01 Cost guide · AI
AI app development cost: how much does an AI app cost to build and run?
AI app development cost is typically ₹2–8 lakh to add an AI feature to an existing app, and ₹3–10 lakh to build a chatbot on your own data. Running costs come on top: LLM usage is roughly ₹3,000–15,000 a month under 5,000 conversations, ₹15,000–60,000 at medium volume and ₹60,000+ at high volume, plus hosting.
Running costs from ₹3,000 a month
03 Build cost
How much does an AI feature cost to build?
An AI feature added to an app you already have typically costs ₹2–8 lakh ($2,500–10,000) to build, before running costs. Where it lands in that band depends less on the AI itself than on the work around it: getting your data ready, testing the answers, and setting limits on what the feature may say and do.
A feature that only reads your content, such as search or summaries, usually sits lower in the band than one that acts for the user, such as an assistant that fills in a form, because acting needs more guardrails and more testing.
- Build
- ₹2–8 lakh
- Running
- From ₹3,000 a month
Common first AI features- Search over your own content. Users ask in plain words and get the passage that answers them, not a list of links.
- Summaries. Long reports, records or threads cut down to what a reader needs first.
- An assistant inside the app. Help with a task on the screen where the user already is, such as filling a form or finding a record.
What pushes the cost up- Data preparation. Content that is scattered, scanned or out of date has to be cleaned and structured before any model can use it.
- Evaluation. A set of real questions with agreed answers to check the feature against, built alongside it.
- Guardrails. Limits on what the feature may answer or do, and what it says when it does not know.
AI we run ourselvesKaipo, our own conversational AI app, has long-term memory, voice notes, images and 23 languages including Hinglish, and routes different models to different jobs, with fallbacks.
Automating a whole workflow rather than adding one feature? See AI automation.
AI chatbot development cost: what a chatbot costs to build
An AI chatbot that answers from your own documents or data typically costs ₹3–10 lakh ($4,000–12,000) to build. Our AI Chatbot package takes 4–8 weeks and covers the chatbot on your website, an admin panel and handover; running costs are billed on top, at cost.
A chatbot for one website in one language, answering from a help centre, sits at the low end. Several sources, more than one language and a handoff to your team move it towards the top. WhatsApp is scoped separately.
- Build
- ₹3–10 lakh
- Timeline
- 4–8 weeks
What moves the price- Number of sources. One help centre is quicker to index and test than a document library, a database and a website together.
- Channels. Each channel, such as the web or WhatsApp, has its own interface, limits and testing.
- Languages. Every language needs its own test questions, and sometimes its own sources.
- Human handoff. Passing a conversation to a person, with its context, adds a flow to design and test.
- Admin panel. What your team can see and change: conversations, unanswered questions and sources.
A chatbot we run ourselvesTailGPT, our own AI agent for websites, answers only from a site and its documents. The range above is for a custom build; TailGPT is free to try.
Our chatbot development services are on AI chatbot development, and every package, AI Chatbot included, is on pricing. Building an AI product rather than one feature? See AI development.
04 Running cost
LLM running cost: what does an AI app cost to run each month?
An AI app typically costs ₹3,000–15,000 a month in LLM usage under 5,000 conversations, ₹15,000–60,000 at medium volume and ₹60,000+ at high volume, plus hosting. The bill is built one conversation at a time, so that is where to start.
The question
What the user types: usually a sentence or two.
Adds: the tokens in the question.
The retrieved context
The passages fetched from your documents because they match the question, sent along so the answer rests on your content.
Adds: every passage sent. It can be the largest share, which makes it the first lever to trim.
The prompt
The instructions and guardrails that tell the model how to answer, plus the conversation so far.
Adds: the instructions, and every earlier turn the conversation carries.
The answer
What the model writes back to the user.
Adds: the tokens in the reply, so longer answers cost more.
- Cost per conversation equals (prompt + retrieved context + answer) × the model's price per token
- Monthly bill equals cost per conversation × conversations a month
Embeddings and vector storage
Indexing your documents is a one-off cost, then a small ongoing one as your content changes.
Both are small next to the build and the model usage, and each appears as its own line in the quote.
Hosting and compute
Hosting and compute run ₹2,000–40,000 a month, depending on scale.
The servers, the search index and the database that run the feature, sized to your traffic.
Monitoring, evaluation and maintenance
Plan for 15–20% of the build cost a year to keep an AI feature answering well.
Answer quality can drift as your content changes and models are updated, which is why an AI feature needs its answers checked again over time.
Channel fees
Messaging channels such as WhatsApp charge their own fees, billed at cost.
They are separate from model usage, and they appear as their own line in the quote.
Who pays
You do, at cost: each running line is listed in the quote and, where possible, billed to your own accounts.
So you can watch the running bill yourself, line by line.
A worked example
A chatbot handling about 3,000 conversations a month sits in the low usage band: typically ₹3,000–15,000 a month in model usage.
Where in that band it lands depends on the model, the context size and caching.
The levers that bring each line down are in the next section. For the models and methods behind these features, see AI and LLM development.
05 Cut the bill
How can you cut AI running costs?
You cut AI running costs by sending less to the model, and by sending it to the smallest model that does the job. Each lever here works on one step of the conversation above.
Each is easiest to set up before launch, while the feature is still being tested against its questions. We route by job on our own Kaipo app.
- Route by difficulty
- Send routine questions to a smaller model, and use a larger one only for the questions that need it.
- Cache what repeats
- Store answers to common questions and reuse repeated parts of the prompt, instead of paying for them every time.
- Trim the retrieved context
- Send the few passages that answer the question, not every passage that matches it.
- Cap conversation length
- A long conversation carries every earlier turn into each new request, so summarise or reset it past a set length.
- Batch what can wait
- Run jobs that are not urgent, such as nightly summaries or re-indexing, in batches rather than one request at a time.
- Track cost per conversation
- Measure it from the first day, so a change that raises the bill shows up in the week it ships.
07 Testing
How are AI features tested before release?
Test an AI feature against a set of real questions with agreed good answers before release. A sound testing plan has four parts; ask for each one when you compare quotes.
An evaluation set
Real questions your users will ask, each with an answer that the people who rely on it agree is right. Test the feature against the set before launch.
Guardrails
Rules on what the feature may answer and do, and what it says when no source covers the question. TailGPT, our own product, refuses prompt-injection and off-topic requests.
Review of failures
Read each wrong or weak answer from testing, and fix the sources, the prompt or the scope before release.
Logging after launch
Log the questions the feature could not answer, for your team to review, as the NFTR Assistant does: it reports unanswered questions to admins.
The QA steps every project goes through are on our process page.
09 FAQ