Generative AI integration: how to add it to your product and what it costs

Generative AI integration means connecting a model that writes text or creates images to your product, your data, and your systems, so it does one job inside a real workflow. At Ronas IT, AI integration starts at $3,000 per AI feature and takes at least 1 week, with the design of any new screens and project management included. If the scope is not written down yet, an optional analysis phase of $1,000 to $2,000 can come first.
This guide is for product owners and business leads who want generative AI in their own software rather than in a separate chat window. It covers where adoption stands, which tasks suit the technology, the four ways to connect a model, how the model fits into your system, how we run such a project, its main risks, and what it costs.
What generative AI integration is
A generative model creates new content, such as text, code, or images, from an instruction. Integration is everything around the model that makes it useful in your business: the data it gets for each request, the code that checks its output, the place where the result appears, and the rules for when a person reviews it.
A chat assistant that your team opens in a browser is not an integration. It knows nothing about your orders, customers, or documents, and its answers stay in the chat. An integrated model works inside your product or internal tools: it drafts a reply in your help desk, fills in fields from an uploaded invoice, or answers a customer from your own product data. When the model also chooses which tools to use and in what order, the system becomes an agent. We cover that case in our guide on how to build a custom AI agent.
Where generative AI adoption stands
Many people already use generative AI, and most companies use AI in some form, but far fewer companies report that AI has added to their profit. The St. Louis Fed found that more than half of US adults aged 18 to 64 used generative AI in 2025, while fewer than four in ten respondents in McKinsey's 2026 survey say AI has contributed to their organization's EBIT.
- People: The Federal Reserve Bank of St. Louis (November 2025) reports that 54.6% of US adults aged 18 to 64 used generative AI, 10 percentage points more than a year before. Use at work grew from 33.3% to 37.4% over the same 12 months.
- Companies: In McKinsey's The state of AI in 2026 survey (August 2026), nearly nine in ten respondents report regular use of AI in at least one business function, 44% say AI is scaling across their enterprise, and 37% report that AI has contributed to their organization's EBIT.
- Pilots: In a July 2024 forecast, Gartner expected at least 30% of generative AI projects to be abandoned after proof of concept by the end of 2025 and named four reasons: poor data quality, weak risk controls, rising costs, and unclear business value.
Each reason in Gartner's list is a question you can answer before the build: which data the model needs, what it may not do, what one request costs, and which number should change. The steps later in this guide are built around those questions.
Tasks that suit generative AI
Generative AI suits tasks that involve free text or images, repeat often, and have a result a person can check quickly. If fixed rules can do the job, rules are cheaper and more predictable.
| Task | Example | What to watch |
|---|---|---|
| Drafting text | A reply to a support ticket, a product description, a follow-up email | Tone and facts; a person approves anything that reaches a customer |
| Summarizing | A long ticket thread, a call transcript, a contract before review | Missed details; the summary links back to the source |
| Extracting fields | Supplier, amounts, and dates from an invoice or a form | Every field is validated in code before it is saved |
| Answering from your documents | A help center assistant, internal search over policies | Answers only from retrieved passages, with the source shown |
| Sorting and routing | Deciding whether an email is a lead, which team a request goes to | A fixed list of allowed values and a fallback to a person |
| Creating images | Storyboard frames, product visuals, design drafts | Consistency between images and control over each detail |
“The best first project is boring. Pick the task your team already complains about, where a wrong draft costs a minute to fix, not a customer. Once that works and people trust it, the harder ideas get easier to fund.”
Roman Surikov, CEO at Ronas IT
Four ways to integrate: an API call, RAG, fine-tuning, or an open model
The simplest start is a hosted model called through an API, with retrieval added when the model has to answer from your own data. Fine-tuning and a model on your own servers solve narrower problems: a format the model keeps getting wrong, or data that may not leave your infrastructure.

| Approach | How it works | When it fits | Main cost after launch |
|---|---|---|---|
| API call to a hosted model | Your backend sends instructions and the user's input to a model such as OpenAI GPT, Anthropic Claude, or Google Gemini | Drafting, summarizing, extracting, and sorting, where the input holds everything the model needs | Model usage, billed by the amount of text in and out |
| Retrieval-augmented generation (RAG) | Your system finds the relevant passages in your documents or records and sends them with the request | Answers from your help center, policies, product data, or anything that changes often | Model usage, plus hosting the search index |
| Fine-tuning | A model is trained further on your own examples of inputs and correct outputs | A fixed format or style at high volume that instructions alone do not hold | Preparing training data, and training again when the task changes |
| Open model on your servers | An open-weight model, such as Qwen, Mistral Small, or OpenAI gpt-oss, runs in your own cloud account or data center | Data that may not be sent to an outside provider | The servers that run the model, and the DevOps work to maintain them |
AWS describes retrieval-augmented generation as making a model reference “an authoritative knowledge base outside of its training data sources” before it answers. The model does not learn your documents: it reads the right passages each time, so an updated policy changes the next answer without any retraining. Our guide to AI-based knowledge management systems covers search across several sources and the rules for who may open what.
