
The best evidence available — a peer-reviewed study of 5,172 agents — measured AI making human agents more productive, not replacing them, with the largest gains going to the least experienced. Air India's published account is of absorbing doubled demand at flat headcount. Both describe capacity, not redundancy. Distress lines, high-stakes advice and anything where a wrong answer harms the caller should stay human, and we decline work that would automate them.
Is an AI voice agent legal for outreach in India?
Yes, within a well-defined set of rules. The agent must disclose that it is an AI at the start of the call (ECI's 2024 advisory on synthetic media), outbound calling must respect TRAI TCCCPR 2018 and DLT registration, personal data must be handled under the DPDP Act 2023, and IT Rules 2021 govern the content itself. Election deployments additionally require registration with the ECI or the state CEO as a political advertiser. AiSewak ships these controls switched on by default rather than as an add-on.
Why are there no McKinsey or Klarna figures on this page?
McKinsey's customer-care productivity figures are quoted everywhere, but their site would not serve us the pages, so nobody here has actually read them — unread means uncited. Klarna's 2024 AI assistant numbers are real and first-party, but the company reversed course in 2025 and began rehiring human agents; citing the launch without the reversal would be selective, and we did not verify the reversal against a primary source. Both are omissions on purpose.
Judge the agent, not the case study
Since we cannot show you a client result, the useful test is the direct one: bring your hardest objection and your actual dialect, and listen to how the agent handles both. That tells you more than any published percentage.
Every AiSewak agent identifies itself as an AI at the start of the call, never asks for an OTP or a payment, and honours DND. Election deployments require ECI / state CEO registration as a political advertiser.