What is a custom AI agent, and why does the price vary so much?
An AI agent is software built around a language model that can decide what to do next and use tools: look up an order, read a file, update a CRM, send a message, then check its own result. A chatbot answers; an agent acts. If the term is new, start with what an AI agent is.
Quotes from AI development companies range widely because “AI agent” covers everything from a single prompt with one tool to a system that runs a whole back-office process. The four things that drive the price are the number of tools the agent can use, how many systems it connects to, how costly a mistake would be, and how much testing that risk demands.
How much does it cost to build, by complexity?
| Agent type | Example | What drives effort | Silvane band |
|---|---|---|---|
| Single-task agent | Reads incoming vendor emails, extracts invoice data and files it | One or two tools, clear right answer | Starter: from ₹50,000 |
| Multi-tool assistant | Answers staff questions from SOPs, looks up stock and raises tickets | Several tools, knowledge base, permissions | Starter to Growth |
| Customer-facing agent | Takes orders or bookings on WhatsApp and voice, updates CRM, collects payment links | Live customers, Hindi and Hinglish, hand-over, guardrails | Growth: from ₹1,20,000 |
| Multi-agent process | Several agents that qualify leads, prepare quotes and chase follow-ups, with human approval steps | Orchestration, monitoring, audit logs | Growth to Scale (custom, from ₹1,00,000/month) |
For an Indian SME, the first agent should almost always be a single-task one. It proves value in weeks, teaches your team how to work with AI, and gives you real usage numbers before you commit to anything larger.
What does the AI model cost per month?
Model providers charge per million tokens, with output tokens costing more than input. Prices as of November 2026 from the providers’ pricing pages:
| Model | Input (per million tokens) | Output (per million tokens) | Good for |
|---|---|---|---|
| Claude Haiku 5.5 (prompts up to 100K tokens) | $0.10 | $0.50 | High-volume, simple steps: classification, extraction |
| Claude Sonnet 5.5 | $2 | $10 | Most agent work: reasoning with tools, customer replies |
| Claude Opus 5.5 | $4 | $20 | Hard reasoning, long multi-step tasks |
| OpenAI gpt-6.1-terra | $1.00 | $5.00 | General-purpose alternative |
| OpenAI gpt-6-luna | $0.10 | $0.50 | Low-cost, high-volume steps |
Agents use more tokens than chatbots because they loop: think, call a tool, read the result, think again. As an illustration, an agent task that uses five model calls of about 4,000 input and 400 output tokens each comes to about $0.06 on Sonnet 5.5, or roughly ₹5.40 at an assumed ₹90 per dollar. A thousand such tasks a month is about ₹5,400. Prompt caching (reusing the same instructions across calls) and the Batch API (50 percent off for non-urgent jobs) can cut that sharply, and routing easy steps to a small model cuts it further.
What about hosting and maintenance?
- Hosting. Many agents run inside n8n or a small cloud server. For an SME this is usually a modest monthly bill. Keeping data in an Indian cloud region helps with DPDP Act 2023 obligations if you process personal data.
- Vector database or search. Agents that answer from documents need a search index. Small ones fit in free or low tiers; large document sets cost more.
- Monitoring. Logs of every decision and tool call, plus alerts when something fails. Non-negotiable for anything customer-facing.
- Maintenance. Models are updated and retired, APIs change, your policies change. Budget for a monthly review of transcripts and failures, either your team’s time or a run plan (from ₹15,000 a month at Silvane).
- Evaluation. A fixed set of test cases you rerun after every change, so improvements do not quietly break something else.
Do data protection rules add cost?
If the agent handles customer names, phone numbers, health or financial details, the Digital Personal Data Protection Act, 2023 applies. In practice that means collecting consent where needed, using data only for the stated purpose, limiting who and what can see it, and keeping logs. Building these in from the start adds a little design and testing time. Retrofitting them after launch costs more. Choosing a model provider and cloud region with clear data-handling terms, and sending the model only the fields it needs, is usually enough for an SME agent.
Build in-house, buy SaaS, or hire an AI development company?
| Option | Upfront cost | Ongoing cost | Fits when |
|---|---|---|---|
| SaaS agent tool | Low | Per seat or per resolution, often in dollars | The task is common (support FAQ, meeting notes) and the tool already does it |
| In-house team | Hiring time and salaries | Salaries | AI is core to your product and you have engineering leadership |
| AI development company or studio | Fixed build fee | Run plan plus usage at cost | You need a specific agent wired into your systems and want it owned by you |
Questions to ask an AI development company
- Will the agent run in my accounts (model API, cloud, n8n), so I own the code and data if we part ways?
- Which model will you use, why, and what is the expected monthly usage cost at my volume?
- How will you test it? Ask to see the evaluation cases and the pass rate before launch.
- What can the agent never do without human approval? (Payments, refunds, medical or legal advice.)
- How are errors logged and who gets alerted?
- What happens when the model provider retires the model you used?
- How is personal data handled under the DPDP Act 2023, and where is it stored?
Our longer checklist for picking a partner is in how to choose an AI automation agency in India. If you are still deciding whether you need an agent or a simpler workflow, read what agentic AI is and what AI automation costs.
A worked example: the full monthly cost of one agent
Take an illustrative distributor in Ludhiana that receives about 1,500 purchase orders a month by email, as PDFs, photos and plain text. An agent reads each order, matches products to the item master, checks stock, drafts a sales order for approval and replies to the customer with expected dispatch.
| Line | Assumption | Monthly cost (illustrative) |
|---|---|---|
| Model usage | 1,500 orders × 5 calls × ~4,000 input and ~400 output tokens on a mid-size model | About $90, or roughly ₹8,100 at ₹90/$; less with caching and small-model routing |
| Hosting | Self-hosted n8n on a small cloud server | A modest server bill |
| Run plan | Transcript review, fixes, new product rules | From ₹15,000 at Silvane, or your own team’s time |
| People | A staff member approves drafted orders | Minutes per order instead of the full entry time |
The point of the exercise is the comparison, not the exact rupee figure. If the agent saves one person most of a working day, the monthly cost is a fraction of that salary, and the order errors it catches are a bonus. If the volume were 50 orders a month, the same agent would be hard to justify.
What makes an AI agent project cost more than quoted?
- Unclear rules. If the team cannot say what the right action is in a case, the agent cannot either. Write the rules down before the build.
- Poor source data. Product names that differ between Tally and the CRM, or customers stored three times, force clean-up work.
- Scope creep. “Can it also do refunds?” is a new tool, new risks and new tests. Add it in phase two.
- Systems without APIs. Old desktop software may need exports or a workaround, which is slower to build and to run.
How Silvane can help
Silvane designs and builds custom AI agents for Indian businesses, inside your own model and cloud accounts, with logs, tests and documentation. If you are not sure where AI fits yet, AI consulting and strategy starts with an audit. A discovery call and AI audit costs ₹199 for founding members (normally ₹15,000) and the fee is credited to the build. Starter builds start from ₹50,000 with a run plan from ₹15,000 a month; Growth from ₹1,20,000 plus from ₹40,000 a month; Scale is custom from ₹1,00,000 a month. Tools and usage are billed at cost, billing is 50/50, everything is built in your own accounts, you get 30 days of free fixes and 30 days’ notice to cancel. GST is extra.
Sources: Anthropic model pricing · OpenAI API pricing · n8n pricing