March 2026 • 12 min read
Dropping a ChatGPT wrapper into your app is not a product strategy. Building AI for the Indian market requires solving for extreme cost sensitivity, high latency on fluctuating 4G networks, and deep vernacular language support. For 99% of startups, use APIs — not custom models. Route simple tasks to cheap models (GPT-4o mini at ₹50/M output tokens), reserve expensive models (Claude Sonnet 4.6, GPT-4o) for complex reasoning, and implement semantic caching to survive at scale.
The first question every Indian PM faces when pitching an AI feature: "Should we build our own model or use an API?" For 99% of startups, the answer is to buy API access. Training a foundational LLM from scratch requires tens of millions of dollars in GPU compute and highly specialised talent that is in short supply globally — let alone in India.
Your competitive moat is not the model itself. It is the proprietary data you feed into the model and the UX you build around it. The real decision is which API to buy.
For English-first applications, the current leaders are OpenAI (GPT-4o, GPT-5.2) and Anthropic (Claude Sonnet 4.6, Claude Haiku 4.5). However, if your core demographic is in Tier 2 or Tier 3 Indian cities, these models are often overkill and expensive for the task at hand.
This is where Indic-first models like Sarvam AI become relevant. Sarvam is trained heavily on Hindi, Tamil, Telugu, Bengali, and other regional languages. When a user speaks colloquial "Hinglish" into a voice search bar, Sarvam processes it with higher contextual accuracy and at a fraction of the cost of routing that same audio to a US-based inference endpoint. Other open-source alternatives include fine-tuned Llama 4 variants for Indic languages, which you can self-host on Indian cloud providers like E2E Networks or Jio Cloud for data residency compliance.
Traditional SaaS features have near-zero marginal costs — serving a dashboard to 1 user costs roughly the same as serving 100 users. Generative AI fundamentally breaks this economic model. Every single prompt a user types costs you actual money in API token usage. If you do not model costs accurately, a viral AI feature can bankrupt your startup in a weekend.
| Model | Input (₹/1M tokens) | Output (₹/1M tokens) | Best For |
|---|---|---|---|
| GPT-4o mini | ₹13 | ₹50 | Classification, routing, simple tasks |
| Gemini 3 Flash | ₹42 | ₹252 | Fast responses, multimodal tasks |
| Claude Haiku 4.5 | ₹84 | ₹420 | Customer support, summarisation |
| GPT-4o | ₹210 | ₹840 | Complex reasoning, data analysis |
| Claude Sonnet 4.6 | ₹252 | ₹1,260 | Coding, long documents, agentic workflows |
| GPT-5.2 | ₹147 | ₹1,176 | Flagship reasoning, multi-step tasks |
Prices exclude 18% GST. Batch API offers 50% discount on both input and output for non-urgent tasks.
Imagine you are building an HR-tech platform and launch a feature that summarises uploaded PDF resumes. Average input: 1,000 tokens (resume text). Average output: 300 tokens (the summary).
| Scale | Using GPT-4o | Using GPT-4o mini | Impact |
|---|---|---|---|
| 10K users/month | ~₹4,600/mo | ~₹280/mo | Negligible either way |
| 1L users/month | ~₹46,000/mo | ~₹2,800/mo | CFO starts asking questions |
| 10L users/month | ~₹4.6L/mo | ~₹28,000/mo | Unit economics shattered vs manageable |
The difference between using GPT-4o and GPT-4o mini for a simple summarisation task is 16x. This is not an optimisation exercise — it is a survival decision.
1. Model routing: Do not use GPT-4o or Claude Sonnet 4.6 for everything. Route simple tasks (classifying a support ticket as "Refund" or "Delivery Issue") to GPT-4o mini or Claude Haiku 4.5. Reserve expensive models strictly for complex reasoning tasks. A well-built routing layer handles 70–80% of traffic at a fraction of the cost with no visible quality drop for end users.
2. Semantic caching: If 5,000 users ask your EdTech chatbot "What is the syllabus for the UPSC exam?", you should only pay the LLM once. Implement a semantic cache (Redis with vector similarity, or a managed service like Momento) that recognises similar questions and serves the pre-generated answer instantly.
3. Hard rate limits: Never release an AI feature without a strict per-user rate limit. Free users get 5 AI generations per day; premium users get 50. This acts as a circuit breaker against abuse and bot attacks, and gives you a natural monetisation lever.
4. Prompt caching (API-level): Both Anthropic and OpenAI now offer prompt caching. If your system prompt and context are consistent across requests, cached input tokens cost up to 90% less. For applications with repeated instructions or RAG context, this is free money — configure it from day one.
5. Batch API for non-real-time: Background tasks like generating weekly reports, bulk email personalisation, or nightly content moderation can use the Batch API at 50% off. Claude Sonnet 4.6 drops from ₹252/₹1,260 to ₹126/₹630 per million tokens in batch mode.
