AI integration
AI integration services for the software you already run
Our AI integration services add large language model features to a product or internal tool you already have: an in-app assistant, search across your documents with cited answers (RAG), summaries of long records, and automatic classification of tickets, emails or forms. A first feature is indicatively ₹60,000–2,00,000 ($2,200–7,500) over 2–4 weeks, tested against your own examples before it ships.
- First feature
- 2–4 weeks
- Indicative cost
- ₹60,000–2,00,000
- Models
- OpenAI, Anthropic and others
- Judged by
- your own test set
Updated · By the Codebrix team
Four LLM features that usually earn their keep
Most AI requests we hear come down to four patterns. Naming the pattern early makes scoping, pricing and testing far easier.
| Feature | What it does | Typical input | How we judge it |
|---|---|---|---|
| In-app assistant | Answers questions and performs simple actions inside your product | A user's question plus their account context | Correct answers, correct refusals, actions only within the user's permissions |
| Document search (RAG) | Finds the right passage in your documents and answers with a citation | Policies, manuals, contracts, help articles | Right passage found, answer faithful to it, link to the source |
| Summarisation | Turns long records into short, structured notes | Call transcripts, case histories, long email threads | Nothing important dropped, nothing invented |
| Classification and extraction | Labels or routes items and pulls fields out of text | Support tickets, inbound emails, uploaded forms | Accuracy against labels your team has checked |
If you need a customer-facing chatbot on your website or app, AI chatbot development covers conversation design, hand-off to staff and channel setup. If the goal is to take whole tasks off your team rather than add a feature to a product, see AI automation.
How document search (RAG) is built
Retrieval-augmented generation means the model answers from your documents rather than from what it absorbed in training. It is the pattern behind most useful internal assistants, and it is where most of the engineering effort goes.
- Ingest: pull documents from where they live, such as drive folders, a CMS or a database, and keep them in sync when they change.
- Split and index: break documents into passages, create embeddings and store them in a vector database, often pgvector inside your existing PostgreSQL.
- Filter by permission: a user should only ever retrieve passages they are allowed to see. This filter runs before the model sees anything.
- Retrieve and answer: fetch the best passages, have the model answer only from them and show the citations.
- Admit gaps: when nothing relevant is found, the assistant says it does not know instead of guessing.
Poor answers are more often a retrieval problem than a model problem, so that is where we look first when quality drops.
Evaluation: how we know it works before your users do
LLM output varies from run to run, so 'it looked fine in the demo' is not a test. Before building, we collect a set of real questions or items from your team, each with the answer a good employee would give. Every change to prompts, retrieval or model is run against that set, and the scores are shown in the weekly demo.
- Correctness and faithfulness: is the answer right, and is it supported by the cited source?
- Refusals: does it decline questions outside its scope or outside the user's permissions?
- Format: does structured output such as JSON, labels or extracted fields parse every time?
- Latency and cost per request, tracked alongside quality.
The same test set protects you later. When a provider releases a new model or retires an old one, we run the set before switching, so an upgrade does not quietly break something that used to work.
Data privacy and where your data goes
Every request to a hosted model sends text to that provider. We design around that fact rather than ignore it:
- Send the minimum: strip or mask personal details the model does not need before a request leaves your system.
- Go through each provider's current data-use and retention settings with you, and configure the options your plan offers.
- Keep prompts, outputs and logs in your own database and cloud account, with a retention period you choose.
- Record who asked what, so access can be audited later.
In India, the Digital Personal Data Protection Act was enacted on 11 August 2023, and the DPDP Rules were notified on 14 November 2025 with an eighteen-month period for phased compliance (PIB backgrounder). We build with the Act in mind, and with GDPR or other laws where your users are elsewhere, but your legal adviser signs off on what is compliant for your business.
