In 2026, adding an AI feature to an app or website you already run costs roughly 2.5 to 5 lakh for a single feature such as search, summaries or document extraction, about $3,000 to $6,000 or £2,200 to £4,400. An in-app assistant runs 4 to 9 lakh, about $4,800 to $10,700 or £3,500 to £8,000, and a first integration of two or three features 6 to 14 lakh, about $7,100 to $16,700 or £5,300 to £12,400. A single feature takes an Indian team three to six weeks; two or three take four to nine weeks with two people.
That is the first number, and it is paid once. The second number is the one this article is really about. After launch, every AI feature costs an amount per active user, every month, on top of a small fixed hosting bill, for as long as people use it, and it rises with how much they use it. In the examples below that amount runs from 19 paise to 128 rupees, and where a feature lands is set by three design decisions: whether you cache, which model each step uses, and how much you send it.
It also comes out of your plan price, so it sets your gross margin on every user who touches the feature. Most integration quotes stop at the build price; this guide puts the running cost beside your plan price, because that is where the decision gets made.
What it costs to build, feature by feature
| Feature | What the work actually is | Cost in India | USD |
|---|---|---|---|
| Search that understands meaning | An index of your records, kept in step with the database, filtered by what each user may see, tuned against real queries | 2.5 to 4 lakh | $3,000 to $4,800 |
| Summaries of records and long documents | The prompt, a cache keyed to each record or document version, invalidation when it changes | 2.5 to 3.5 lakh | $3,000 to $4,200 |
| Extracting fields from documents or photos | Reading the file, mapping it to your form, a confidence score, a person checking the doubtful ones | 3 to 5 lakh | $3,600 to $6,000 |
| Recommendations from your own data | Events you may not be logging yet, a model of what goes with what, a fallback for new users | 3 to 5 lakh | $3,600 to $6,000 |
| Drafting replies or documents | Context from the record, a draft the user edits, logging of what they kept | 2.5 to 4 lakh | $3,000 to $4,800 |
| An in-app assistant | Answers from your help content and the user's own data, with that user's permissions, and a hand-off to support | 4 to 9 lakh | $4,800 to $10,700 |
Dollar and pound figures use 84 rupees to the dollar and 113 to the pound, as across our cost guides. The rupee figures are what established Indian teams charge at a blended 16,000 rupees a person-day.
Below about 2 lakh, what you are buying is a demo. A feature can be wired to a model in an afternoon. What turns it into something you can put in front of paying customers is the evaluation set, the permission checks, the feature flag and the logging, and that is where the days go.
A document summary feature, where generative AI reads each incoming contract, report or claim and writes a summary without acting on it, is priced on the summary and extraction rows above: 2.5 to 5 lakh, about $3,000 to $6,000. If it must act on what it reads, it is an agent, priced in our AI agent cost guide.
The assistant sits above a custom chatbot over your own content, which our AI chatbot cost guide prices at 3 to 8 lakh, and above the answering tier in our AI agent cost guide at the same 3 to 8 lakh, for one reason: it reads the signed-in user's own data, so it has to respect your existing permission model on every answer. A chatbot on a marketing site answers everyone the same way. An assistant inside your product must never show one customer another customer's invoice. The retrieval and permission work behind it is set out on our LLM and generative AI apps page.
Why your codebase sets the price, not the model
Calling a model is the cheap part of the build, as it is with AI agents. The days go into the connection between the model and software that was never designed to have one.
- Data access. The feature needs to read records through a layer that applies the same permissions the rest of your app does. If your permissions live in forty controllers rather than one place, this is the largest line in the estimate.
- Keeping things in step. A search index or a summary cache is a copy of your data, and a copy goes stale. Deciding when it refreshes, and making sure a deleted record disappears from it, is real engineering.
- A clean boundary. The AI runs best as a separate service your app calls. An old monolith with a working API is often easier than a new app with no boundary at all, which surprises the owners of both.
- Mobile apps are no different. The model is called from your server, not from the phone, so a Flutter, React Native or native iOS or Android app falls in the same ranges. The new screens are part of the interface work; the only addition is a store release and its review time.
- The evaluation set. Fifty to two hundred real examples with known good answers, so a prompt change or a model update can be scored before it ships rather than discovered by customers.
- The flag and the control group. Every feature should ship switched off, then on for a share of users, with one metric agreed in advance. That is also your rollback plan. We have had prototypes that did not move the metric; we said so at the end of the second week, and the client had spent two weeks rather than two months.
How we handle each of these is set out on our AI integration page. The point for budgeting is that two quotes for the same feature can differ by two times and both be honest, because one team has looked at your code and the other has not.
