The short version: Pick one repeated task, give the AI only the data each user may see, call the model from your own backend, check every result, keep the old way as a fallback, release to a few users first, and measure what each request costs before you scale.
This guide shows how to integrate AI into an app you already have, mobile or web: the steps from idea to launch, the parts of the system, which features to start with, whether to use an AI model's API or host your own, what it costs each month to call a model, and what goes wrong.
We are FNA Technology, a software development company with our team in Mumbai, India. We add AI features to existing web apps, portals and internal tools. If you already know you want it built for you, see our AI integration services.
The steps from idea to launch
1. Pick one job for the AI
Start with one task your users already do by hand and repeat often, such as summarising a support ticket, reading the fields off an uploaded invoice, or finding the right help article. Write down who does it, what goes in, what a good result looks like, and what happens today when it goes wrong.
"Add AI to the app" is not a job. If you cannot say which screen changes and how you will know it worked, the feature is not ready to build.
2. Collect real examples with the right answer
As a starting rule of thumb, gather 30 to 100 real inputs from your own data, each with the result a careful person would give, including awkward ones: typos, missing fields, long inputs, and questions your data does not answer. This list is how you will judge every version, so keep it and add to it.
3. Check the data and permissions
List the data the feature needs and where it lives. Then check who may see it: if a user cannot open a record in your app today, the AI must not read it for them either. Leave out personal details the task does not need before anything is sent to a model.
If the data is missing, out of date or locked in a system with no way to read it, fix that first. No model fixes bad data.
4. Choose the model and how to call it
For most first features, use a hosted model through its API, such as one from OpenAI, Anthropic or Google. You pay per use, there are no servers to run, and you can try two or three models on your real examples before choosing. Use the smallest model that passes your examples; bigger models cost more per request and are often slower.
5. Call the model from your backend, not the app
Your mobile app or web page should call your own backend, and the backend calls the model. The backend checks who the user is, collects the permitted data, adds the instructions, calls the model and checks the result. The API key for the model stays on the server; a key inside a mobile app or a web page can be copied and used by anyone.
Put the AI code in one separate service, not spread across your app. Then a new model, a price change or a new provider means changing one service, not your whole codebase.
6. Check every result before it is used
Ask the model to return its result in a fixed format, and reject anything that does not match before your app shows it, saves it or acts on it. Start with features that suggest, such as a draft a person edits or a summary a person reads. Anything that changes records, sends messages or moves money should wait for a person to confirm until the feature has proved itself.
7. Plan for failure
Set a time limit on every call. If the model is slow, down or returns something that fails the checks, fall back to the screen the user had before the AI was added. The AI should be an extra step, never the only way to finish a task.
8. Test on your examples, then launch in stages
Run your real examples, plus attempts to break it: text that tells the AI to ignore its rules, questions about records the user may not see, and very long inputs. Then release behind a switch: to your own team first, then a small group of users, then everyone. The switch should turn the feature off at once without a new app release, which matters for mobile apps waiting on app store review.
9. Measure quality and cost after launch
Log each request: how long it took, how much text went in and out, what it cost, and whether the user kept, edited or rejected the result. Set limits per user and per day. Run your examples again before any change to the model or the instructions goes live, because those changes can change every answer.
How the parts fit together
This is the shape of most AI integrations in an existing app. Each arrow is a call your own code makes; nothing goes from the user's phone or browser straight to the model.
User's app or web page
|
Your backend: who is the user, what may they see
|
AI service: permitted data + instructions + limits
|
AI model (hosted API, or your own model)
|
Checks: right format, allowed action, sources
|
Shown to the user, saved, or sent for a person to approve
(if any step fails: the old screen, without AI)
When the answers have to come from your own documents, such as policies, manuals or contracts, the AI service also searches those documents first and passes only the matching passages to the model. That pattern is called RAG; our RAG chatbot architecture guide explains it in detail.
Which AI features to add first
Good first features have a narrow job, a clear way to tell right from wrong, and a person who checks the result:
| Feature | What it does | A good first use |
|---|---|---|
| Search in plain language | Finds records or articles from a question, not exact keywords | Help centre or product catalogue search |
| Answers from your documents | Answers with a link to the source passage | Staff questions about policies or manuals |
| Summaries | Shortens long tickets, notes or threads | A summary at the top of each support ticket |
| Drafts | Writes a reply or description for a person to edit | Suggested replies that an agent approves |
| Reading documents | Pulls fields out of invoices, forms or emails | Invoice fields filled in for finance to check |
| Tagging and routing | Labels incoming work and sends it to the right team | Sorting support requests by topic and urgency |
| Recommendations | Suggests products, content or next steps | "Related items" on a product page |
| An assistant | Answers questions and looks things up through your APIs | Order status and account questions in your app |
A chat assistant is not always the best start. If users need an exact action with no doubt, such as a refund, let the AI suggest it and let your normal app logic confirm and carry it out.
