Compliance-first delivery
What ChatGPT integration services involve
ChatGPT integration services connect OpenAI’s models to your own software and data. There are four routes: calling the OpenAI API (now the Responses API) from your app or website; building an app that runs inside ChatGPT on the MCP-based Apps SDK; deploying the same models through Azure OpenAI in Microsoft’s South India region; or wiring a no-code automation. The right route depends on who uses it, where the data must live and how much control you need.
“Integrate ChatGPT” can mean four different projects
When a business asks for ChatGPT integration services, it usually wants one of four things. A support assistant on its website. Drafting help inside Gmail, Outlook or a CRM. A way for staff to ask ChatGPT about company systems without copying data around. Or bulk work on spreadsheets and documents. Those are four different builds with different costs, risks and data rules, and the first job is to say which one you are buying.
The second job is to build it like software, not like a demo. That means API keys on the server, never in a browser; personal data minimised before it reaches the model; an eval set of real examples that every prompt or model change must pass; a spending cap; and a plan for when OpenAI changes or retires something. The Assistants API, which many 2024 integrations were built on, was switched off on 26 August 2026. Integrations built on it simply stopped.
A note on our own stack, because it matters for trust: the AI in Radiatus’s own products runs on the Claude API, not OpenAI. We build on OpenAI for clients who choose it, keep the model layer swappable, and test both families against the same evals. Our public MCP server works inside ChatGPT, which is a real ChatGPT integration we run ourselves. For the wider picture of what we build with language models, see generative AI applications and AI automation.
ChatGPT integration services we build
Five kinds of project. Each starts with a written list of data sources, users, what the model may and may not do, and how success will be measured.
Assistants in your product or website
The problem
Visitors and customers ask the same questions all day, and a FAQ bot cannot answer anything it was not scripted for.
How we do it
We build on OpenAI’s Responses API with retrieval over your own content, structured outputs where the answer feeds another system, citations back to the source, and a clear handover to a person. Keys and rate limits live on your server; the browser never sees them.
What you get
- Assistant grounded in your content
- Citations and “I don’t know” handling
- Human handover with the transcript
- Eval set and cost dashboard
Apps that bring your systems into ChatGPT
The problem
Your team already uses ChatGPT, but to use company data they copy and paste it in, which is slow and leaks data you would rather control.
How we do it
We expose the data and actions your team needs as an MCP server with per-tool permissions, then connect it to ChatGPT as an app built with OpenAI’s Apps SDK, which is based on the Model Context Protocol. The same server works with Claude and other MCP clients.
What you get
- MCP server with scoped tools
- Sign-in so users see only their data
- Usage and error logging
- Works in ChatGPT and other MCP clients
Gmail, Outlook, Excel and CRM integrations
The problem
Sales and support staff spend their day drafting similar emails, summarising long threads and tidying CRM notes.
How we do it
We add drafting and summarising where the work already happens: an add-in or extension for the mail client, a Sheets or Excel function for bulk text work, or an automation that writes a draft into Zoho, HubSpot or Salesforce for a person to send. Drafts, never auto-sends, until the evals say otherwise.
What you get
- Draft-for-review in the tool staff use
- Bulk spreadsheet processing via Batch
- CRM notes and summaries
- Per-user and per-month spend caps
Assistants API migration and hardening
The problem
An integration built in 2024 on the Assistants API stopped working when OpenAI switched it off on 26 August 2026, or it still runs on a model that is due to retire.
How we do it
We move it to the Responses API, re-create threads, files and tools in the new model, add the eval set it never had, and pin model versions so the next change is a planned upgrade instead of an outage.
What you get
- Migration to the Responses API
- Eval set built from real past cases
- Pinned models and a deprecation watch
- Before-and-after cost comparison
Azure OpenAI for enterprise controls
The problem
Security or procurement wants Microsoft’s enterprise terms, private networking and the models in an Indian Azure region.
How we do it
We deploy OpenAI models through Microsoft Foundry in Azure’s South India region (Chennai) where the model you need is offered there, with private endpoints, managed identities and logging into your own subscription. Microsoft lists no Azure OpenAI availability in Central India, so we check each model per region first.
What you get
- Foundry deployment in your subscription
- Private networking and managed identity
- Region check per model
- Logs and cost in your Azure account
Which ChatGPT integration route fits your use case?