Where the money goes: a worked estimate
A B2B software product used by 8,000 people a month, with case records and attached documents. The owner wants two features: search that understands meaning across records and documents, and a summary of each record's history when it opens.
| Work package | Person-days |
|---|---|
| Discovery: read the codebase and data model, score the candidate features, model running cost on real volumes, agree the metric | 4 |
| Data access: a read-only layer onto records and documents, with each user's permissions applied | 6 |
| Search: the indexing pipeline, keeping it in step with the database, permission filters | 8 |
| Search quality: an evaluation set of real queries, tuning, and a fallback to the old search | 4 |
| Summaries: the prompt, a cache keyed to record version, invalidation when a record changes | 5 |
| Interface: results and a summary panel inside the existing screens | 4 |
| Feature flags, a control group, and logging of cost and outcome on every request | 3 |
| Security review and a data-flow note for your customers' questionnaires | 2 |
| Staged rollout and handover to your team | 2 |
| Total | 38 |
That is about 6 lakh, roughly $7,200 or £5,400, over four to five weeks with two people. Each feature costs less here than in the table above because discovery, the data access layer, the interface and the flags are built once and shared by both. Note that the two AI features themselves, search and summaries, are 13 of the 38 days. The rest is the connection to software that already exists, which is exactly the part a quote written without reading your code cannot see.
The second number: cost per active user
Ordinary software has almost no marginal cost. One more user costs you a little hosting and nothing else. An AI feature changes that: every search, summary and answer is a metered request, so the feature has a cost per active user per month that never goes away.
The method is two multiplications. How many times does an active user trigger the feature in a month, and what does one request cost? Here it is for each feature, at usage levels we see often and illustrative model prices: a small model at 15 cents per million input tokens and 60 cents per million output tokens, and a mid-tier model at $3 and $15, the same mid-tier price our running cost worked example uses and a deliberately cautious one, since the newest mid-tier models list at about $2 and $10. Prices move, in both directions; the method does not, and it is the same whether the feature calls OpenAI, Anthropic or Google. The provider and the model tier change the meter, not the build.
Each request is assumed to take, in tokens in and out: summaries 2,500 and 250 on the small model, plus 10 percent for retries; extraction 3,000 and 300; drafting 3,000 and 400; the assistant 8,000 in per turn with the whole conversation resent, or 2,500 trimmed, and 350 out, on the mid-tier model. Extraction, drafting and the assistant include the 30 percent allowance for retries and checks from our running cost example. A cached summary is rebuilt about once every ten opens, and search uses the all in figure from our integration page, about $20 per 50,000 searches.
| Feature, usage per active user per month | Expensive design | Cheaper design | Difference |
|---|---|---|---|
| Search, 30 searches | About 1 rupee | About 1 rupee | None worth chasing |
| Summaries, 40 record opens, small model | 1.94 rupees, regenerated on every view | 19 paise, cached per record version | 10 times |
| Extraction, 5 documents | 7.37 rupees on a mid-tier model | 34 paise on a small model | 22 times |
| Drafting, 20 drafts | 32.76 rupees on a mid-tier model | 1.51 rupees on a small model | 22 times |
| In-app assistant, 40 turns, mid-tier model | 127.76 rupees, full history resent each turn, no caching | 55.69 rupees, trimmed context | 2.3 times |
In dollars, that runs from a fraction of a cent a user for cached summaries to about $1.52 for the assistant built the expensive way, or about 66 cents trimmed. Run every assistant turn on the small model, once your evaluation set shows quality holds, and the same 40 turns cost roughly 2.5 rupees trimmed and 6 with full history. The other way round matters too: both summary columns assume the small model, and on the mid-tier model each would cost about 20 times as much.
These figures assume no provider prompt caching. OpenAI caches a repeated prompt prefix automatically and Anthropic does when the request asks for it, billing the cached part at about a tenth of the input price while a conversation stays active, so an assistant that resends its history within a session costs less than the first column shows. New passages retrieved for each turn are still billed in full.
The summary rows reproduce the figures on our integration page: summarising a record on every open at 200,000 opens costs about $115 a month, and about $11 when the summary is cached by record, because a record that has not changed does not need summarising again. Search is cheap either way, which is one reason it is usually the right first feature.
For the worked example above, 8,000 active users make about 240,000 searches and 320,000 record opens a month. Search costs about 8,000 rupees, summaries about 15,500 regenerated on every open or about 1,550 cached, and monitoring about $30, or 2,500 rupees. That comes to about 26,000 rupees a month, roughly $310 or £230, regenerating every summary, or about 12,000, roughly $145 or £107, caching them: 3.26 rupees a user against 1.52. On top sits a fixed 4,200 to 12,600 rupees a month, $50 to $150, for the search index and the AI service's hosting, which does not grow with users. Neither total will trouble anyone. On the mid-tier model the same summaries would push it over 3 lakh a month, which is why the model each step uses belongs in the design. The assistant row is where the meter starts to matter.
