Skip to content

Bananaflow MCP is live.

Read the docs

AI in Call Centers: What Changes for a BPO

Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.Jaseem Abbas

Published
Cover for "AI in Call Centers: What Changes for a BPO": a white isometric line diagram of a call center floor: two rows of three seats on a faint floor outline, the back row holding a lit puck, the AI agent, and two empty, faded seats, and the front row a person standing on each seat, beside a lit pill reading Seats for the calls people keep, drawn over a soft, out-of-focus salvia in muted sky blue, its slender petals blurred into veils, with an aqua haze at one end and a dusty rose glow at the other.

AI in a call center means software that answers or places some calls itself and helps people on the rest, so the floor needs fewer people on repeat calls and more on hard ones. This guide is for the person who runs the floor, not the brand that buys from it. It covers which calls move first, what happens to seats, shifts, QA, training and attrition, how client contracts are priced and where people go next.

Picture the head of operations at a 300-seat BPO in Noida. A retail client asks, "What is your AI plan for our line?" Her real question is smaller: which of my 300 seats does this touch first, and what do I tell the client about the price?

The Noida, Pune and Chennai floors in this guide are made up. They show how to think, not what a real BPO did.

What does AI in a call center actually do?

AI in a call center does two jobs. It takes whole calls itself, or it helps a person or a manager do their work. These are different products with different risks, and clients often mix them up in one request.

  • Job: AI agent takes the call

    What it looks like
    A voice agent places or answers a call, follows the client's script, books, confirms or answers, and hands to a person when it cannot finish
    What changes for the floor
    Fewer people needed on that queue; a new service line to price
  • Job: AI helps the person or the manager

    What it looks like
    Summaries, suggested replies and scoring of every recorded call. The industry calls the first two agent assist
    What changes for the floor
    Shorter wrap-up, and QA works on exceptions, not samples

Our guide to conversational AI explains how a voice agent works. How a brand should choose one is in our guide to customer care AI in India. This post stays on the operator's side.

How big is India's BPO industry, and why is AI a live question now?

India's business process management (BPM) industry is large, and its clients are now asking what AI does to the work they outsource. NASSCOM's Strategic Review 2025 press release reported India's tech industry revenue at over $282.6 billion in FY2025 (an estimate), of which BPM was $54.6 billion, and a tech workforce of 5.8 million. Its 2026 review expects revenue of $315 billion in FY26 and says BPM exports grew 7%.

Read those numbers with care. The $54.6 billion is NASSCOM's FY2025 estimate, from its press release on the Strategic Review, and NASSCOM's own BPM page labels the same figure FY24. The Strategic Review 2026 executive summary gives about 6 million direct jobs for the whole tech sector, not BPM alone. We found no current primary figure for BPM headcount, so we do not state one.

The government is watching the same shift. In June 2026 India's IT ministry opened a consultation on AI in IT services and said generative AI is gaining traction across IT and BPM processes, and the consultation named large-scale upskilling and reskilling of the IT and BPM workforce.

For an operator, the pressure shows up in client reviews: what does your AI do on our line, how do you score it, and how will you price it?

How is AI used in a BPO?

AI is used in a BPO on the calls themselves (reminders, confirmations, first answer, status questions, lead callbacks), on quality (scoring and reports across all calls, not a sample) and in planning (which queues are spiky, and how many seats the people-only calls need). The rest of this guide follows from those three. Our industry page for BPOs shows the product side.

Which call types move to AI first?

The calls that move first are the ones that repeat, can be answered from data or a document, are cheap to fix if wrong and arrive in spikes. This is our own test, not a published standard. Score each client queue on four questions:

  1. Does one script cover about 80% of the calls?
  2. Is the answer in data or a document? A delivery status, an EMI date, a calendar slot.
  3. Is a wrong answer cheap to fix? A wrong reminder time costs a callback. A wrong refusal of a refund costs a customer.
  4. Is the volume spiky? A queue that doubles on Mondays is where seats sit idle the rest of the week.

High on all four: move it first. Low on the third or fourth: AI helps the person instead of replacing the call.

A BPO's call types sorted by what moves to AI first (high-volume, routine, rule-bound calls) and what stays with people (complex, emotional, judgement calls), with a handoff path between them.

What is the usual order on an Indian floor?