Keeping the running cost predictable
Model APIs bill per token, roughly a fragment of a word, with input and output priced separately and usually quoted per million tokens. As of November 2026, both Anthropic and OpenAI also offer batch processing at a 50% discount for requests that do not need an instant reply (Anthropic pricing, OpenAI pricing). Prices change often, so check those pages before you budget.
Most of the running cost is decided by design choices made early:
- Use a small, cheap model for classification and routing, and keep larger models for hard reasoning.
- Cache the parts of a prompt that repeat, such as long instructions or a fixed set of documents.
- Send overnight jobs, like summarising yesterday's tickets, through batch processing.
- Set per-user and per-day limits, with an alert when spend crosses a threshold you choose.
Our AI chatbot development cost guide shows how to estimate monthly API spend from expected usage.
Models, stack and how we fit into your codebase
We work with OpenAI and Anthropic models, use LangChain where its abstractions save time, and store embeddings in a vector database or in PostgreSQL. Calls go through a thin layer in your code that names the model in one place, so switching provider is a configuration change plus a test run, not a rewrite.
The feature can live inside your existing Next.js, Node.js or Python codebase as a pull request your team reviews, or as a small separate service with its own API if your core system is hard to change. If you want people working inside your team rather than a scoped project, you can hire AI developers through us.
Frequently asked questions
What do AI integration services usually include?
Scoping the feature against a real business task, building the retrieval, prompts and code inside your product, an evaluation set to measure quality, cost and rate limits, logging, and handover. A first feature is indicatively ₹60,000–2,00,000 ($2,200–7,500) and takes 2–4 weeks. The fixed price follows a discovery call.
Will the AI make things up?
Language models can produce confident wrong answers. We reduce that by answering from your retrieved documents, showing citations, instructing the model to say when it does not know, and testing against real questions before launch. For anything with legal, financial or medical weight, a person reviews the output before anyone acts on it.
Is our data used to train the AI model?
That depends on the provider, the product and the plan you use, and those terms change over time. Before anything is sent, we go through each provider's current data-use and retention settings with you, configure what your plan allows, and keep prompts and outputs stored in your own systems.
Which model do you use, OpenAI or Anthropic?
Whichever performs best on your evaluation set at a cost you accept. We often run two or three models against the same questions and show you the results side by side. The code names the model in one place, so you can switch later when prices or quality change without rebuilding the feature.
How much will the AI cost us each month to run?
It depends on how many requests you make, how long they are and which model handles them. Providers bill per million tokens, with input and output priced separately. During scoping we estimate monthly spend from your expected usage, set hard limits, and use caching and smaller models to keep it down.
Can you add AI to an old system without rebuilding it?
Usually. If the system has an API or a database we can read, we build the AI feature as a separate service that sits beside it. Users then reach it through your existing interface, a chat channel or email, depending on where they already work, while the old system keeps running unchanged.
Sources
Related
AI chatbots
An AI chatbot development company for assistants grounded in your content
AI chatbot development company in Noida. Assistants that answer from your own documents, hand off to people and are tested before they go live.
Read moreHire developers
Hire AI developers who build, test and ship LLM features
Hire AI developers who build LLM features with retrieval, agents and evals in Python or TypeScript. Paid first milestone, your keys, rate after a call.
Read moreCost guide
AI chatbot development cost: build price and monthly running cost
AI chatbot development cost in two parts: our indicative build range and a worked formula for monthly LLM token and WhatsApp running costs.
Read moreAI automation
An AI automation agency for the repetitive work behind your business
AI automation agency that moves email triage, lead follow-up, CRM updates and WhatsApp replies to AI workflows, with people approving the risky steps.
Read moreSaaS development
SaaS development with multi-tenancy, billing and admin built in from the start
SaaS development company building multi-tenant products with Stripe or Razorpay billing, onboarding, admin and analytics. Milestone pricing, your code.
Read moreB2B & SaaS
A B2B software development company for portals, dashboards and integrations
B2B software development company for customer portals, dashboards, integrations, SSO and audit logs, built so your SOC 2 preparation has evidence to use.
Read more