A custom GPT for your business is a tailored version of ChatGPT trained on your company’s data, tone, and workflows. You can build one using OpenAI’s GPT Builder (no code, free on Plus plan) or the OpenAI Assistants API (requires development, ₹50,000–₹2,00,000 to build and integrate).
GPT Builder vs Assistants API: Two Paths to Custom AI
OpenAI provides two distinct routes to building a custom GPT for your business, and choosing the right one depends on your technical resources and customisation requirements. GPT Builder is a no-code interface available to ChatGPT Plus subscribers (approximately ₹1,700/month) that lets you create a custom GPT by describing its purpose, uploading documents to its knowledge base, and configuring its behaviour through a conversation with the GPT Builder itself. The result is a shareable custom GPT that anyone with the link (and optionally a ChatGPT account) can use. For small Kerala businesses with simple, internal-facing use cases, this is a powerful and accessible option that can be configured in under two hours.
The Assistants API path offers far greater integration flexibility. Instead of living inside ChatGPT’s ecosystem, an Assistants API implementation embeds your custom AI directly into your website, WhatsApp, mobile app, or any other channel you control. Your customers never need to have a ChatGPT account — they interact with your branded AI interface. The technical requirement is a backend server that calls the OpenAI Assistants API, manages thread creation for each conversation, and handles tool calling for integrations. Development cost ranges from ₹50,000 for a basic web chat integration to ₹2,00,000+ for a multi-channel deployment with CRM integration.
The decision between the two paths comes down to three questions. First: do your users have (or want to create) ChatGPT accounts? If not, you need the Assistants API. Second: do you need the AI to access live data — current inventory, CRM records, booking availability — rather than just static uploaded documents? Live data access requires the Assistants API with function calling. Third: is this for internal team use or customer-facing deployment? Internal use cases (HR policy lookup, sales proposal drafting, code review assistance) are often well-served by GPT Builder, while customer-facing applications benefit from the controlled, branded experience the Assistants API enables.
Building With GPT Builder: The No-Code Approach
Starting with GPT Builder is straightforward. Log into ChatGPT Plus, navigate to ‘My GPTs’ in the sidebar, and click ‘Create.’ The GPT Builder will ask you what your GPT should do — describe your business, its purpose, and the types of questions it should answer. The Builder immediately suggests a name, profile image, and basic system prompt, which you can accept or modify. This conversational setup process typically takes 20–40 minutes for a simple business use case.
The knowledge base upload is where GPT Builder adds genuine business value. Upload your product catalogues, service guides, FAQ documents, pricing sheets, and any other reference materials you want the GPT to draw from when answering questions. Supported file types include PDF, Word, Excel, and plain text. The GPT uses retrieval over your uploaded files to ground its responses in your actual business information, significantly reducing hallucination compared to a generic ChatGPT conversation. For a Kerala Ayurveda clinic, uploading treatment protocols, contraindication guidelines, and patient preparation instructions creates a genuinely useful patient information assistant.
Configuring actions — the equivalent of tool calling in GPT Builder — allows your custom GPT to call external APIs to retrieve live data. With some JSON schema configuration (no coding required, but some technical comfort needed), you can connect your GPT to a Google Calendar to check availability, a Notion database to look up project information, or a custom API endpoint to retrieve product inventory. This actions capability is what separates a powerful GPT Builder deployment from a basic document chatbot. For Kerala businesses with existing web APIs or Zapier integrations, actions unlock significant automation possibilities without moving to the full Assistants API.
Writing an Effective System Prompt for Your Business
The system prompt is the most important configuration element in any custom GPT — more impactful than the documents you upload or the actions you configure. A well-written system prompt defines: the AI’s identity and persona (name, role, tone of voice, communication style), the scope of topics it should and should not address, specific response format rules (use bullet points for lists, always include price in INR, mention availability as a next step), escalation instructions (when to recommend speaking to a human), and any compliance requirements (do not make medical diagnoses, always recommend consulting a doctor).
For a Kerala jewellery brand’s custom GPT, an effective system prompt would specify: ‘You are Lakshmi, the AI assistant for [Brand Name]. You help customers in both English and Malayalam. Always quote gold purity (22K or 18K) when discussing any jewellery piece. Provide prices in INR. For custom order inquiries, always collect: the occasion, budget, preferred metal type, and the customer’s phone number for our design team to follow up within 24 hours.’ This level of specificity produces consistent, on-brand responses across all customer interactions.