Use a model's API or host your own?
| Hosted model through an API | Your own model on your servers | |
|---|---|---|
| Upfront work | None beyond your code | Servers with graphics cards, setup and monitoring |
| How you pay | Per request, only for what you use | For the servers all the time, busy or idle |
| Model quality | The largest, newest models | Open-weight models you can download and run |
| Where your data goes | To the provider, under its business terms | Stays inside your environment |
| Who keeps it running | The provider | Your team |
Start with the API unless one of these is true: your data may not leave your own environment by law or contract, you have steady, high volume where your own servers would cost less than paying per request, or you need control over the model that a provider does not offer. Many teams end up with both: a large hosted model for the hard requests and a smaller, cheaper one for simple, high-volume tasks.
Hosting your own model is not automatically more private. You still have to secure the servers, the logs and the network. Equally, a hosted API is not automatically a privacy risk: both OpenAI and Anthropic state that business API data is not used for training by default. Read the terms for the plan you actually use, including how long requests are kept.
What it costs each month to call an AI model
Model providers charge by tokens, small pieces of text, and charge separately for the text you send and the text the model writes back, usually more for the text it writes. So the monthly bill for one feature is roughly:
Requests per month × (tokens sent × price per token sent + tokens returned × price per token returned)
A worked example, with example numbers only and not a quote for any model: a feature called 20,000 times a month, sending 2,000 tokens and getting back 500 tokens each time, on a model priced at USD 2 per million tokens sent and USD 10 per million tokens returned, costs 20,000 × (0.004 + 0.005) = USD 180 a month. The same feature on a model ten times cheaper costs USD 18; on one ten times dearer, USD 1,800. Check the current price on each provider's own pricing page, such as OpenAI's and Anthropic's, before you budget, because prices and models change often.
What makes the bill bigger than expected:
- Sending the whole conversation history or long documents with every request
- One user action that triggers several model calls, as agents often do
- Retries after errors, with no limit
- No limit per user, so a bug or abuse runs up the bill
- Using the largest model for tasks a small model passes
What brings it down: send only the text the task needs, cache results for repeated requests, route simple tasks to a smaller model, and set daily limits. Hosting for the AI service, logs and any search index come on top of the model fees.
The cost of building the integration is separate. At FNA, AI integration projects typically cost between USD 5,000 and USD 200,000, depending on size, and most take around two months.
What goes wrong, and how to prevent it
| What goes wrong | What it looks like | How to prevent it |
|---|---|---|
| Confident wrong answers | An invented refund policy, a wrong total, a source that does not say what the answer says | Answer from your data with sources, allow "not found", test on real examples |
| Another user's data shown | A crafted question returns a record the user may not open | Check permissions before the data reaches the AI, not only in the screen |
| Hidden instructions | Text in a message or document tells the AI to ignore its rules | Treat user text and documents as data, check results, confirm actions |
| The app breaks | Screens hang when the model is slow or down | Time limits and a fallback to the old screen |
| Unwanted actions | The AI changes records or sends messages on its own | Suggest first; a person confirms anything that changes data |
| Runaway bills | Costs jump after a release or from one heavy user | Cost logging, limits per user and per day, alerts |
| Answers change after an update | A new model version or prompt changes results quietly | Run the real examples before every change goes live |
| Nobody owns it | Errors pile up after launch with no one checking | Name one owner who reads the logs and the rejected results |
| Arabic misunderstood | Questions in Gulf dialect, or Arabic and English mixed in one message, get wrong or off-topic answers | Test on formal Arabic, Gulf dialect and mixed Arabic and English, with a native speaker reviewing the answers; see our Arabic AI chatbot page |
| No way to a person | The user is stuck with an assistant that keeps guessing | When it is not confident, it says so and passes the user to a person, with the conversation so far |
Personal data needs extra care: send only the fields the task needs, decide how long requests and answers are kept in logs, and tell users when content was written by AI.
How we do it at FNA
How we add AI depends on the project and how complex it is, so the set-up is agreed after discovery rather than taken from a fixed template. We use OpenAI or Claude models, chosen by how complex the job is, and you own the source code. We can build a working demo of your problem in one day, so you see the approach before the full build.
Add AI to Your Software
AI Integration Services
AI features added to the web apps, portals and internal tools you already run, without rebuilding them.
RAG Development
Assistants that answer from your own documents, with sources and only the documents each user may see.
AI Chatbot Development
Chatbots for your website, app or WhatsApp that answer from your content and hand off to your team.
Frequently Asked Questions
Talk to an engineer
If you have an app and a task you want AI to take on, tell us what it is and which systems hold the data. An engineer will tell you whether AI fits, what the first version would look like and what it would cost to run.