Start from who will use it and where the data must live. The technology follows.
| Route | Who uses it | Best for | Watch out for |
|---|---|---|---|
| OpenAI API in your app or site (Responses API) | Your customers or staff, inside your software | Website assistants, product features, back-office processing | You own security, evals and costs; you pay per token |
| An app inside ChatGPT (Apps SDK on MCP) | People already working in ChatGPT | Letting staff query or update company systems from ChatGPT | Needs a ChatGPT plan and admin approval; tool permissions must be tight |
| Azure OpenAI (Microsoft Foundry) | Your app’s users, under Microsoft’s terms | Enterprises standardised on Azure, with private networking needs | Model availability differs by region; check before you design |
| No-code automation (Zapier, Make, n8n) | Back-office workflows | Simple “when X arrives, summarise and file it” jobs | Fragile at volume; secrets and data flow through a third party |
Most real projects combine two: an API integration for customers and an app for staff, sharing one MCP server or one set of tools underneath. For bigger automation jobs, our n8n automation page explains when a workflow tool beats custom code.
ChatGPT integration with WhatsApp, Excel, Gmail, CRMs and websites
These are the integrations people search for most. Each has a sensible default design and one common mistake.
| Integration | Usual design | Common mistake |
|---|---|---|
| Website | Chat widget calling your server, which calls the Responses API with retrieval over your pages and documents | Putting the API key in the page’s JavaScript, where anyone can copy it |
| Gmail or Outlook | An add-in that drafts or summarises the open thread and leaves the draft for the person to send | Auto-sending replies before anyone has measured how often they are wrong |
| Excel or Google Sheets | A custom function or script for small jobs; the Batch API, at half price, for thousands of rows | Calling the API row by row in real time, slowly and at full price |
| CRM (Zoho, HubSpot, Salesforce) | A webhook or scheduled job that writes summaries and suggested next steps into the record | Giving the model write access to every field instead of a few named ones |
| Your business’s own assistant on the WhatsApp Business Platform, answering about your products and bookings | Trying to put “ChatGPT itself” on WhatsApp: since 15 January 2026 Meta’s terms keep general-purpose AI assistants off the platform |
WhatsApp has its own message charges and opt-in rules. Our guide to WhatsApp AI chatbot rules and costs in India covers them, and our WhatsApp Business API page covers the integration itself.
What changed in OpenAI’s platform, and why old integrations break
- The Responses API is the main API. OpenAI recommends it for new work and folded the useful parts of the Assistants API into it.
- The Assistants API is gone. OpenAI announced its deprecation on 26 August 2025 and shut it down on 26 August 2026. Anything still calling it fails. OpenAI publishes a migration guide.
- Apps inside ChatGPT are built on MCP. OpenAI’s Apps SDK uses the open Model Context Protocol, so a well-designed MCP server can serve ChatGPT, Claude and developer tools from one codebase. Read our MCP server development page for the security model.
- Models retire on a schedule. OpenAI publishes deprecation dates. Pin model versions, watch that page, and re-run your evals before switching, not after users complain.
Is your data safe with a ChatGPT integration?
It can be, if the settings and contracts are right. What OpenAI’s data controls guide says, as of November 2026:
- Training: data sent to the API has not been used to train OpenAI’s models since 1 March 2023, unless you opt in.
- Abuse-monitoring logs: kept for up to 30 days unless law or harm prevention requires longer.
- Zero data retention: available for eligible endpoints, including the Responses and Chat Completions endpoints, with OpenAI’s approval. Some features keep application state by design, so check each one you use.
- Data residency: OpenAI lists India among its data residency regions, alongside the US, Europe, the UK, the UAE, Japan, Singapore, South Korea, Australia and Canada. In India it requires approval for abuse-monitoring controls and a modified retention amendment. The pricing page adds a 10% uplift on regional-processing endpoints for models released on or after 5 March 2026.
Under India’s DPDP Act your company remains the data fiduciary, and OpenAI and any integrator are processors acting under contract. Show a clear notice where personal data goes to an AI service, send only the fields the task needs, and mask the rest. The rule-by-rule build list is in our DPDP compliance checklist.