For the model itself, our technology radar lists the OpenAI API, Anthropic Claude, and Google Gemini as our default hosted models, and Ollama as our default way to run open-weight models on your own servers.
The approaches combine. A support assistant can use RAG over the help center and an API call to a hosted model, and a later version can move to an open model if the data rules change. That is one more reason to keep the model behind your own backend, as the next section explains.
How the model fits into your system
The model sits behind your backend, and the app never calls it directly. The app sends a request to your server, the server adds the data and instructions, calls the model, checks the result, and only then returns it or saves it. That keeps the provider's key off users' devices and gives you one place to control cost, logs, and access.
“Give the model the smallest job you can, such as filling in a few fields from a user's request. Anything that must be exact, like prices, permissions, or which records change, belongs in code that behaves the same way every time. That also makes the feature easier to test, because most of it does not depend on the model.”
Alexander Storozhevsky, Lead developer at Ronas IT
- Versioned instructions: each version of the prompt is kept, so you can compare them and roll back.
- Structured output: the backend asks the model for named fields and validates them in code before anything is saved. A value outside the allowed list goes to a person or falls back to a safe default.
- Background jobs: long requests, such as analyzing a video or generating images, run in a queue, and the app shows progress instead of holding an open request. In a video analysis feature we built, the API marks the analysis as pending at once and notifies the user when the result is ready.
- Streaming: for chat-like features, the answer appears while the model writes it, so the user does not wait for the full reply.
- Usage logs and limits: each call is recorded with its feature and user, so you see what each feature costs before the bill grows. Limits per user or per day cap that cost.
- Timeouts and fallbacks: if the provider is slow or down, the feature fails in a clear way instead of hanging.
On one client project, users describe the workout they want in their own words. The model, Google Gemini, only turns that text into filters: a duration, plus muscle groups and equipment that it picks from lists the backend inserts into the prompt. Ordinary code then builds the workout from the app's own exercise catalog. Each prompt is a separate record in the database, edited in the admin panel. One is active at a time, and an earlier one can be switched back on. If the call fails, the error is reported, and the request still returns empty filters instead of an error response. The tests replay saved model responses to check how the backend handles a valid answer and an empty one, and they also check a failed call.
How we integrate generative AI, step by step
We run an integration in six steps. Projects without a written scope can start with an optional analysis phase that takes at least one week and ends with user stories, a technical feasibility check, an architecture review, and a scope document that lists the budget and the timeline.

- Choose one task and its result. Name the workflow, the people in it, and the number that should change, such as time per ticket or the share of invoices entered by hand. Our article on integrating AI into existing business processes shows how to map a process and find the steps worth automating.
- Try the real model on real inputs. Experiments show what the model does well before anyone designs screens. On an AI tool for a creative agency, we tried GPT-4o and DALL-E 3 on storyboards. Generating the whole scenario as one picture kept it consistent but gave no control over details, so we decided each frame would be generated on its own, and added a character editor to keep characters consistent between frames.
- Collect a test set. Save typical inputs with the output you expect, including the hard and unusual ones; we build it from anonymized or synthetic data, not from your customers' records. Score every later change to the prompt, the model, or the retrieval on this set.
- Build the integration in sprints. Our sprints are usually one to three weeks long, and each one ends with something you can open and check, such as a working feature on a test server. Our engineers build the integration with coding agents, AI tools that write and test code. An engineer reviews and accepts every change and checks it against the test set.
- Add limits and review. Decide which outputs a person approves, what the model may never do, and the spending limit per user or per day.
- Roll out in stages. Give the feature to one team or a share of users first, then widen the group. Each week, collect the outputs that were wrong and save them as new cases in the test set.
Data, security, and compliance
The main risks are sending data the provider may keep, text in the input that changes the model's behavior, and answers that sound right but are wrong. Each one is handled by how you build the integration, whatever model you choose.
- Provider data terms: read what the provider stores, for how long, and whether it trains on your data. OpenAI's API documentation is one example: API data is not used for training unless you opt in, and abuse monitoring logs are kept for up to 30 days by default.
- Prompt injection: OWASP puts prompt injection, in which input changes “the LLM's behavior or output in unintended ways,” first in its Top 10 for LLM applications and notes that RAG and fine-tuning do not fully prevent it. One of the measures OWASP lists against it is to limit the model's access to the minimum its task needs.
- Wrong answers: ground answers in retrieved passages, show the source, and test on the saved set before each release.
- Provider account: the feature calls the model through an API key kept on your backend. Model usage is not part of our price: you pay the model provider directly, on an account in your name. Restrict the keys that already exist in the same cloud project to the APIs they need: Truffle Security found that public Maps keys could call Gemini once their project turned on the Gemini API.
For a structured review of risks, the US National Institute of Standards and Technology published the Generative Artificial Intelligence Profile of its AI Risk Management Framework in July 2024. If your data falls under rules such as HIPAA or GDPR, ask your lawyer which providers and data flows are allowed before the build starts.