System prompt testing is non-negotiable before deployment. Test every common customer query type, several unusual edge cases, and deliberately awkward inputs. Check that the GPT follows your formatting instructions consistently, stays within scope boundaries, and handles queries in both Malayalam and English as intended. Common system prompt failures include the GPT inventing prices for products not in its knowledge base (fix with: ‘Never quote prices for items not listed in your knowledge documents’), and the GPT giving medical or legal advice when the business does not want it to (fix with explicit prohibitions). Budget 3–6 hours for systematic prompt testing before sharing the GPT with customers.
The Assistants API: When You Need Custom Integration
The OpenAI Assistants API introduces three components that enable custom integration: Assistants (the configured AI with system prompt and tools), Threads (individual conversation sessions), and Messages (individual turns within a thread). Your server creates a thread when a customer starts a conversation, adds their message as a thread message, runs the assistant on the thread, and polls for the response. This server-side architecture gives you complete control: you can store all conversation data in your own database, integrate with any of your existing systems through function calling, and present the AI through any interface you build.
Function calling — the mechanism by which the Assistants API triggers your custom code — is where the real business logic lives. You define functions that the assistant can call: check_inventory(product_id), create_booking(date, time, patient_name), get_order_status(order_number), send_whatsapp_confirmation(phone, booking_details). When a customer asks ‘Is the Kanchi silk saree in red available?’, the assistant calls your check_inventory function with the appropriate parameters, receives the live inventory response from your database, and incorporates that real data into its reply. This creates genuinely useful AI — not just a sophisticated text generator.
Development timeline for an Assistants API project depends on the number of function integrations. A basic web chat with no function calling takes 2–3 weeks. Each additional function integration (CRM lookup, booking system, inventory check) adds 1–2 weeks. A full-featured customer support system with 4–5 functions, conversation memory, and a management dashboard for your team takes 6–10 weeks. Work with experienced AI services consultants to scope function requirements precisely before development begins — function scope creep is the primary source of budget overruns in Assistants API projects.
Cost Breakdown and Ongoing Maintenance
GPT Builder costs are straightforward: ChatGPT Plus subscription at approximately ₹1,700/month. If you want to share your custom GPT with customers who do not have ChatGPT accounts, you need a ChatGPT Team plan at approximately ₹4,000–₹5,000/month for up to 10 users. There is no additional per-query cost within your subscription’s usage limits. For internal team use where everyone has a Plus subscription, GPT Builder is the most cost-effective option available. The primary limitation is that all users must have ChatGPT accounts and be comfortable using the ChatGPT interface.
Assistants API costs are usage-based: you pay for tokens consumed at GPT-4o rates (approximately ₹0.40–₹2.00 per 1,000 tokens depending on context length). For a Kerala business handling 1,000 customer conversations per month averaging 2,000 tokens each, monthly API cost is approximately ₹800–₹4,000. Add hosting for your server (₹1,000–₹3,000/month on AWS or DigitalOcean) and you have a predictable monthly operating cost that scales linearly with usage volume — no surprise charges if a viral moment sends 10x normal traffic.
Ongoing maintenance for a custom GPT involves two activities: knowledge base updates and system prompt refinement. Knowledge base documents should be updated whenever your product catalogue, pricing, policies, or service offerings change. System prompt updates are needed when you identify response patterns that need correction or when you want to add new response capabilities. Budget 2–4 hours per month for maintenance of a moderately active GPT Builder deployment, or ₹5,000–₹15,000/month in developer time for a maintained Assistants API deployment. The businesses that get the most value from custom GPTs are those that treat maintenance as an ongoing investment rather than a one-time setup task. AI & Machine Learning retainer arrangements handle this efficiently.
Frequently Asked Questions
Is a GPT Builder custom GPT secure enough for confidential business data?
GPT Builder custom GPTs store your uploaded documents within OpenAI’s infrastructure, which means the data is subject to OpenAI’s privacy policies. For truly confidential business data — client contracts, financial records, proprietary formulas — avoid uploading to GPT Builder. Instead, use the OpenAI Assistants API with your own database backend, giving you full control over where your data lives.
How many documents can I upload to a custom GPT knowledge base?
GPT Builder currently allows uploads up to 512 MB total per custom GPT, supporting PDF, Word, TXT, and code files. For the Assistants API, you can attach up to 20 files per assistant with a 512 MB limit per file. For larger knowledge bases — such as an Ayurveda clinic’s complete treatment database — a RAG system with a vector database is more appropriate than file uploads.
Can a custom GPT handle Malayalam language queries from Kerala customers?
GPT-4o, which powers custom GPTs, handles Malayalam reasonably well for common business queries — product information, appointment scheduling, FAQ responses. Pure Malayalam text generation is less reliable than English. For a Kerala business, the most effective approach is a bilingual custom GPT that responds in the language the customer uses, with English as the default.