Two years ago the enquiry was "we want a chatbot". Now it is "we want an AI agent" — something that does not just answer, but looks things up, calls your systems and finishes a task. The question that follows is always the same, and it deserves a straight answer rather than a discovery call.
So: here are the ranges we quote in 2026, what sits inside each one, and — the part almost nobody puts in a proposal — what it costs to keep running every month after you have paid for it.
What actually counts as an "AI agent"?
The word has been stretched to cover everything from a prompt with a nice UI to a fleet of coordinating processes. Since the price follows the definition, it is worth fixing the definition first. We scope three tiers, and almost every enquiry lands in one of them.
| Tier | What it does | Typical build |
|---|---|---|
| Assisted | One task, one tool, no autonomy. Answers from your documents, drafts a reply, classifies a ticket. | ₹1.5L – ₹4L |
| Agentic | Plans across several tools, reads and writes to your systems, retries when a step fails. | ₹4L – ₹12L |
| Multi-agent | Several specialised agents with a supervisor, a queue, and a human review step before anything commits. | ₹12L – ₹35L |
What is inside a build quote?
A tier-2 agent at, say, ₹7L is not seven lakh of prompt engineering. Roughly, on a 6–10 week build, the effort distributes like this — and the two lines people try to cut are the two that decide whether it works.
- ~30%
- Integration: auth, APIs, and the data that turns out to be messier than described
- ~25%
- Agent logic: tools, planning, retries, guardrails, failure paths
- ~25%
- Evaluation and observability: test sets, traces, regression runs
- ~20%
- Interface, deployment, handover and documentation
The running cost nobody puts in the proposal
This is the part to model before you sign anything. Do the arithmetic yourself — it takes five minutes and it changes decisions. Take your expected volume, multiply by tokens per interaction, multiply by your model's price.
Assume a support agent with retrieval:
conversations / month 8,000
LLM calls per conversation 3 (route, retrieve+answer, verify)
input tokens per call ~3,500 (system + retrieved context)
output tokens per call ~400
input = 8,000 x 3 x 3,500 = 84,000,000 tokens
output = 8,000 x 3 x 400 = 9,600,000 tokens
At a mid-tier model around $1.00 / $5.00 per million tokens:
input 84.0M x $1.00 = $ 84
output 9.6M x $5.00 = $ 48
------
$ 132/month (~₹11,500)
Now re-run it with a frontier model at $5 / $25 per million:
$ 660/month (~₹57,000) for exactly the same product.Two things fall out of that arithmetic. First, the model you pick matters more than almost any code you write. Second, tokens are usually not your biggest monthly line — the infrastructure and the people around the agent are.
| Line item | Typical monthly range | Notes |
|---|---|---|
| LLM tokens | ₹8,000 – ₹60,000 | Swings 5× on model choice alone |
| Vector store / search | ₹2,000 – ₹15,000 | Often free at small corpus sizes |
| Hosting and queues | ₹3,000 – ₹20,000 | Agents are long-running; serverless gets awkward |
| Observability and evals | ₹4,000 – ₹25,000 | Tracing, stored runs, scheduled regression suites |
| Human review | ₹0 – ₹80,000 | The real cost, and the one always left out |
What makes the number go up?
- The state of your data. An agent that reads a clean Postgres schema is a different project from one that reads twelve years of inconsistently-named spreadsheets. This is the biggest single swing factor and it is never about AI.
- Write access. Reading is cheap. The moment an agent can create an invoice or refund a payment, you need approvals, audit trails and rollback — and the cost roughly doubles.
- Compliance. Health, financial or personal data brings data-residency, retention and access-logging requirements that are engineering work, not paperwork.
- Languages. Hindi and regional-language support is usually fine for understanding and considerably harder to evaluate. Budget for the test sets, not the translation.
- Latency targets. "Feels instant" on a voice channel is a different architecture from a batch agent that has 30 seconds. Sub-second forces smaller models, caching and streaming, and that is design work.
Should you build it or buy it?
Honest answer: if an off-the-shelf tool does 80% of what you want, buy it. The case for building is not capability, it is the last mile — the part specific to your business that no vendor will ship for you.
| Buy a platform | Build your own | |
|---|---|---|
| Time to first value | 1–4 weeks | 6–12 weeks |
| Year-one cost | Subscription, scales with seats or volume | Build cost + run cost, flat after launch |
| Fits your workflow exactly | ||
| You own the data and the logic | ||
| Switching cost later | High | Low |
“Buy to find out whether the workflow is worth automating. Build once you know it is, and know exactly what "it" means.”
Frequently asked questions
How much does it cost to build an AI agent in India?
A scoped single-task agent typically costs ₹1,50,000 to ₹4,00,000. A multi-tool agent integrated with your internal systems is ₹4,00,000 to ₹12,00,000. Multi-agent workflows with human review start around ₹12,00,000. Running cost is separate and usually ₹8,000 to ₹1,20,000 a month.
How long does it take to build one?
Four to six weeks for a tier-1 assisted agent, six to ten for a tier-2 agentic build, and three months or more for multi-agent systems. The variable is almost never the AI work — it is how long it takes to get clean access to your data.
Is it cheaper to use an open-source model?
Sometimes, at high and steady volume. Below roughly a few million tokens a day, self-hosting costs more once you count GPU time, ops and the engineer maintaining it. Start on a hosted API, measure for a quarter, and move only if the arithmetic says so.
What is the ongoing maintenance?
Budget 15–20% of the build cost per year. Models get deprecated, your own APIs change, and prompts that worked on last year's model need re-validating. An agent is a system you operate, not a deliverable you receive.
Can an agent replace a support team?
It can absorb the repetitive half and hand the rest over cleanly, which is a real and measurable saving. Teams that aim for full replacement generally end up paying twice — once for the agent and once for the people cleaning up after it.
The short version
Pick one workflow you can measure. Scope the smallest agent that changes it. Insist that evaluation and observability are in the quote. Model the monthly run cost before you sign, using your own volumes rather than ours. If the numbers still work, build it — and if they do not, you have saved yourself ₹7L and a difficult quarter.
If you want a second opinion on a scope or a quote you have already been given, send it over — we will tell you what we think, including when the answer is that you do not need us.