The Indian Small Business AI Starter Kit

You don't need a tech team, a CTO, or a six-month project to start using AI in your business. You need to pick the right job, hand it to an agent that lives where your customers already are (WhatsApp, mostly), and stop it from making expensive mistakes. This guide shows you exactly where AI pays off, what it costs in rupees, and how to roll it out in 30 days.

Where AI actually pays off

AI earns its keep when a task is repetitive, text-heavy, and time-sensitive — the kind of work that quietly eats your evenings. It does not pay off for one-off creative decisions, anything legally binding, or work where being wrong 5% of the time is unacceptable (think final invoices or medical advice).

The sweet spot for an Indian SMB is the front office: answering the same WhatsApp questions, qualifying enquiries, drafting content, and chasing follow-ups. These are high-volume, low-stakes, and directly tied to revenue or hours saved.

A simple test before you automate anything:

• Do I (or a staff member) do this task more than 10 times a week?
• Is it mostly reading and writing text?
• Would a "good enough" draft that I review be acceptable?
• Is a mistake recoverable (not a refund, not a legal promise)?

Three or more "yes" answers means it's a strong AI candidate.

The 6 highest-ROI agent use-cases

Start with support or lead capture — they touch revenue fastest and the value is obvious within a week. Follow-ups are the quiet money-maker: most small businesses lose deals simply because nobody chased the quote.

What it realistically costs in rupees

There are three cost layers, and people usually forget the middle one.

1. The agent itself — a hosted subscription on a marketplace typically runs ₹500–₹3,000/month per agent depending on volume and capability. Source code is a one-time buy if you want to own and self-host it.
2. The "plumbing" — the WhatsApp Business API isn't free at scale. Meta charges per-conversation (often a few paise to a couple of rupees, varying by category and India rates), routed through a provider like a BSP. Budget a realistic ₹500–₹2,000/month for a small business; in-app chat, Telegram, and email avoid this entirely.
3. The AI usage — if the agent is priced all-inclusive, this is already covered. If you self-host, you pay the model API directly (usually small for text — tens to a few hundred rupees a month at SMB volumes).

Hypothetical example: a boutique using one WhatsApp support agent might spend ~₹1,500/month subscription + ~₹800/month WhatsApp conversations = ~₹2,300/month, plus GST. If it saves 1.5 hours a day of replying, the maths is not close.

A few money rules:

• Pay per outcome or per month, not per "seat" you won't fill.
• Watch the channel cost — WhatsApp is where your customers are, but it's the priciest pipe. Use email/in-app for internal reports.
• Remember 18% GST on most SaaS subscriptions; if you're registered, you reclaim it as input credit.

Buy vs build vs hire

Three honest paths. Most SMBs should buy.

• Buy a hosted agent if your need is standard (support, leads, content). You subscribe, configure your FAQs and tone, connect WhatsApp, done. Browse what's available at AgentDukaan.
• Hire done-for-you setup when your workflow is specific (e.g. clinic scheduling with your exact slot rules) but you don't want to babysit servers.
• Build only if you have a developer and genuinely need to own the code — for example, deep integration with your billing system. If you go this route, the Build Your First AI Agent: Idea to Live in a Weekend and Deploy & Host AI Agents the Right Way guides will save you weeks.

Rule of thumb: buy to learn, build later if you must. Don't make your first AI project a custom build.

A 30-day adoption roadmap

One agent. One channel. One measurable outcome. Resist the urge to automate everything in week one.

Week 1 — Pick & set up

• Choose your single highest-pain use-case (likely WhatsApp support or lead capture)
• Subscribe to a hosted agent from AgentDukaan
• Feed it your real FAQs, pricing, timings, and tone (Hinglish if that's how you talk)
• Connect your channel and run 10 test conversations yourself

Week 2 — Go live, supervised

• Turn it on for real customers, but keep a human reviewing replies daily
• Collect every reply the agent got wrong; fix the FAQ/prompt, don't blame the bot
• Add a clear "type human to reach our team" escape hatch

Week 3 — Measure

• Track one number: enquiries answered, leads captured, or hours saved
• Compare to the week before you started
• Tighten the prompt based on real transcripts

Week 4 — Expand or refine

• If it's working, add a second use-case (e.g. follow-ups on top of support)
• If it's shaky, stay on one agent and keep improving — don't pile on more

Use this ready-made prompt to configure a WhatsApp support agent's persona. Paste it into the agent's instructions and fill the brackets:

You are the WhatsApp assistant for [BUSINESS NAME], a [TYPE OF BUSINESS] in [CITY].
Reply in the same language the customer uses, including Hinglish.
Be warm, brief (2-4 lines), and use the customer's first name if known.

You CAN answer: timings, location, pricing, product/service details, order
status, and booking requests, using only the FACTS below.
You must NOT: promise discounts, confirm refunds, or invent information.

If you are unsure, or the customer asks for anything involving money owed,
complaints, or custom requests, say: "Let me connect you to our team" and tag
the chat as HUMAN.

FACTS:
- Timings: [...]
- Location: [...]
- Pricing: [...]
- Top 5 FAQs and answers: [...]

For more reusable prompts, the 40 WhatsApp Customer-Support Prompts (Copy-Paste) is the natural next step.

Common mistakes to avoid

• Automating everything at once. One agent, one job. Sprawl kills adoption.
• Going live with no human review. Supervise for the first 2 weeks; an unwatched bot quoting wrong prices is worse than no bot.
• No human escape hatch. Always let customers reach a person. Frustrated customers who feel trapped don't come back.
• Forgetting the WhatsApp/channel cost. The subscription is rarely the full bill.
• Ignoring the DPDP Act. You're collecting customer phone numbers and chats. Only collect what you need, tell people you're using an assistant, and don't feed customer data into tools you don't trust. Keep a simple record of what you store and why.
• Treating a wrong answer as the bot "failing." It's almost always a missing or unclear FAQ. Fix the input.
• Buying on a flashy demo. Test it with your awkward, real customer questions before you commit.

Next steps

• Pick your single highest-pain task using the 4-question test above.
• Spend 30 minutes writing out your real FAQs, pricing, and timings — you'll need them whichever path you choose.
• Run 10 honest test conversations before going live, and keep a human in the loop for two weeks.

When you're ready to see what's available, browse hosted agents at AgentDukaan or ask the Help Center which one fits your use-case. Start with one, measure it, and grow from there — that's how AI actually sticks in a small business.