“What does it cost?” is the first serious question in every AI agent conversation, and it deserves a better answer than “it depends”. So here are the actual shapes: what vendors in India charge, what building it yourself really costs once you count everything, and the arithmetic that tells you whether the whole thing pays back in months or never. The cost-per-ticket groundwork from an earlier post feeds straight into this — if you know what a ticket costs you, every number below gets easy.
One framing before the numbers. An AI agent is not a software licence; it is a replacement for work you are already paying humans to do, done worse in some places and better in others. So the honest unit of comparison is not the fee against your budget — it is the fee against the ledger it touches.
The four pricing models
Indian vendors and automation firms mostly price one of four ways, and the ranges below are what mid-market D2C brands commonly see in 2026:
- Per resolution: ₹8-25. You pay when a ticket closes without a human. Cleanest alignment, easiest ROI math, and the number to negotiate hardest — definitions matter (what counts as resolved, what happens to reopened tickets).
- Per order: ₹1-4. Common for logistics-heavy flows that touch every order rather than every ticket — confirmation, address verification, tracking updates. Priced against freight economics, so a ₹2 fee against a ₹150-₹250 RTO it prevents is not a close call.
- Monthly seat or platform fee: ₹15,000-60,000 per month at typical mid-market scope. Predictable, budget-friendly, and misaligned at scale — you pay the same in the month it answers 400 tickets and the month it answers 4,000.
- Outcome share: a percentage of measured savings. The most aligned and the rarest — it requires agreed measurement, which requires trust, which requires a diagnostic first. Where you can get it, take it.
Beware the hybrid that quietly stacks: a platform fee plus per-resolution plus per-message rates plus integration costs. Any one model is defensible; three at once deserves a spreadsheet and a hard question.
What building actually costs
The build option is real and sometimes right, so it deserves honest accounting rather than a scarecrow. The line items that surprise teams:
- The model bill is the small part. API costs for a mid-size brand’s ticket volume are typically thousands of rupees a month, not lakhs. Fine.
- Prompts, evals and guardrails are the real work. A WISMO answerer that reads courier scans across four carriers, composes a dated reply and stays inside policy is weeks of prompt engineering plus a permanent evaluation harness — every model update and courier format change re-opens it.
- Someone is on call. The failure modes are silent: a broken webhook means the agent confidently answers from stale data. Detection, alerting and a human who owns it — forever.
- Integration and maintenance drift. Your OMS changes, WhatsApp template rules change, courier APIs change. The build is not a project; it is a pet.
Fully loaded, a competent in-house build for one workflow lands around ₹8-15 lakh and 3-5 months to production-grade trust, then ongoing maintenance at a fraction of a senior engineer. That is a sound investment when the workflow is core, stable and differentiating. For “answer refund status queries”, it is usually an expensive way to learn what a vendor already ships.
The crossover
Build beats buy when volume is huge, workflows are stable, and the capability is strategic. Buy beats build when you are proving value on messy, evolving processes — which describes nearly every brand’s first three automation projects. At India’s mid-market ticket volumes (1,000-5,000 tickets a month), the buy curve starts lower and stays lower for the first 18 months or so; the build curve starts with a lump and flattens slowly. Most teams that build first spend their learning budget on infrastructure instead of on earning trust in the workflow — which was the actual scarce resource.
The payback worksheet
Take a mid-size D2C brand: 9,000 orders a month, ~2,400 tickets at festive peak, loaded cost ₹65 per ticket (the four-layer math is in the support cost post). The current support ledger is about ₹1.56 lakh a month. Suppose automation covers 62% of volume — the routine, answerable categories: WISMO with clean scans, refund status, COD confirmations — at a blended cost including human review time of roughly ₹0.9 lakh a month. Net saving: ₹0.66 lakh a month. Against a setup and integration effort priced in lakhs, payback lands at 3-5 months, and every month after compounds.
Now the honesty checks that make the worksheet trustworthy. Did you include the human review time in the agent column? (The guardrails approach keeps humans in the loop by design — cost them.) Did you assume zero attrition savings in year one? (You should; support teams shrink by choice, slowly.) Did you count the second-order wins — RTO prevented by better confirmation, refunds disputed less, festive season survived without emergency hiring? (Count them as upside, not base.) A worksheet without those three conservative adjustments is a pitch deck, not a decision.
When the answer is “not yet”
Three honest disqualifiers. If your ticket volume is under ~600 a month, the absolute savings rarely justify any fee or build — fix the product and the process, revisit later. If your data is a mess — no order IDs on tickets, no scan history access, processes that live in one person’s head — the agent inherits the mess at machine speed; diagnose first. And if nobody internal will own the automation — review its drafts, set its limits, read its weekly report — you are not buying an agent, you are buying a liability with a demo. The MSME guide covers the smaller-scale version of the same decision.
The bottom line
What an AI agent costs in India: ₹8-25 per resolution or ₹1-4 per order on the common models, ₹15-60k a month on seats, or a share of savings if you can get it. What it pays back: 3-5 months at mid-market volume when the ticket ledger stings and the data is ready, and never when neither is true. Any vendor who cannot help you compute the second number — including us — has not earned the first one.