While urban centres in India have 5G connectivity, the reality for a massive portion of users is fluctuating, congested 4G that frequently drops to 3G inside buildings or during train commutes. Generative AI is inherently slow. If an Indian user taps "Generate" and stares at a frozen screen for 8 seconds while an API call routes to US-East-1 and back, they will assume the app has crashed and force-close it.
To solve this, you must engineer perceived speed:
Token streaming is non-negotiable. Do not wait for the LLM to generate the entire 500-word paragraph before showing it. Stream the response word-by-word (the "typing" effect). The time-to-first-token (TTFT) should be under 800 milliseconds. As long as the user sees text appearing on screen, their tolerance for waiting increases dramatically.
Choose the right inference region. Both AWS and GCP now have Mumbai (ap-south-1) and Hyderabad regions. If your LLM provider offers regional inference, use it. The round-trip latency from Mumbai to US-East-1 is ~200ms; Mumbai to Mumbai is ~5ms. That 200ms difference compounds on every single streamed token.
Pre-compute where possible. If your app shows AI-generated recommendations on a home feed, generate them asynchronously via batch API and cache the results. The user sees instant content; the AI computation happened hours ago. Only use real-time inference for genuinely interactive features like chat or search.
Building AI in heavily regulated sectors requires navigating immense trust deficits. Here is how three major Indian players approach it differently.
When dealing with people's life savings, AI hallucinations are unacceptable. Groww has implemented AI for semantic search — users can type "Show me good mutual funds for saving tax" and get relevant results. But the trust hurdle is so high that the AI is strictly sandboxed. It cannot execute trades or offer direct financial advice. It acts solely as a navigation co-pilot, surfacing relevant UI elements and official mutual fund documents. The final transaction remains entirely manual. This "AI as search, not as advisor" pattern is the safest approach for fintech products operating under SEBI regulations.
Many startups make the mistake of slapping a "Chat with AI" button on their homepage. CRED takes the opposite approach. They use ML deeply, but it is entirely invisible to the user. ML models power real-time fraud detection at the transaction layer and dynamically generate hyper-personalised, variable-reward cashbacks based on spending velocity and patterns. The AI solves business logic problems — fraud prevention, personalisation, churn prediction — rather than being a UI feature. This is often the higher-ROI approach: AI that improves margins invisibly rather than AI that users interact with directly.
Implementing AI symptom checkers in India requires navigating regulatory grey areas. The Medical Council of India has strict rules regarding diagnostics. Practo positions AI not as a replacement for doctors, but as a "triage assistant." The AI gathers structured patient data before the consultation, summarising symptoms and history for the human doctor. It is positioned as a co-pilot for the physician, not an autonomous diagnostician for the patient. This framing — "AI prepares, human decides" — is the template for any health-adjacent AI feature in India.
Before adding AI to your roadmap, run it through a JTBD audit. Are you adding AI because your investors want an "AI narrative," or because it genuinely reduces friction for the user? If an AI chatbot takes 5 text prompts to accomplish what a standard UI button can do in 1 click, you have degraded the user experience.
Use AI to collapse complex workflows (turning unstructured voice input into structured forms), parse unstructured data (extracting information from PDFs, images, or voice), and personalise at scale (adapting content, recommendations, and notifications per-user without manual segmentation). If the task is simple and structured, a well-designed UI will always beat a chat interface.
No. For 99% of Indian startups, training a foundational model from scratch requires tens of millions of dollars in GPU compute — capital that is far better spent on product, distribution, and data. The correct strategy is to buy API access (OpenAI, Anthropic, Sarvam) and build proprietary workflows and prompt chains on top of them. Your moat is the data and UX, not the model weights.
For most Indian products: use GPT-4o mini or Claude Haiku 4.5 as your default model (they are the cheapest options that still produce quality output). Upgrade to GPT-4o, Claude Sonnet 4.6, or GPT-5.2 only for tasks that require complex reasoning, long-context analysis, or high-quality writing. Many production systems route 80% of requests to the cheap model and 20% to the expensive one. The cost difference is 10–20x.
If your AI feature processes financial data (transaction history, portfolio details, PAN/Aadhaar), you need to consider data residency. Both AWS (Mumbai) and GCP (Mumbai) support Indian data residency. For API providers: Anthropic and OpenAI process data on US/EU servers by default, but Anthropic now offers data residency controls via the inference_geo parameter. For strict compliance, consider self-hosting open-source models (Llama 4, Mistral) on Indian cloud infrastructure using E2E Networks or Jio Cloud.
Sarvam AI is specifically trained on Indic languages and is viable for production use cases where vernacular language support is the primary requirement — voice search, customer support in Hindi/Tamil/Telugu, and regional content generation. For English-first applications or complex reasoning tasks, the global models (Claude, GPT-4o) remain significantly better. The practical approach is to use Sarvam for Indic-language tasks and global models for everything else.
Don't blow your AWS credits on unoptimised API calls. Let our advisory team review your AI product spec, model your INR costs at scale, and ensure your architecture is ready for the Indian market.
Book Free Strategy Call →