The usual order, from our view of the work, is this:

  1. Outbound reminders and confirmations. These follow a fixed script. See our guides to COD order confirmation calls, doctor appointment reminder calls, insurance renewal reminder calls and payment reminder calls.
  2. Overflow and after-peak first answer. Calls that would wait or ring out at night get a first answer and a callback.
  3. Lead callbacks and first qualification. New leads need a call in minutes, as our guide to speed to lead explains. The first call checks fit with a checklist such as BANT.
  4. Status questions. Where is my order, when is the technician coming.
  5. Stays with people: complaints and escalations, collections disputes, retention and save desks, distress calls, and anything with a regulator's script that needs judgement.

The fifth group matters most for a collections process like the one in Pune. A reminder before a due date can follow a script. A call with a borrower who disputes a charge or is in trouble should not. The buyer's side of the same fit question is in the fit table of the customer care AI guide.

What about a Tamil and Telugu support process?

Add a fifth question for an Indian floor: does the AI handle your callers' language and accent? A Chennai process serving Tamil and Telugu callers is not an English process, and a Hindi demo tells you nothing about either. Test each language with your own recordings before you move volume, as our guide to testing a voice agent explains.

What happens to seats and shifts?

Seats and shifts should be sized to the calls people keep, not to the old peak. That is the clearest change on the floor, and it arrives queue by queue.

  • Lever: Seats and shifts

    Before
    Seats sized to the peak; night and Sunday shifts staffed for thin volume
    After
    Seats sized to the calls people keep; AI takes the peaks and the thin hours where the client allows it
  • Lever: QA

    Before
    A QA analyst samples a few calls per agent a month
    After
    Every AI call scored; people's calls still sampled; QA reviews the lowest scores
  • Lever: Training and attrition

    Before
    New hires learn the repeat calls first; repetitive queues drive people out
    After
    New hires start on harder calls sooner; the effect on attrition is not proven in India
  • Lever: Contract and pricing

    Before
    Per seat (FTE) or per hour
    After
    A mix: per seat for people, per minute or per call for AI, an outcome fee on some lines

Take the Noida floor. The retail client's night shift has 20 people answering order-status calls that arrive in thin bursts. After a move, AI gives the first answer at night and logs a callback for anything it cannot finish. The night shift shrinks, but not to zero, because escalations still need someone who can decide.

Three practical points:

  • Move a slice, not a queue. Keep the old staffing for a trial window and compare. Do not cut seats on a forecast.
  • Plan the handoff as a seat. Whoever takes a handed-over call needs to see what the AI collected. See warm transfer vs cold transfer.
  • Re-read average handle time. The calls left to people are harder, so their handle time rises even when the floor is healthier. See AHT in a call center, and the telecaller cost guide for the cost maths of a seat against a minute.

How does QA change when AI takes calls?

QA changes from sampling a few calls per person to scoring every AI call and then reviewing the worst ones. A scoring step on every call can return tags, a sentiment and a summary, and the analyst starts from the lowest scores. Bananaflow's call QA works this way. Our guide to auditing calls in a BPO covers sampling and rubrics for human calls.

What does not go away:

  • Calibration. A scorer is only as good as its rubric. Have your QA lead score the same set of calls and compare, every few weeks.
  • Sampling of people's calls. They are the hard calls, and the ones clients escalate over.
  • A client-facing view. Write the rubric and the pass mark down, because the client will ask.

What happens to training and attrition?

The best evidence we found says AI that assists a person can help new hires ramp up faster. It does not prove that AI fixes attrition in India. Be careful with both claims.

The study is "Generative AI at Work", by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, a working paper from the National Bureau of Economic Research. It looked at 5,179 customer support agents at one firm and found that an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal gains for the most experienced. It also reported better customer sentiment and higher employee retention.

Know its limits before you quote it:

  • It covers one firm, in chat support, not voice.
  • The AI assisted the person. It did not take the call.
  • We did not check where the agents sat, so we do not claim it applies to India.

For attrition, we found no current India figure in a primary source. Job sites and vendor blogs quote numbers without method, so we do not repeat them. Use your own exit data by queue, and read our guide to call center metrics for how attrition and occupancy are counted. The cost of a monthly exit is part of the telecaller cost guide.

Our hypothesis: repetitive queues are where people tire fastest, so taking some of those calls off people may help retention. Test it by comparing exits on a queue where AI took the repeat calls with a similar queue where it did not.

For training, the first week changes. New hires used to start on the easiest repeat calls, which AI may now handle. Rework the first month in your call center training manual so it does not teach calls the floor no longer takes, and use good AI call transcripts as examples.

How do client contracts and pricing change?

