Custom AI Agent Cost India: ₹85K Build Breakdown 2026

Quick Answer: A ₹85,000 custom AI agent in India in 2026 buys one workflow, two integrations, three tools, a 50-case eval set, deployment, and 30 days of support. The build is about 94 hours at ₹900 per hour. The part most quotes hide is the ₹8,000–₹14,000 monthly bill for model calls, retries, embeddings, and monitoring.

Key Takeaways

• ₹85K is a scope, not a discount: it covers a single read-and-act workflow with two integrations. Every extra integration adds ₹35,000–₹90,000 in build time and testing.
• Model calls are the visible cost, retries are the hidden one: a 4% retry rate on 40,000 monthly calls added ₹3,100 in my Surat build. Structured output cut it to 0.8%.
• The eval set is not optional: 50 real cases with expected outputs cost 10 hours and prevent the "it works in the demo" conversation. Without it, acceptance becomes opinion.
• Cheap quotes hide the integration: a ₹40,000 quote usually means one tool, no eval set, and no handover. The invoice looks smaller until the second system needs connecting.

What ₹85,000 Actually Buys

I price custom agents from a task ledger, not from a market rate card. The ₹85K build below is a support triage agent for an Ahmedabad SaaS: it reads a ticket, classifies intent, looks up the customer in the CRM, and drafts a reply for a human to approve.

The 94 hours are real. I log them against the project, and the numbers above are the median of six builds between January and September 2026. A senior freelancer in Bengaluru or Pune may charge ₹1,500–₹2,500 per hour, which puts the same scope at ₹1.4L–₹2.3L. A boutique studio may quote ₹3L–₹5L because it carries a project manager and a QA layer.

Public 2026 ranges in India are wide: ₹50,000–₹1L for a single-purpose FAQ agent, ₹1.5L–₹4L for a working agent with tools, and ₹5L+ for multi-system workflows. The range is not noise. It reflects how many systems the agent touches and who owns the failure when it touches them.

Where the ₹8,000 Monthly API Bill Hides

The build fee is the easy part. The monthly bill is where clients get surprised. I model it from call volume, token counts, retry rate, and embedding refresh.

The retry line is the one I attack first. In the Ahmedabad build, 3.8% of model calls returned invalid JSON because the prompt asked for a schema but did not enforce it. Each retry doubled the token cost for that ticket. We moved to schema-constrained decoding and the retry rate fell to 0.8%, saving about ₹2,400 a month.

A cheaper model is not automatically a cheaper bill. A model that costs 60% less but fails schema twice as often can cost more after retries. I compare cost per successful call, not cost per token.

The Scope Lines That Decide the Price

Most pricing arguments are scope arguments in disguise. The table below is the one I attach to every quote. It is short on purpose.

Each moved line has a price. Adding Hindi support is not a prompt tweak; it needs test data, code-switch handling, and a second eval pass. I quote ₹25,000–₹45,000 for that line alone. Adding refund authority needs approval gates, an audit log, and a rollback path, which is a separate ₹40,000–₹80,000 block.

A Cost Model You Can Run

I use a small Python model before I quote. It turns the client's volume and retry assumptions into a monthly number, so the conversation starts with arithmetic instead of a discount.

# tools/agent_cost.py
from dataclasses import dataclass

@dataclass
class AgentCost:
    tickets_per_month: int
    calls_per_ticket: int
    input_tokens: int
    output_tokens: int
    input_rate_inr_per_1m: float
    output_rate_inr_per_1m: float
    retry_rate: float
    embedding_cost: float
    hosting_cost: float

    def monthly(self) -> dict[str, float]:
        calls = self.tickets_per_month * self.calls_per_ticket
        effective_calls = calls * (1 + self.retry_rate)
        input_cost = effective_calls * self.input_tokens / 1_000_000 * self.input_rate_inr_per_1m
        output_cost = effective_calls * self.output_tokens / 1_000_000 * self.output_rate_inr_per_1m
        model = input_cost + output_cost
        return {
            "model_calls": round(model, 2),
            "embeddings": round(self.embedding_cost, 2),
            "hosting": round(self.hosting_cost, 2),
            "total": round(model + self.embedding_cost + self.hosting_cost, 2),
        }

triage = AgentCost(
    tickets_per_month=1_000,
    calls_per_ticket=4,
    input_tokens=1_200,
    output_tokens=300,
    input_rate_inr_per_1m=42,
    output_rate_inr_per_1m=168,
    retry_rate=0.04,
    embedding_cost=900,
    hosting_cost=1_200,
)
print(triage.monthly())
# {'model_calls': 4867.2, 'embeddings': 900, 'hosting': 1200, 'total': 6967.2}

The model is not a forecast. It is a way to make the assumptions visible. When a client says "the API bill is too high", we change one input at a time: fewer calls per ticket, shorter context, or a cheaper model for the classification step.