What does ChatGPT integration cost to run?
Three illustrative workloads, priced with OpenAI’s standard list prices read on 1 November 2026: GPT-6.1 Sol at $2.00 input and $10.00 output per million tokens, GPT-6.1 Terra at $1.00 and $5.00, and GPT-6 Luna at $0.10 and $0.50. Not client work; change the volumes to yours.
| Workload | Assumptions | GPT-6.1 Sol | GPT-6.1 Terra | GPT-6 Luna |
|---|---|---|---|---|
| Email drafting for a sales team | 2,000 drafts a month; 3,000 tokens in, 500 out each | $22.00 | $11.00 | $1.10 |
| Website assistant with retrieval | 8,000 conversations of 4 turns; 5,000 tokens in, 300 out per turn | $416.00 | $208.00 | $20.80 |
| Spreadsheet clean-up via Batch | 50,000 rows; 400 tokens in, 50 out; Batch at half price | $32.50 | $16.25 | $1.63 |
Uncached prices, before any regional-processing uplift and before tax. Prompt caching cuts the website figure substantially when instructions and retrieved passages repeat; our AI agent cost guide shows that arithmetic. Prices change, so check OpenAI’s pricing page before budgeting.
Two things follow. For most business integrations the API bill is small next to the build and the staff time saved, so choose the model on quality, measured on your evals, not on price alone. And the biggest lever is design: retrieval that sends five relevant passages instead of a whole document, and Batch for anything that can wait a few hours.
Integration checklist: what a production build includes
- API keys on the server, one per environment, with a monthly spend limit on each.
- An organisation-owned OpenAI or Azure account; never a developer’s personal one.
- Data minimisation and masking before any call; no secrets or passwords in prompts.
- Structured outputs (JSON schemas) wherever the answer feeds another system.
- An eval set of real cases, run on every prompt, retrieval or model change.
- Logging and tracing with personal data redacted, kept where your policy says.
- Rate limits per user, and a fallback message when the API is slow or down.
- Prompt-injection defences: content from emails, web pages and uploads is treated as data, never as instructions, and never widens what a tool may do.
- Pinned model versions and an owner for the deprecation calendar.
- A human handover path, and a way for users to flag a wrong answer.
Questions to ask a ChatGPT integration company
- Which of the four routes are you proposing, and why not the others?
- Whose OpenAI or Azure account will this run in, and who holds the keys?
- How will you measure accuracy before launch, and what score must it reach?
- What will it cost per month at our volume, with the arithmetic shown?
- How is personal data minimised, and which retention settings will you request?
- What happens when OpenAI retires the model this is built on?
- Could we switch to another model provider later without a rebuild?
- Who owns the code, prompts, eval set and logs at the end?
Typical ChatGPT integration projects
Illustrative examples, not client case studies. Each starts small and grows only when the evals support it.
| Who | What we integrate | How it usually works |
|---|---|---|
| D2C or e-commerce brand | Product and order questions on the website | Retrieval over the catalogue and policies, order look-up through a read-only tool, handover to the support inbox |
| B2B sales team | Drafts and summaries in Gmail or Outlook, notes into the CRM | An add-in drafts from the thread and CRM record; the salesperson edits and sends; summaries are written back to the deal |
| Operations team with spreadsheets | Classification and clean-up of product or vendor data | Batch jobs over thousands of rows, results written to a new column with a confidence flag for review |
| Company already on ChatGPT | An internal app so staff can query orders, stock or tickets from ChatGPT | An MCP server with read-only tools first, sign-in per user, and logging of every call |
| Enterprise on Azure | An internal assistant on company documents | Azure OpenAI in South India with private networking, document retrieval and access rules matching existing permissions |
A first ChatGPT integration, step by step
A fixed-scope project, usually four to six weeks, that ends with a working integration and the numbers to decide what comes next.
Pick the route and write the rules
Users, data sources, what the model may and may not do, which route fits, and a cost estimate at your volume.
Eval set and baseline
Fifty to a few hundred real examples with the answers you expect, and how long people take today.
Build with guardrails
Server-side integration, retrieval or tools, spend caps, logging and the handover path, tested daily against the evals.
Pilot and decide
A limited group uses it with feedback buttons. You get accuracy, cost and time-saved figures, and a recommendation.
