'Can we add AI to it?' is the most common line in briefs this year, from Mumbai startups and US founders alike. The honest answer is usually yes, and it's often cheaper than people fear. But the cost isn't only the build. AI features have a running bill that grows with every user, and that's the part most estimates skip.
Below I break down the AI features I actually get asked for, what they typically cost to build with a senior independent developer versus an agency, what they cost to run each month, and when on-device AI can bring that running cost close to zero.
The AI features people actually ask for
- In-app assistant or chat. Answers questions about your product, orders, courses or content. The most requested and the most over-scoped.
- Smart search. Search that understands meaning, so 'cheap red running shoes' finds the right products even if the words don't match.
- Summaries and writing help. Summarise a long document, a call, a chat thread or reviews. Draft replies or descriptions.
- Scan and extract. Read receipts, IDs, invoices, menus or forms from a photo and fill in the fields.
- Recommendations. 'You might like' for content, products or lessons, based on behaviour.
- Voice. Speak instead of type, or have the app read content aloud, often in Hindi or other Indian languages for Indian audiences.
What it costs to build
These assume an existing app with a backend, and a clear scope. Building AI into a brand-new app costs about the same on top of the base build.
| AI feature | Senior independent (USD) | Senior independent (INR) | US agency (USD) | Typical time |
|---|---|---|---|---|
| Summaries or writing help | $800 – $2,500 | ₹65,000 – ₹2,00,000 | $5,000 – $15,000 | 1–2 weeks |
| Scan and extract from photos | $1,200 – $4,000 | ₹1,00,000 – ₹3,25,000 | $8,000 – $25,000 | 2–3 weeks |
| Smart search over your data | $2,000 – $6,000 | ₹1,60,000 – ₹5,00,000 | $12,000 – $40,000 | 2–4 weeks |
| Assistant that knows your content | $3,000 – $10,000 | ₹2,50,000 – ₹8,00,000 | $20,000 – $80,000 | 3–6 weeks |
| Recommendations | $2,500 – $8,000 | ₹2,00,000 – ₹6,50,000 | $15,000 – $50,000 | 3–5 weeks |
| Voice input and read-aloud | $1,000 – $4,000 | ₹80,000 – ₹3,25,000 | $8,000 – $25,000 | 1–3 weeks |
What pushes a feature to the top of its range: lots of your own data to connect (documents, catalogues, order history), strict accuracy needs (health, legal, money), multiple languages, and the safety work to stop the assistant saying things it shouldn't.
What it costs to run
Cloud AI models charge per token, which is roughly per word in and out. Prices vary a lot between models and change often, so I won't quote a price list that'll be stale next month. What matters is the shape of the bill:
- Small, fast models cost very little per request. They're fine for summaries, tagging, extraction and simple chat.
- The biggest frontier models can cost many times more per request. Worth it for complex reasoning, not for every tap.
- Your bill grows with usage, not with users signing up. Ten thousand users who each ask two questions a month cost far less than one thousand power users chatting daily.
| Scenario | Monthly usage | Rough monthly AI bill |
|---|---|---|
| Small app, light use | 2,000 active users, 5 questions each | Under $20 with a small model |
| Growing app, regular use | 20,000 active users, 15 questions each | $50 – $400 depending on model |
| Heavy daily use | 50,000 active users, 60 questions each | $500 – $5,000+ depending on model |
Good engineering keeps this low: send less text per request, cache common answers, use a small model by default and a big one only when needed, and set per-user limits. I've seen these choices cut AI bills by more than half without users noticing any difference.
If you charge a subscription, check the AI cost per paying user against the price. A ₹99 a month plan can't fund unlimited calls to the most expensive model.
On-device AI: the free option
This is the biggest shift of 2026. Apple's Foundation Models framework gives iOS apps direct access to the same on-device model behind Apple Intelligence, and at WWDC26 Apple extended it with image input and the option to plug in cloud models such as Claude or Gemini through the same API. On Android, Google offers Gemini Nano on supported phones through its on-device AI APIs.
On-device AI has no per-request bill, works offline and keeps data on the phone, which also helps with privacy laws like India's DPDP Act. The trade-offs: the models are smaller, so they're better at summarising, tagging and short replies than at deep reasoning, and they only run on newer devices. In India, where many users are on budget Android phones, you'll usually need a cloud fallback.
My usual setup is hybrid: on-device for quick, private tasks when the phone supports it, and a cloud model for heavier requests or older phones. It keeps the running bill small without leaving anyone out.
Is the AI feature worth it?
Before I quote, I ask clients three questions:
- What job does it do for the user? 'Fill in this form from a photo of the invoice' is a feature. 'Add AI' is not.
- What happens when it's wrong? A wrong product suggestion is fine. A wrong medicine dose or a wrong refund amount isn't. The higher the stakes, the more checking and the higher the cost.
- Does it pay for itself? More sign-ups, fewer support tickets, longer sessions or a paid tier. If you can't name one, start smaller.
Usually the best first AI feature is a small, boring one that saves users a few taps every day. Scan and extract, or smart search, beats a chatbot nine times out of ten.
How I build AI features
I add AI to new and existing apps built in Flutter, React Native and native iOS and Android, with the backend and admin controls to manage cost and quality. You get a fixed scope and price up front, usage limits built in, and a dashboard that shows what the AI is costing you. If you're curious how this fits into a full build, see app development cost in the USA and UK or in India.
Have an AI feature in mind? Tell me what it should do and I'll come back with a build cost and an estimate of the monthly running cost.
