A 100-agent floor at Rs 31,500 a seat costs about Rs 31.5 lakh a month in people before telephony. Run it as a Human + AI floor and by month six the AI finishes 70 to 80% of calls; the people left are the ones on the hard calls plus a few overseeing the AI lane, and the AI's minutes arrive on the bill in their place at Rs 2.50 falling to Rs 1.75 a minute. Whether that's cheaper comes down to cost per outcome at month one, three and six, which rarely makes the first slide.
The numbers I'll use
All of these are public or on our pricing page, so you can check them. A standard English or Hindi agent seat in India costs Rs 28,000 to 35,000 a month loaded; I'll use Rs 31,500. Our AI is Rs 2.50 a minute at the pilot rate and steps down with committed volume to Rs 1.75 at the largest slab. The handover between the AI and a person is not charged. Setup is priced once per integration and sits on the quote. On a typical programme the AI finishes up to 40% of calls by month 3, 50% by month 4 and 70 to 80% by month 6. I'll keep the arithmetic rough on purpose; the ROI calculator does it precisely with your numbers.
Month one
The floor is all still there: 100 seats, Rs 31.5 lakh. The AI is taking a small share while it learns the book, say a fifth of the calls, and its minutes are on the bill at the pilot rate. So month one costs more than the floor did, which is why the pilot carries no commitment, and why anyone who shows you a saving in week one is showing you the bot's containment and calling it the floor.
Month six
By month six the AI is finishing three quarters of the calls. The people left on the queue are the ones taking the handovers plus a handful overseeing the AI lane; on a 100-seat floor that tends to be somewhere between a fifth and a third of the original bench, depending on how hard the queue is. The minutes the AI handled are on the bill. On the queues we run, the total at month six lands below the old people bill, and cost per outcome lands further below it, because the AI also reached the calls nobody was getting to. I won't put a percentage on that here; the calculator does it with your volume, your agent cost and the published rate, and the answer moves a lot with the queue.
The three places the maths goes wrong
Comparing per-minute to per-seat is the usual one. A seat does a number of outcomes a month; a minute does a fraction of one. Divide the whole bill by outcomes or you are comparing a car to a litre of petrol. Then there's counting the bot's containment as the floor's result. If the bot kept 85% of calls and dropped the rest onto your floor, your floor's handle time went up and your renewal rate did what it did; the 85% is the bot's number, not yours. And people forget setup, and the hours their own team spends reading transcripts in the first weeks. Setup is a one-time line; the calibration weeks are a real cost in your team's hours and the biggest lever on how fast the ramp climbs.
What I'd ask any vendor, including us
Three columns: cost per outcome at month one, three and six, with the AI share they are assuming in each. If they won't give you the month-one column, they're hiding the ramp. If the month-six share is above 80%, ask which calls they think need a person and where those go. And ask whether the handover is billed; some vendors charge the transfer leg as minutes on both sides. Put your own numbers in the calculator first so you know what the answer should look like before you hear theirs.