AINPM.in / Field guide
Field guide · Updated 3 Nov 2026

A field guide to AI-native product management in India

AI-native product management is building products where an AI model is the core of the experience, not an add-on. This guide covers what changes for product managers, why India is different, the AI rules to track, a playbook from problem to production, and where to start for your role.

01 · What changes

What changes for a PM? Six shifts

AI-native product management isn't adding a chatbot to a roadmap. It changes what a spec is, how quality is proven, and what a unit of work costs.

Static PRDsEvals are the spec

Acceptance criteria become test sets and graded rubrics. If you can't measure "good", you can't ship a probabilistic product.

Features & screensAgents & workflows

Design the job to be done end-to-end — tools, memory, hand-offs and where a human approves — not just the UI.

Fixed-cost software₹ per outcome

Every task has an inference bill. Model choice, context size and caching are product decisions with margin impact.

Annual roadmapsLearning loops

Models improve monthly. Plan in short bets, instrument everything, and re-run evals whenever the model underneath changes.

Compliance at the endTrust by design

Consent, labelling, audit trails and kill-switches belong in the first sprint — regulators in India now expect them.

English-first UXBharat-first UX

Voice, vernacular and low-bandwidth are defaults, not localisation tasks — Indic models make this practical today.

02 · India lens

Why is India a distinct place to build AI products?

Sovereign compute, Indic foundation models, population-scale digital rails and the world's largest pool of capability centres — a combination no other market has.

38,000+ GPUs

IndiaAI compute

Subsidised national compute under the ₹10,372 crore IndiaAI Mission, with 20,000+ more GPUs announced in February 2026.

Source ↗
22 languages

Indic models

BharatGen Param2 (17B) spans 22 Indian languages; Sarvam launched 30B and 105B models plus speech and vision at the AI Impact Summit.

Source ↗
2,117 GCCs

Capability centres

2.36 million professionals in India's GCCs (FY2026) — increasingly owning global AI products, not just delivery.

Source ↗
DPI rails

Public infrastructure

UPI, Aadhaar, DigiLocker, Account Aggregator and ONDC give AI agents identity, consent and payment rails to act on.

03 · Regulation radar

Which AI rules must an Indian product manager track?

India has chosen principles-based guidelines plus sector regulators over a single AI Act. That means your obligations depend on your domain — and several dates are close.

  1. RBI FREE-AI framework

    Seven "sutras" for responsible AI in regulated finance — trust, people first, fairness, accountability, understandable by design, safety and resilience.

  2. MeitY India AI Governance Guidelines

    Principles-based, techno-legal approach relying on existing law and sector regulators rather than a standalone AI statute.

  3. DPDP Rules notified

    Data Protection Board provisions active; the rest of the Digital Personal Data Protection regime phases in.

  4. IT Rules amendment on synthetic content

    AI-generated media must be prominently labelled; takedown windows cut to 3 hours (2 hours for deepfake intimate imagery).

This summary is for orientation only and is not legal advice. Read the full regulation guide

04 · Playbook

How do you take an AI product from problem to production?

Eight steps I use to take an AI product from problem to production. See them applied in the SEBI RAG case study →

Frame the job, not the model

Start from a user outcome and a baseline. Is AI 10× better here, or just novel?

Write evals before prompts

Golden datasets in the languages and dialects your users actually speak; rubric-graded, versioned, owned by the PM.

Choose models by ₹ per task

Compare frontier APIs, open-weight and Indic models on quality, latency, cost and data residency.

Ground it in your data (RAG)

Retrieval over trusted sources with citations — essential for BFSI, health and government use cases.

Design agents with guardrails

Scoped tools, approval gates, audit logs and a kill-switch. Human-in-the-loop where money or rights move.

Consent & privacy by design

Purpose-limited data, itemised notices and consent-manager readiness for DPDP.

Ship small, observe everything

Trace every call, sample outputs for review, track cost and quality drift per cohort.

Measure outcomes, then scale

Task success, time saved and trust signals — not just DAUs. Re-run evals on every model change.

05 · Start here

Where should you begin?

  • Build one eval set for a feature you own this month — even 50 labelled examples change the conversation.
  • Learn the cost model: tokens, context windows, caching and model routing.
  • Prototype yourself with AI coding tools — PMs who can ship a working demo move faster than specs.
  • Map your data flows against DPDP before May 2027.