What generative AI integration costs
At Ronas IT, each generative AI feature in an existing product starts at $3,000, and an analysis phase before the build is optional. The table lists our starting prices for the parts of such a project and for narrower or different options, with when each one applies.
| What you need | When you need it | Starting price | Timeline |
|---|---|---|---|
| Analysis phase: the tasks for the model, a check that they are technically feasible, an architecture review, and a written scope with its budget and timeline | Optional, before the build, if you have no written scope | $1,000 to $2,000 | At least 1 week |
| AI integration: one generative AI feature inside your mobile app, web product, or internal process, such as a chatbot that answers questions and qualifies requests | For each AI feature, the design of its new screens included | From $3,000 per AI feature | At least 1 week |
| API development: an API that lets the integration reach one of your systems | Only if a system the model needs has no API | From $2,500 | At least 2 weeks |
| Machine learning: a custom model trained on your anonymized or synthetic data | When the task is a prediction, such as demand or churn, not generated content | From $15,000 | At least 6 weeks |
| Support of the live product, the AI integration included: monitoring, fixes, and updates | After launch, as an optional monthly plan: one app with its backend, or a whole system with several client apps | From $1,000 per month; from $2,000 with several client apps | Ongoing |
A project where one AI feature needs a system with no API yet starts at $5,500: from $2,500 for that API and from $3,000 for the feature. Project management and the design of any new screens the feature needs are included in both.
A feature that touches regulated data or many outside systems takes more engineering work, and its estimate is higher. Our breakdown of AI development cost shows what goes into these prices.
After launch you also pay the model provider for usage and pay for hosting. Usage grows with the number of requests and the amount of text in each, so RAG, which adds passages to every request, costs more per answer than a short prompt. In McKinsey's 2026 survey, about 20% of respondents said AI operating costs, including token costs, limited their AI use. The usage logs from the architecture section show this cost per feature before it grows. If you run an open model on your own servers, you pay for those servers instead of a usage fee, and their setup is DevOps work that we estimate by its scope.
Two later costs come from the provider's schedule rather than your traffic. Hosted models are retired: Anthropic's model deprecations page says requests to a retired model fail and promises at least 60 days' notice for publicly released models. Each switch means running the test set again and adjusting the prompts. We plan the architecture so that the model can be replaced from the first release, which lets the move to a newer model be part of our monthly support. In RAG, the search index belongs to the embedding model that built it, so a new embedding model means embedding every existing document again, not only new files. Our guide to an AI knowledge management system explains why and how to plan that re-index.
What to do next
Pick one task from the table of tasks above, name the number it should change, and collect real examples of its inputs with the results you expect. Then check which of the four approaches it needs and which systems the model has to read from. With those answers, our optional analysis phase can turn the idea into a scope, a budget, and a timeline.
Frequently Asked Questions (FAQs)
What is generative AI integration?
It is the work of connecting a generative model, such as a language model or an image model, to your product or internal systems. The model gets your instructions and the data it needs for each request, and your software checks, stores, and shows what the model returns.
How much does it cost to integrate generative AI into an app?
At Ronas IT, integrating generative AI into a mobile app, a web product, or an internal process starts at $3,000 for each AI feature, such as a support chatbot or a document assistant. The design of new screens and project management are included. Model usage fees and hosting are paid separately once the feature is live.
How long does generative AI integration take?
One AI feature takes at least 1 week at Ronas IT, and an analysis phase, if one comes first, adds at least one more week. A task with several features, a RAG pipeline, or a new API takes longer, and the estimate names the timeline. A sprint here usually runs one to three weeks; at its end the feature is ready to try on the development server.
Should we use RAG or fine-tuning?
Start with RAG when the model has to answer from your documents or data that changes, because the model reads the current version at request time. Fine-tuning trains a model further on your examples. It is worth trying only when careful instructions and examples in the prompt still do not give the format or behavior you need.
Will the AI provider train its models on our data?
It depends on the provider and the plan, so read its data terms before you send real data. For example, OpenAI's API documentation says that since March 1, 2023, data sent to its API is not used to train its models unless you opt in. The same page sets a default retention of up to 30 days for abuse monitoring logs.
Can generative AI work with our CRM, ERP, or other existing systems?
Yes. Your backend reads the records the model needs through the system's API, sends them with the request, and writes the checked result back. If a system has no API, one may have to be built first. At Ronas IT, building an API for an existing system starts at $2,500.
How do we stop the model from making things up?
You cannot stop it fully, but you can make wrong answers rare and easy to catch. Give the model the source passages and ask it to answer only from them, check its output in code, test every change on a saved set of real examples, and send answers that reach customers or change records to a person when the model is unsure.
Which AI model should we choose?
Choose by results on your own test examples, by the price per request, and by the provider's data terms. Keep the model behind your backend, so you can switch providers or add an open model later without changing the app.