Put it against your plan price
A cost per user means nothing until it sits beside revenue per user. Here is the assistant, the most expensive feature in the table, against three typical plan prices.
| Plan price per user per month | Assistant, full history | Assistant, trimmed | A heavy user, full history | A heavy user, trimmed |
|---|---|---|---|---|
| A consumer app at 199 rupees | 64% of the price | 28% | 3.2 times what they pay | 1.4 times what they pay |
| An Indian business seat at 999 rupees | 13% | 6% | 64% | 28% |
| A US seat at $20, or a UK seat at £15 | 8% | 3% | 38% | 17% |
A heavy user here means five times the typical usage, 200 assistant turns a month.
Read the heavy user columns first. The typical user is not the risk. The risk is that AI costs scale with use while most plans charge per seat, so your most enthusiastic customer, the one you would put in a case study, can cost more to serve than they pay. On a consumer plan the trimmed design does it too: a heavy user costs about 278 rupees a month against a 199 rupee plan. There, design alone will not save the margin. A smaller model for the assistant, a monthly cap on turns or a separate paid tier has to be part of the plan, decided before the build.
Software has traditionally run at gross margins around 80 percent, and Bessemer Venture Partners puts AI companies at 50 to 60 percent against 80 to 90 for SaaS. A feature costing 13 percent of the plan price takes 13 points of margin, permanently, from a typical user of it, and more from the heavy ones. Free tiers are worse: a free user with an AI feature is no longer free to serve. The other costs that grow per user on a subscription product are in our SaaS development cost guide.
None of this is a reason not to build. It is a reason to model the number before the build, and to choose the design with it in mind.
Three design decisions that move the meter
- Cache by record version, not by view. If the input has not changed, the output has not either. On a record opened about ten times between edits, that is the tenfold saving on summaries. Adding the cache is quick; making it refresh whenever a record changes is the harder part, and both sit inside the five summary days in the worked estimate. It also makes the feature faster.
- Give each step the smallest model that does it well. Extracting an invoice number or classifying a ticket rarely needs the model that writes a careful reply. Moving a whole feature to the small model is where the twentyfold differences in the table come from; splitting one feature's steps between two models, as in our running cost example, saves a third to a half. Either way, check quality on your evaluation set before switching rather than after.
- Send less. Most of an assistant's bill is input: the conversation so far and the documents retrieved for each turn. Summarising old turns and retrieving fewer, better passages halves the meter and usually improves the answers.
The fuller list of levers is in what an AI agent costs to run. For integration, caching moves to the top, because features inside an app tend to repeat the same request far more than a support agent does.
What the same integration costs in the US and the UK
UK agencies publish figures of £15,000 to £40,000 for a single AI feature in an existing product. Contract AI engineers charge a median of about £600 a day before any agency margin, and senior London specialists £800 to £1,000 and up. Guides from two US firms put a basic integration at $10,000 to $30,000 and a single API feature at $5,000 to $20,000, but both firms also deliver from teams in India, so those ranges already blend in offshore rates.
The 38 person-days in our worked estimate come to about 6 lakh in India, roughly $7,200 or £5,400. At a contractor's £600 a day, before any agency margin, the same days cost about £23,000 in the UK, and at the $150 to $250 an hour US agencies bill for AI engineers, against $80 to $150 for general development, $45,000 to $76,000. The work is identical. The difference is where the people live, and the fully loaded comparison, including what it takes to work well across time zones, is in our guide to hiring developers in India. Those multiples compare project rates; the hiring guide compares a dedicated engineer by the month, which is why its ratios are higher.
The costs that are not in the quote
| Cost | Typical size | When it surprises you |
|---|---|---|
| The meter itself | From paise to over a hundred rupees per active user a month, as above | The month usage takes off |
| Evaluation runs | Small, a few thousand rupees a month | When someone skips them and quality slips unnoticed |
| Model retirement | A few days of work, every time a provider retires a model, often on a few months' notice | Usually in the middle of something else |
| Security questionnaires | A day or two per enterprise customer who asks where their data goes | The first large deal |
| Rate limits at peak | Queueing and retry logic, a few days | The campaign that triples usage for a week |
| The control group period | Two to four weeks before you know if it worked | When the launch date was promised to customers |
| Maintenance | 15 to 25 percent of the build a year, the usual rule from our app maintenance guide | Every year |
For clients in India, GST at 18 percent is added to the development invoice and is claimable as input credit if you are registered. Clients in the US, the UK and elsewhere pay no Indian GST, because services exported from India are zero rated. Overseas AI providers stop charging GST once you add your GSTIN to your billing settings, but you then pay the 18 percent yourself under reverse charge and claim it back in the same return. Without a GSTIN the provider adds it and you cannot.
