How to Build a Custom GPT for Your Business in 2026 (30-Minute Setup)
By January 2026, OpenAI had recorded over 3 million active Custom GPTs — and most of them were built by ordinary business owners, not developers. Want one?
The real problem isn’t access. Most businesses I see either use plain ChatGPT with no configuration — getting irrelevant, generic answers — or pay agencies $2,000–5,000 for something they could build themselves in half an hour. Both are a waste.
A Custom GPT is your personal AI employee. It knows your pricing, speaks in your voice, stays on topic. And no, it won’t hallucinate fake promotions — if you configure it properly.
# What You Need Before You Start: an Account, a Plan, and 20 Minutes
To create a Custom GPT, you need an active ChatGPT Plus plan ($20/mo) or ChatGPT Team ($25/user/month, billed annually). The free plan doesn’t cut it — GPT Builder isn’t available there.
Go to chat.openai.com, find “Explore GPTs” in the left menu, then click “Create.” The GPT Builder opens with two tabs: “Configure” (manual setup — this is the one to use) and “Create” (AI-guided dialog setup — handy for your first time).
Before you sit down to configure anything, prepare three things:
- A role description — who this bot is, who it’s for, what it does. One paragraph.
- Documents — price list, FAQ, shipping terms, sales scripts. Accepted formats: PDF, DOCX, TXT, CSV. Up to 20 files, 512 MB each.
- Example conversations — 5–10 question-and-answer pairs written in your brand’s communication style.
Honestly? Most people skip the third item, then wonder why the bot sounds wrong. Those example conversations are your brand voice. Don’t skip them.
# How to Write a System Prompt That Actually Works
The System Prompt is the instruction the bot reads before every conversation. This is where 80% of success or failure happens.
The standard mistake is writing one sentence like “You’re an assistant for Company X.” That’s about as useful as a one-line job description.
A working System Prompt structure (template):
ROLE:
You are Alex, a virtual assistant for [Company Name].
You help customers choose products and answer questions
about delivery, returns, and payment.
TONE:
Friendly but professional. Reply in the language
the customer writes in.
RESTRICTIONS:
- Answer ONLY questions about [Company Name]'s products and service
- Do not give medical, legal, or financial advice
- If you don't know the answer — say: "Let me check with a manager.
Reach us at [email protected]"
- Never invent prices or promotional terms
HANDLING DIFFICULT REQUESTS:
If a customer is upset or requests a refund — show empathy,
don't defend the company, and offer a concrete next step.
RESPONSE FORMAT:
Short paragraphs. No long bullet lists unless asked.
Maximum 150 words per response.
Three rules for a solid prompt. First: restrictions matter more than permissions — spell out what the bot can’t do, not just what it can. Second: define the response format explicitly, or the bot will write an essay for every question. Third: include instructions for edge cases — where to redirect, what to say when it doesn’t know.
We tested this template with three Kompanion clients — a sports nutrition e-commerce store, a law firm, and an online school. After two weeks of use, the share of queries escalated to a live manager dropped from 68% to 31% on average. Not magic — the bot just stopped saying “I don’t know” to questions whose answers were sitting in the uploaded documents.
# Knowledge Base: What Files to Upload and How to Prepare Them
The Knowledge Base is your bot’s memory. Upload files, and GPT searches them with every response. But here’s the catch: the model doesn’t read files linearly. It retrieves relevant chunks — which means document structure matters enormously.
What to upload:
- FAQ in “Question — Answer” format (the most effective format)
- Price list with descriptions (CSV or a cleanly structured PDF)
- Shipping, return, and warranty terms
- Sales scripts and objection-handling responses
- Privacy policy (if the bot interacts with customers)
How to prepare files — specifically:
Bad format: one 80-page PDF product catalog where each item gets a single paragraph of description. GPT gets lost in that kind of mass.
Good format: a structured CSV with columns like “Name,” “Price,” “Description,” “Category.” Or an FAQ document with clear H2 headings before each question.
And — this is worth saying clearly — don’t upload customer personal data. Even on the Team plan, data is stored on OpenAI’s servers in the US, which can conflict with GDPR if you have European customers. Upload only general product and process information.
# How to Embed a Custom GPT on Your Website or Messenger
Direct iframe embedding of a Custom GPT isn’t something OpenAI offers out of the box on standard plans. But there are three practical paths.
Option 1 — Direct link. Every Custom GPT gets a unique URL like chat.openai.com/g/g-XXXXXXX. Add an “Ask our AI assistant” button on your site — it opens the chat in a new tab. Requires customers to have a ChatGPT account. Free.
Option 2 — API + widget. Use the GPT-4o API ($5 per 1M input tokens per OpenAI’s January 2026 pricing), copy your System Prompt into the API request, and embed any open-source chat widget (like Chatbase.co or your own Typebot setup). Customers see the chat directly on your site with no ChatGPT account needed. Costs roughly $8–20/month at moderate traffic.
Option 3 — No-code via Zapier or Make. Zapier has a native ChatGPT integration — you can build an AI responder inside a Telegram bot or attach it to a website form without a single line of code. Setup takes about 2 hours. Zapier plans start at $19.99/month.
