Apply for AI Grants India

What human-centric AI means in practice

Building human-centric AI consumer products means treating the model as one component of a product—not the product itself. The goal is to help people complete meaningful tasks with less effort while preserving their agency, privacy, dignity, and ability to challenge an outcome.

For Indian teams, this requires more than adapting a global interface. Products may serve users across languages, literacy levels, bandwidth constraints, device types, and trust environments. A voice assistant for a Bengaluru professional, a health-information tool for a rural user, and a finance app for a first-time digital customer need different safeguards and interaction patterns.

A useful product promise should answer three questions:

• Whose problem is being solved? Define a specific user and context rather than targeting “everyone”.
• Where does AI add value? Use AI for ambiguity, personalisation, generation, or pattern recognition—not as decoration.
• What remains under human control? Make approval, correction, escalation, and deletion straightforward.

Start with a problem, not a model

Begin with interviews, diary studies, support-ticket analysis, and observation of the existing workflow. Document what users currently do, where they lose time, what errors cost them, and what they fear might happen if automation fails. Include people who are likely to be underserved: users with disabilities, regional-language speakers, older adults, and customers with low digital confidence.

Turn research into a narrow job-to-be-done and a measurable outcome. “Use AI to improve customer experience” is too vague. “Help a Hindi-speaking customer understand a refund status in under two minutes, without exposing account information” is testable.

A strong discovery process produces:

• A primary user segment and clearly bounded use case.
• Situations where the system must refuse, defer, or involve a person.
• A baseline workflow and success metric.
• A list of high-impact failure modes.
• Consent, data-retention, and deletion requirements.

Teams building agentic products should also map every action an agent may take. The principles used in building distributed systems with AI agents are relevant here: define permissions, isolate tools, record events, and design for partial failure rather than assuming autonomous execution will be reliable.

Design the interaction around user agency

AI interfaces should make uncertainty visible without overwhelming people. Tell users when content is generated, identify the source of retrieved information where possible, and avoid presenting guesses as facts. Confidence scores alone are rarely meaningful; plain-language explanations and actionable next steps are more useful.

Design for correction. Users should be able to edit a generated answer, undo an action, change preferences, report a harmful output, and reach a human or conventional workflow. Avoid dark patterns such as preselected data-sharing consent, confusing cancellation flows, or making it harder to opt out than to enrol.

For voice and conversational products, test turn-taking, interruptions, accents, code-switching, and noisy environments. Indian users commonly mix English with Hindi or other regional languages, so benchmark performance on real speech rather than polished demonstrations. A practical implementation reference is building a voice agent with Whisper and ElevenLabs, but production teams must add consent, recording controls, fallback channels, and retention limits.

Accessibility belongs in the first prototype. Support screen readers, keyboard navigation, captions, readable contrast, scalable text, low-bandwidth modes, and non-voice alternatives. Test with disabled users; automated checks cannot evaluate whether an AI interaction is genuinely understandable.

Build privacy and safety into the architecture

Treat personal data as a product liability as well as an engineering input. Before collecting data, specify why it is needed, how long it will be kept, who can access it, and how users can delete or correct it. Minimise prompts and logs, redact sensitive fields, encrypt data in transit and at rest, and separate production data from development environments.

For Indian deployments, align your controls with the Digital Personal Data Protection Act, 2023 and applicable rules, while obtaining current legal advice as requirements evolve. Do not copy GDPR language blindly; provide consent and notices that users can understand in the languages and contexts you support.

Create a risk register covering:

• Hallucinated or outdated advice.
• Privacy leakage through prompts, memory, logs, or retrieval.
• Prompt injection and malicious uploaded content.
• Discrimination across language, gender, caste, age, disability, or geography.
• Unsafe automation, including payments, account changes, or health recommendations.
• Over-reliance, emotional dependency, or misleading claims about sentience.

Use layered controls: input validation, retrieval permissions, model and tool restrictions, output checks, rate limits, human review, incident response, and account-level controls. High-impact decisions should not be delegated to a general-purpose model without domain-specific review and a meaningful appeal process.

Evaluate with Indian users and real failure cases

A polished demo says little about reliability. Build an evaluation set from anonymised, representative interactions and include adversarial cases, ambiguous requests, code-mixed language, spelling variation, and low-quality audio. Measure more than accuracy:

• Task completion and time saved.
• Factuality, citation quality, and refusal quality.
• Performance across languages, regions, devices, and user groups.
• Privacy leakage and unsafe-action rates.
• Correction, escalation, retention, and complaint rates.
• Cost, latency, battery use, and failure recovery.

Run offline tests before launch, then conduct moderated usability sessions and a limited pilot. Keep a versioned record of prompts, models, datasets, policies, and evaluation results so regressions can be traced. Open-source projects can accelerate experimentation; teams may find useful patterns in building high-performance AI applications with open-source tools and building open-source AI tools for Indian developers. Open source does not remove the need for licensing, security review, or user safeguards.

Launch responsibly and improve continuously

Start with a constrained release. Limit capabilities, user volume, data access, and irreversible actions until evidence supports expansion. Publish clear product documentation covering intended use, known limitations, data practices, and support routes. Give internal support teams a playbook for harmful outputs, privacy requests, model outages, and urgent escalation.

After launch, monitor cohorts rather than relying only on averages. A strong overall metric can hide poor performance for a minority language or older device. Establish thresholds that trigger rollback or human review, and notify users when a material model or policy change affects behaviour.

Measure whether the product creates durable value, not merely engagement. Useful indicators include successful task completion, reduced confusion, repeat use for the intended job, lower support burden, and the rate at which users can identify and correct errors. If an AI feature increases time spent but worsens decisions, remove or redesign it.

A practical 90-day build plan

• Days 1–20: Interview users, define the job, map risks, establish a baseline, and write data and escalation requirements.
• Days 21–45: Prototype the smallest useful workflow with clear disclosure, correction, accessibility, and human fallback.
• Days 46–70: Build evaluation sets, test representative Indian languages and devices, conduct red-team exercises, and complete privacy and security reviews.
• Days 71–90: Run a controlled pilot, monitor incidents and subgroup outcomes, publish limitations, and decide whether to expand, revise, or stop.

Human-centric AI is not a branding layer added after model selection. It is a product discipline that connects user research, interface design, engineering, governance, and measurement. Teams that make the system understandable, bounded, and correctable are more likely to earn trust—and more capable of discovering where AI genuinely improves consumer experiences.

FAQs

How do I choose the right AI feature?

Choose a task with clear user value, sufficient data or context, tolerable risk, and a conventional fallback. If a deterministic workflow works equally well, use it instead of adding AI.

Should a consumer AI product disclose that it uses AI?

Yes. Disclose AI involvement at the point where it affects a user’s decision or trust. Explain what the system can do, what it cannot do, and how the user can get help or challenge an outcome.

How can a small Indian team manage safety work?

Start with a narrow scope and a documented risk register. Use established model and infrastructure controls, test with representative users, log incidents, and bring in legal, accessibility, domain, or security expertise for high-impact use cases. A smaller scope is often the strongest safety control.

When should an AI action require confirmation?

Require confirmation before irreversible, costly, sensitive, or externally visible actions—such as sending a message, making a payment, changing an account, or submitting a formal application. Show the exact action and key details before execution.

Apply for AI Grants India

If you are building a responsible AI consumer product in India, AI Grants India can help you identify funding, mentorship, and ecosystem support. Prepare a concise application covering the user problem, evidence of demand, technical approach, safety plan, evaluation method, and expected impact.