AI integrations we run ourselves
One works inside ChatGPT today. The other runs on the Claude API, which is where our production habits (evals, traces, spend caps) come from.
Radiatus MCP Server
Free public MCP server: Radiatus knowledge and developer tools for AI assistants Our free, public, read-only MCP server at radiatus.com/mcp. It works inside ChatGPT as well as Claude, VS Code and Cursor, with scoped tools and usage logging. Visit radiatus.com
Appointment Reminder App
WhatsApp & SMS reminders that grow into an AI inbox Its AI inbox and agent answer replies to WhatsApp and SMS reminders on the Claude API, with human handover. The same integration patterns apply on OpenAI. Visit appointmentreminderapp.comWhat we build ChatGPT integrations with
Versions and product names as of November 2026. We pin them per project.
OpenAI Responses API
The main API for text, tools, file search and structured outputs. The Assistants API it replaced was shut down on 26 August 2026.
Apps SDK and MCP
Apps that run inside ChatGPT are built on the Model Context Protocol, so one server can also serve Claude and IDEs.
Azure OpenAI in Microsoft Foundry
OpenAI models under Microsoft’s terms, deployable in the South India region where the model is offered.
PostgreSQL with pgvector, or a managed vector store
Hybrid search over your documents with metadata filters, so answers cite the right source.
Eval harness and tracing
Real-case evals in CI, OpenTelemetry traces with personal data redacted, and cost per request.
Tools we use alongside it
- OpenAI Responses API
- Batch API
- Apps SDK
- MCP
- Azure OpenAI
- Microsoft Foundry
- pgvector
- OpenTelemetry
- Langfuse
- Google Apps Script
- Office add-ins
- n8n
Facts, versions and vendor details checked November 2026.
Guides and related pages
Guides
Technologies
Glossary
Where we work
Questions about ChatGPT integration
They connect OpenAI’s models to your own software and data: an assistant on your website or in your product, drafting and summarising inside Gmail, Outlook or a CRM, bulk processing of spreadsheets and documents, or an app that lets staff use company systems from inside ChatGPT. A good integrator also handles security, data residency, cost limits and testing.
Add a chat widget that talks to your own server, and have the server call OpenAI’s Responses API with retrieval over your pages and documents. Never put the API key in the website’s code. Add citations, a polite “I don’t know” path, rate limits, a spending cap and a handover to a person, and test it on real questions before launch.
Yes. For Excel or Google Sheets, a custom function handles small jobs and OpenAI’s Batch API handles thousands of rows at half price. For Gmail or Outlook, an add-in drafts or summarises the open thread for the person to send. For Zoho, HubSpot or Salesforce, a webhook writes summaries and suggested next steps into the record.
You can build your own business assistant on the WhatsApp Business Platform using OpenAI’s models, answering questions about your products, orders and bookings. What Meta’s terms do not allow, since 15 January 2026, is offering a general-purpose AI assistant on the platform. Opt-in, message charges and a route to a human still apply.
Not by default for the API. OpenAI says API data has not been used for training since 1 March 2023 unless you opt in. Abuse-monitoring logs are kept for up to 30 days, and zero data retention is available for eligible endpoints with OpenAI’s approval. Set these options before any real data flows.
Yes, for eligible customers. OpenAI lists India among its data residency regions for the API, subject to approval of abuse-monitoring controls and a modified retention amendment, and offers India residency for ChatGPT Enterprise and Edu. Regional-processing endpoints carry a 10% price uplift on models released on or after 5 March 2026.
The build is quoted after scoping; Radiatus does not publish rates. Running costs are usually modest: in our illustrative examples, 2,000 email drafts a month cost about $11 on GPT-6.1 Terra, and a website assistant handling 8,000 conversations about $208 before caching. The API bill goes to your own OpenAI or Azure account.
Yes. OpenAI shut the Assistants API down on 26 August 2026. We migrate the integration to the Responses API, re-create its tools and files, add an eval set built from real past conversations, and pin model versions so the next platform change is a planned upgrade rather than an outage.
Have a ChatGPT integration in mind?
Tell us who will use it, which systems it must reach and roughly how many requests a month. A senior engineer will reply with the route we would choose, the cost at your volume and what we would test first.