Client contracts change because a seat is no longer a good unit when AI does part of the work. If a client pays per seat and the BPO uses AI to do the same calls with fewer seats, the BPO loses revenue for being efficient. That is why the pricing model is the first conversation to have, not the last.

NASSCOM's Strategic Review 2026 describes the shift directly. In its words, providers "re-engineered revenue models, moving away from FTE delivery to outcome-based, risk-sharing constructs as AI-driven productivity materialised". Our view of the models:

  • Model: Per seat (FTE)

    Client pays for
    A staffed seat for a shift
    Who carries the productivity risk
    The client, who pays for idle time
    Fits
    Steady, complex queues
  • Model: Per hour or per minute

    Client pays for
    Talk or login time
    Who carries the productivity risk
    Shared
    Fits
    Variable volume; the natural unit for AI minutes
  • Model: Per transaction or per call

    Client pays for
    Each handled call
    Who carries the productivity risk
    The BPO
    Fits
    Repeatable calls with a clear end
  • Model: Per outcome

    Client pays for
    A result: a booking, a confirmed order, a resolved ticket, a payment promise
    Who carries the productivity risk
    The BPO
    Fits
    Lines where the outcome is measurable and the BPO controls most of it
  • Model: Hybrid

    Client pays for
    Base seats plus a per-call or outcome fee
    Who carries the productivity risk
    Shared
    Fits
    The common transition step
Three ways a client pays a BPO: per seat (FTE), per minute and per outcome, and what each one rewards.

What does each model reward?

  • Per seat rewards headcount. More seats staffed means more revenue, so AI works against the BPO. The client pays for idle seats.
  • Per minute rewards time. It suits AI, since an AI call has a clear length, but it can reward slow calls unless a quality measure sits in the contract.
  • Per outcome rewards the result. The BPO earns when the order is confirmed and carries the risk when it is not. It fits poorly where the client's offer or lead quality decides the result.

A sensible path is a hybrid first. For the retail client in Noida, a base fee covers the people who stay on escalations and a per-confirmed-order fee covers the AI's routine work. Agree the unit, how an outcome is counted and who audits it in writing. For the cost maths, see the telecaller cost guide and cost per call in India. We keep rupee figures out of this post on purpose.

Bananaflow bills by the minute used, so any outcome fee is a contract between you and your client.

Is AI replacing BPO jobs?

AI is replacing some repeat calls. It is not removing the need for people, and the forecasts that exist say both things at once. We found no credible figure for how many BPO jobs change, so we do not give one.

Two forecasts from Gartner, an analyst firm, are worth reading together. Both are predictions and Gartner states them as such.

  • In March 2025, Gartner predicted that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention, with a 30% reduction in operational costs. The word "common" matters.
  • In January 2026, Gartner predicted that by 2030 the cost per resolution of generative AI will exceed $3, higher than many offshore human agents' cost, that full automation will be prohibitively expensive for most organisations, and that firms will have to maintain or even rehire human agents because of rules on the right to speak to a person.

What an operator can take from the two: routine volume will be automated over time, doing so is not free, and people stay for calls that need judgement, empathy or a rule. The second forecast compares AI with offshore human agents, the model many Indian BPOs run.

Where do people move?

In our view, people move to work the AI creates or cannot do. For what this means for the person in the seat, see our guide to the voice process in a BPO. We cannot say how many.

  • Escalation desk. Handed-over calls are harder and need someone who can decide.
  • QA and calibration for AI calls. Someone sets the rubric and reviews the lowest scores.
  • Writing and testing agent scripts. Your best trainers know what a good call sounds like. See call flow design.
  • Language QA. A native speaker hears that a Tamil or Telugu phrase is wrong or rude.
  • Client success and workforce planning. Reading reports with the client, and forecasting people and AI together.

MeitY's consultation, cited above, named reskilling as a goal. Tell your team early which queues move first and what paths exist.

What does not change?

The rules, the brand voice and the need to measure results all stay.

  • Consent and calling rules. The same rules apply to an AI call as to a person's. See TRAI rules for AI calls, DPDP rules for call centers and is call recording legal in India.
  • The client's brand voice. The AI speaks for the client, so script and tone come from them.
  • A person on complaints and escalations.
  • Measuring resolution, not answered calls. An AI that answers everything and resolves little is a worse floor with a better dashboard.
  • Carrier and number costs, which sit outside the AI.
  • The client's audit rights over recordings.

This is not legal advice. Check with counsel before you change who handles a client's calls.

How should a BPO start?

A BPO should start with one client queue that passes the four-question test, run AI beside people on a slice, score every call and agree the pricing unit before scaling. A small, measured start beats a floor-wide plan.