When not to pay for an integration
- When the software is a SaaS product you subscribe to. Check whether the vendor already ships the feature, or has it on the roadmap. Switching on theirs costs a fraction of anyone integrating around them.
- When you want a chat widget on a marketing website. That is a platform purchase in most cases, and the plan fees and per-resolution pricing are in our AI chatbot cost guide.
- When the feature should act, not just read. Refunds, bookings and updates in other systems make it an agent, with limits, approvals and an audit log, priced in our AI agent cost guide.
- When the AI is the product. A new product built around a model is an MVP, not an integration, and the right starting point is our MVP development cost guide or our LLM and generative AI apps page.
- When your own engineers have the time. A team with a few free weeks and someone who has built an evaluation set before can ship search on its own, and the method in this guide works just as well for costing it in house.
- When nobody can say what number should move. A feature without a metric agreed in advance cannot be judged, and cannot be switched off with confidence either. An AI readiness audit finds that number in two weeks.
How to read an AI integration quote
- Ask whether the team read your code before quoting. If the estimate arrived before anyone saw your data model, it is a guess.
- Ask for the running cost per active user, feature by feature, on your volumes. A quote with only a build price has left out the number that stays.
- Ask how permissions are handled in the data layer. The answer should name your existing permission model, not invent a new one.
- Ask for the evaluation set and who owns it after handover.
- Ask what happens when the provider retires the model. One service to update is the right answer. Forty call sites is the wrong one.
- Ask for the rollback plan. It should be a switch, not a deployment.
- Ask what leaves your systems. The quote should say, per feature, what the model provider receives, what is stored and for how long. Major providers do not train on data sent through their APIs by default, and our AI integration page sets out how we map the rest.
The wider checklist for choosing who builds it is in how to hire an AI development company in India, and if the method question is retrieval against fine-tuning, that is in RAG against fine-tuning.
Questions people ask
How much does AI integration cost?
Two numbers. Building one feature into an app you already run costs 2.5 to 5 lakh with an Indian team, about $3,000 to $6,000, and an in-app assistant 4 to 9 lakh. After launch each feature has a running cost per active user per month, from about a rupee for search to over a hundred rupees for an assistant built carelessly. The worked estimate above, two features for a product with 8,000 users, is 38 person-days and about 6 lakh.
How much does it cost to integrate ChatGPT into a website or app?
On a marketing site, a ChatGPT style chat widget is usually a platform plan at 2,500 to 30,000 rupees a month plus 8 to 170 rupees per resolved conversation, or 3 to 8 lakh to build over your own content, and our AI chatbot cost guide shows where one becomes cheaper than the other. Inside your own product, answering from each signed-in user's data with their permissions, it is an integration: 4 to 9 lakh for an assistant, or 2.5 to 5 lakh for a narrower feature such as search or summaries.
What does an AI feature cost to run per user?
At the usage and illustrative prices in this article, about a rupee a month for search, 19 paise to 2 rupees for summaries, 34 paise to 7 rupees for document extraction, 1.5 to 33 rupees for drafting, and 56 to 128 rupees for an in-app assistant on a mid-tier model, or a few rupees on a small one. The spread inside each feature is set by caching, which model each step uses and how much context is sent.
Do we need to rebuild our app to add AI?
No. The AI runs as a separate service beside your existing backend, called from the screens that need it, behind a feature flag. With the flag off your app behaves exactly as it does today, which is also how you roll back.
How long does an AI integration take?
Three to six weeks for a single feature: a week to read the code and choose it, a week to prototype on your data, and two to four weeks to integrate, measure and roll out. Two or three features take four to nine weeks with two people; the worked estimate above runs four to five.
Which AI feature should we add first?
Search, most often. It is cheap to run, easy to measure and fails safely, since an unhelpful result costs nothing. Summaries come next where people read a lot. An assistant is the most visible and the most expensive to run, so it is rarely the right first move.
Will AI features hurt our margins?
They can. Software has traditionally run at gross margins around 80 percent, and an AI feature costing 13 percent of your plan price takes about 13 points on a typical user and more on heavy ones. Model the cost per active user before the build, check what your heaviest users would cost, and design caching and model routing in from the start.
Is it cheaper in India than in the UK or US?
For the same work, yes, by about four times against UK contractor day rates, more against UK agencies, and six to ten times against US agency rates for AI work. Our worked estimate is about 6 lakh in India, roughly £23,000 at UK contractor day rates and $45,000 to $76,000 at US hourly rates. The running cost per user is the same wherever the team sits, because it is paid to the model provider.
Getting a real number
Ranges end where your codebase begins. We read your code and data model, score the candidate features, and model the running cost per active user on your real volumes, then quote the build as a fixed price after a one-week discovery. Sometimes the answer is that your vendor already ships what you want. What that process looks like is on our AI integration page.