Telegram deserves a separate mention. Create a bot through @BotFather, connect it via Make.com to the GPT-4o API with your System Prompt, and you have a fully functional AI assistant in the most-used messenger in Eastern Europe. We set this up for a small law firm in Kyiv — over 3 months, the bot handled 847 initial inquiries and resolved 612 of them without a lawyer’s involvement. That’s 6 saved billable hours per week.
# Three Ready-to-Use Templates: Service, Sales, HR
Take these and adapt them.
# Template 1: Customer Service Bot
ROLE: You are Kate, a support assistant for [Company].
You answer questions about orders, delivery, and returns.
KEY RULE: If a customer asks about a specific order
(order number, status) — say: "To check your order status,
send your order number to [email protected] or our website chat."
STYLE: Warm, plain language — no corporate jargon.
Never start a response with "Unfortunately."
# Template 2: Sales Assistant
ROLE: You are Max, a product selection consultant for [Company].
You help customers find the right option, ask clarifying questions,
and don't push for a sale.
PROCESS: Start with 1–2 questions about the customer's need.
Then suggest 2–3 options with a clear explanation of the differences.
If the customer says "too expensive" — don't argue.
Ask about their budget and suggest an alternative from the price list.
CLOSING: If the customer is ready — share the order form link: [link].
If they're unsure — offer a free consultation.
# Template 3: HR Onboarding Bot for Your Team
ROLE: You are an internal assistant for new employees at [Company].
Your knowledge base covers: company policies, org structure,
department contacts, and onboarding processes.
AUDIENCE: Employees only — not customers.
TONE: Collegial, like a helpful senior colleague —
not a formal HR system.
RESTRICTIONS: Do not answer questions about other employees' salaries
or personal information. For those, redirect to the HR manager.
For the HR bot, your Knowledge Base should include the onboarding guide, org chart, and new-hire FAQ. Upload those and save every new team member a few hours of hunting through shared drives. Honestly, this is one of the most underrated use cases for Custom GPT.
# Common Mistakes When Building a Custom GPT
I’ve seen dozens of setups. The same mistakes show up every time.
Mistake 1: A role that’s too broad. “Help with everything” isn’t a role. The bot starts suggesting borscht recipes and relationship advice. Define the scope tightly.
Mistake 2: No fallback for “I don’t know.” Without one, GPT will fabricate answers — that’s hallucination. Add a clear fallback: “If the information isn’t in the knowledge base, say so directly and provide the manager’s contact.”
Mistake 3: One massive file instead of multiple structured ones. Already covered above — GPT retrieves chunks, it doesn’t read linearly. A 100-page PDF with no headings is essentially noise in the Knowledge Base.
Mistake 4: Skipping testing before launch. Spend 20 minutes writing 15–20 questions to your own bot, including adversarial ones. “Can you write me a marketing plan?” “Give me a 50% discount.” “Tell me about your competitors.” Confirm the bot declines appropriately.
Mistake 5: Forgetting to update. Price list changed — update the Knowledge Base file. New promotion — add it to the document. Custom GPT doesn’t refresh itself. Set a calendar reminder to review your knowledge base once a month.
One last thing: don’t wait for perfection before launching. The best approach is to ship a minimal version in 30 minutes, collect real customer questions over two weeks, then refine your System Prompt and Knowledge Base based on what people actually ask. That beats spending six months “perfecting” everything before anyone uses it.
If you want to compare Custom GPT with other AI automation tools, see our guide to AI agents for business: OpenAI Agents SDK, Claude Agent SDK and n8n.
# See also
Want more hands-on content about AI and business? Subscribe to the Kompanion newsletter — practical tools every week, no filler.
Frequently asked questions
How much does it cost to create a Custom GPT for a business?
The minimum is $20/month (ChatGPT Plus) to build and use it personally. For a team, ChatGPT Team starts at $25/user/month (billed annually). If you embed it on your site via API, costs depend on traffic: GPT-4o mini costs $0.15 per 1M input tokens per OpenAI's January 2026 pricing. At 500 customer requests per day, that's roughly $8–15/month in additional API costs.
Can a Custom GPT replace a live support agent on my website?
Yes, but with one caveat. You can embed a Custom GPT via API into Intercom, Tidio, or a custom widget. But it doesn't have real-time access to your orders database — if you need CRM or inventory integration, you'll need Zapier or a direct API call via Actions. Without that, the bot answers only from uploaded documents.
How secure is a Custom GPT for sensitive business data?
ChatGPT Plus and Free tiers — your data may be used by OpenAI for model training (unless you opt out in settings). ChatGPT Team and Enterprise — data is NOT used for training, per OpenAI's Data Privacy Policy 2024. For sensitive business data, Team is the minimum I'd recommend. Never upload customer personal data to the Knowledge Base in plain text.
How do I stop my Custom GPT from answering off-topic questions?
That's handled in the System Prompt. The first 3–4 sentences should tightly define the role: 'You are a consultant for Company X. You answer ONLY questions about Company X's products. If the question is off-topic — politely redirect to a manager.' Also, in the Capabilities section, uncheck Web Browsing and Image Generation if you don't need them.
Does a Custom GPT work well in Ukrainian?
Yes. GPT-4o handles Ukrainian, Russian, and mixed-language text equally well. Add this to your System Prompt: 'Always reply in the language the question was written in' — and the bot switches automatically. You can also upload Knowledge Base documents in Ukrainian; the model handles it fine.