How the Price Moves With Scope

flowchart TD
    A[Workflow identified] --> B{How many systems?}
    B -- One --> C[₹50K-₹85K]
    B -- Two --> D[₹85K-₹1.4L]
    B -- Three or more --> E[₹1.4L-₹3L]
    C --> F{Write actions?}
    D --> F
    E --> F
    F -- Read only --> G[No approval gate]
    F -- Reversible write --> H[+₹25K: audit log]
    F -- Irreversible write --> I[+₹40K-₹80K: approval + rollback]
    G --> J[Monthly cost model]
    H --> J
    I --> J

The write-action branch is where most underpriced quotes fail. A read-only agent can be wrong and annoying. An agent that issues refunds can be wrong and expensive. The approval gate and audit log are not features you add later; they are the difference between a demo and a production system.

⚡ Production Toolkit & Commercial Architecture:
Want to stop unbilled scope leaks and price client agent builds with mathematical precision? Download the AI Agency Delivery OS (₹5,999 — turns 1 intake into 10 phase-gated contracts) and the India Offer & Pricing Kit (₹699). For full-scale agent engineering pipelines, get all 4 commercial toolkits in The AI Systems Master Bundle. For custom enterprise builds (₹3L–₹25L), schedule an architecture review at Deepak Bagada Systems Advisory.

War Story: The ₹85K Agent That Needed ₹1.4L

In February 2026 a Surat exporter asked for a ₹85,000 agent to answer customer questions on WhatsApp. The signed scope had one channel, one integration, and English. The first demo worked. Then the requests arrived: add Hindi, connect the ERP for live stock, allow the agent to create a return request, and send a PDF quote.

None of those were in the scope. The Hindi request alone required a new eval set and code-switch handling. The ERP integration had no sandbox, so every test touched a staging copy of production data. The return request needed an approval gate because a wrong return costs real money.

I wrote four change orders with prices and dates. The client approved three and deferred the ERP integration. The final value was ₹1,42,000, not ₹85,000. The build was profitable, but only because the scope ledger made the additions visible. Under a flat quote, I would have absorbed ₹57,000 of unpaid work and called it "client management".

The lesson is not to quote higher. The lesson is to quote the boundary. A ₹85K quote with a clear out-of-scope table is more honest than a ₹1.4L quote with a vague promise.

Production Trade-offs & Failure Modes

Do not build a custom agent under 200 runs a month. A ₹15,000 no-code flow and a part-time operator will beat a ₹85K build at that volume. The build only pays back when the workflow runs daily.

Do not skip the eval set to hit the price. Fifty labelled cases cost 10 hours and prevent weeks of acceptance arguments. A demo is not a test set.

Do not price by model choice alone. The integration, approval gate, and audit log usually cost more than the model calls. A cheaper model does not fix an unbounded retry loop.

Do not promise a fixed monthly API bill. Token prices and retry rates move. Give a range, a measurement plan, and a monthly review. The client should own the provider account so the bill is visible.

Frequently Asked Questions

What does a ₹85,000 custom AI agent include in India?

It includes one workflow, two integrations, three read tools and one draft tool, a 50-case eval set, deployment, monitoring, a runbook, and 30 days of business-hours support. It does not include extra channels, regional languages, write actions, or a 24×7 SLA.

How much does a custom AI agent cost in India in 2026?

Single-purpose agents start around ₹50,000–₹1L. A working agent with two or three integrations runs ₹1.5L–₹4L. Multi-system workflows with approval gates and compliance start at ₹5L. The build is usually 40–50% of the first-year cost once model calls and maintenance are counted.

What is the monthly running cost of an AI agent?

For 1,000 tickets a month with four model calls per ticket, I model ₹7,000–₹14,000 including retries, embeddings, hosting, monitoring, and messaging. A 4% retry rate can add ₹3,000 a month, so schema-constrained output is the first cost fix.

Why do AI agent quotes vary so much in India?

Quotes vary because scope varies, not because one vendor is cheating. A quote with one integration, no eval set, and no handover is not the same product as a quote with three integrations, approval gates, and a regression suite. Compare the scope table, not the headline number.

📚 Authoritative Primary Sources & Research Citations

• Anthropic Claude Tool Use & Agent Best Practices (2026): Evaluation metrics on stateful tool call retries and schema constraint enforcement.
• Model Context Protocol (MCP) Working Group: modelcontextprotocol.io — Least-privilege agent tool contracts.
• SaaS Next Engineering Ledger (2026): Empirical cost benchmarks across 40+ client agent production deployments in Gujarat and global teams.

Bottom Line

A ₹85K agent is a real product if the scope is narrow and the handover is honest. Price the integrations, the eval set, the write actions, and the monthly retries before you sign. If you are budgeting a build, start with the AI Agency Delivery OS and the AI automation pricing model, explore the complete AI Systems Master Bundle, and review the custom AI development service I run from Junagadh. When you have the workflow and volume, send them to me and I will tell you whether ₹85K is the right tier.