  1. Pick the queue. One with a repeatable script and a low cost of error, such as order confirmation.
  2. Write the success measure first. Confirmed orders, bookings or tickets resolved, plus the handoff rate and complaints about the AI.
  3. Test before a real call. Run it by voice and text with recordings of your own callers' accents. See testing a voice agent.
  4. Set the limits. Decide what the AI must never say and when it hands over. See voice agent guardrails.
  5. Run a slice. Send a small share of the queue to AI, keep people on the rest and compare results, QA scores and complaints.
  6. Agree the pricing unit with the client before volume grows.
  7. Then add queues, one at a time, with the same checks.

How Bananaflow fits

Bananaflow is an AI voice agent platform. You build one agent per client, each with its own draft, published versions and version history. Campaigns run from a CSV with retries, calling time slots, concurrency limits and an A/B split between agent versions. You bring your own numbers and carrier, including VICIdial and Asterisk. For how contact center platforms are sold as a service, see the glossary entry on CCaaS.

After each call you get a transcript, a recording, extracted variables and an outcome code, and a scoring step returns tags, sentiment and a summary. A call can be transferred to a person on Twilio, Plivo, Telnyx, Cloudonix and Asterisk. The agent speaks English and 10 Indian languages and can switch language during a call. Recordings are stored in India, and call data is not used to train models.

Lists, consent and TRAI, DLT and DPDP duties stay with your team, so check numbers against the DND registry before you call. See the BPO page for the product.

What should you remember about AI in call centers?

AI in call centers moves the repeat calls and leaves the hard ones, so plan around that split. Size seats to the calls people keep, score every AI call, train new hires on the calls the floor still takes and settle the pricing unit with the client before volume moves. Tell your team where people go. Keep the rules, the brand voice and a person on escalations where they were.

FAQ

Frequently asked questions

  • Is AI replacing BPO jobs?

    It is taking over some repeat calls, not the need for people. Gartner forecasts that by 2029 agentic AI will resolve 80% of common customer service issues, and also forecasts that by 2030 the cost per AI resolution will exceed that of many offshore human agents, so firms will have to keep or rehire people. Both are predictions. For an operator, the practical change is fewer people on routine queues and more on escalations, QA and script work.

  • How is AI used in a BPO?

    In two ways. AI agents take whole calls on routine queues such as reminders, confirmations, status questions and after-hours first answer. AI also helps people and managers: summaries after a call, scoring of every call for QA and live prompts to a person on the line, which the industry calls agent assist. The two can run on the same client account.

  • Which calls should a call center automate first?

    Start with calls that repeat, can be answered from data or a document, are cheap to fix if wrong, and arrive in spikes. In practice that means outbound reminders and confirmations, overflow and night first answer, lead callbacks and status questions. Complaints, disputes, retention calls and anything a regulator scripts stay with people. This test is our own framework, not a published standard.

  • What is outcome-based pricing in BPO?

    The client pays for a result, such as a confirmed order, a booked visit or a payment promise, instead of a staffed seat or a call minute. The BPO carries the risk of not reaching the result. NASSCOM's 2026 Strategic Review says providers are moving from FTE delivery to outcome-based, risk-sharing constructs as AI-driven productivity grows.

  • Does AI reduce attrition in a call center?

    The one strong study we found says an AI assistant increased employee retention at a single US support firm, and we found no Indian figure. Taking the most repetitive queues off people may help, but that is a hypothesis for your floor, not a proven result. Measure your own exit data by queue before and after, and compare.

  • Can AI handle Hindi and other Indian languages on calls?

    Some platforms can. Bananaflow supports English and 10 Indian languages (Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil and Telugu) and can switch language during a call. Test every language and accent you serve with your own recordings before you move a queue, and keep a person ready for calls the agent cannot finish.

  • How does QA change when AI takes calls?

    Every AI call can be scored, so QA no longer relies on a small sample for those calls. The QA team spends its time on the lowest scores, on checking that the scoring matches human judgement, and on the calls people still handle, which are still sampled. Calibration between the scorer and your QA lead stays necessary.

Written by

  • Jaseem Abbas, founder of Bananaflow, in glasses and a dark top against a pale background.

    Jaseem Abbas

    Founder, Bananaflow

    Jaseem founded Bananaflow to build voice AI that understands how people in India really speak, including the switch between languages in the middle of a call.

Similar articles

Bring us one process. We'll pilot it on your real calls, with your team as the fallback.

Plan a pilot