Creating a Custom GPT: How to Build Your Own AI Assistant

A Custom GPT is no black magic. But a generic, one-size-fits-all response tool no longer cuts it for most businesses. In this guide, I'll walk you through, step by step, how to build a specialised AI assistant that actually solves a recurring problem. As a practical example, I'm using a customer service GPT, because that's the use case most SMEs try first.

What Is a CustomGPT?

Custom GPTs are personalised versions of ChatGPT that OpenAI introduced in November 2023. You give them a name, define their instructions, upload specific knowledge, and decide which tools they're allowed to use, such as web search, image generation, or data analysis (OpenAI, 2023). The difference from a normal chat: a Custom GPT remembers its role, its expertise, and its tone, so you don't have to explain that again in every conversation (OpenAI, n.d.-c).

What started as a simple idea has since become a serious business tool. According to OpenAI's first "State of Enterprise AI" report, weekly usage of Custom GPTs and Projects grew roughly 19-fold in 2025, and they now account for about 20 percent of all Enterprise messages. Spanish bank BBVA alone runs more than 4,000 of its own GPTs (OpenAI, 2025). Uptake is growing in Switzerland too: the share of SMEs using AI tools rose from 22 to 34 percent between 2024 and 2025, with customer relationship management among the areas seeing the sharpest growth (AXA Schweiz, 2025).

CustomGPT: Why Customer Service Is a Good Place to Start

A Custom GPT pays off wherever you're handling the same type of request over and over, and re-supplying the same context every time (OpenAI, n.d.-c). Customer service is the textbook example. A practical case from Swiss SME consulting illustrates this well: at a business with three employees, roughly 40 percent of some 80 weekly support emails came down to the same five questions, about delivery times, returns, payment methods, product availability, and pickup options (Massive Mind, 2026). That kind of repetition is exactly the use case a Custom GPT is worth building for, not the creative one-off.

Other typical use cases for Custom GPTs include content creation in your own tone of voice, data analysis built around your KPIs, an internal knowledge base for your team, or educational content adapted for teaching. The customer service case is a particularly good starting point because success there is measurable: fewer follow-up questions, shorter response times, less burden on staff.

Practical Example: Building a Customer Service GPT Step by Step

Let's say you run a small online shop and want to stop answering the same recurring questions about shipping, returns, and payment by hand every time. The plan: a Custom GPT that you share via a link, for example on your contact page or in the order confirmation, so customers ask it directly instead of writing you an email. Here's how you do it.

• Get access. You need a ChatGPT Plus, Team, or Enterprise account. Go to "Explore GPTs" and select "Create" to open the GPT Builder (OpenAI, n.d.-a).

• Create the GPT. In the Builder you'll see two tabs: "Create" is the conversational mode, where you describe in your own words what you want your GPT to do, and ChatGPT drafts a first version for you, for example, "Create an assistant that answers questions about shipping, returns, and payment for my online shop." "Configure" is the manual view, where you enter the name, description, instructions, and knowledge base directly. For a customer service GPT, it's worth switching to "Configure" once you have a first draft, because it gives you more precise control than the conversational mode (OpenAI, n.d.-a).

• Set the name and description. The name should make immediately clear what your GPT does. "Sample Shipping Support Assistant" is more precise than "Customer Service Bot" and helps users place the tool right away (OpenAI, n.d.-a).

• Write the instructions, the heart of the matter. This is where you define role, task, tone, and boundaries as precisely as possible. For a customer service GPT, that includes: a clear role definition ("You answer questions about shipping, returns, and payment for [shop name]"), the permitted tone, what the GPT should explicitly not do (for example, promise discounts or cancel orders), and a clear escalation rule: for anything beyond the stored standard questions, or anything that turns emotional, the GPT refers the person to a real point of contact (OpenAI, n.d.-b).

• Build the knowledge base. Upload the documents your GPT should be able to draw on, such as your shipping terms, your returns policy, and an FAQ list. Important: knowledge is reference material, not behavioural rules. Tone and approach belong in the instructions, not in the uploaded files (OpenAI, n.d.-a).

• Set conversation starters. Sample questions like "Where's my order?" or "How do returns work?" show at a glance what the GPT is for, and make it easier to get started.

• Deliberately limit capabilities. Only turn on what you actually need. For a customer service GPT that handles customer data, it's usually wiser to switch off web browsing and image generation rather than leaving everything on by default.

• Test before customers see it. Write ten to fifteen realistic test questions along with the correct answers, and check whether your GPT answers them reliably. Only refine and go live afterwards (OpenAI, n.d.-c).

Don't Forget Data Protection

Once a Custom GPT is working with customer data, it stops being a purely technical project. Currently, only about a third of Swiss SMEs have clear internal rules on what data staff are allowed to enter into AI tools at all; among micro-businesses with fewer than ten employees, that figure drops to just 23 percent (AXA Schweiz, 2025).

For a customer service GPT, that means in practice: customers should know they're talking to a bot, you should document where support data is stored, and with US providers you'll typically need to sign a data processing agreement (Massive Mind, 2026). I'm not a lawyer. If your GPT handles sensitive or larger volumes of customer data, it's worth a brief check-in with a legal specialist in data protection before you go live.

Common Mistakes

Working with several Custom GPTs has surfaced a few recurring pitfalls:

• Vague instructions. "Be friendly and helpful" isn't enough. Define concretely what a good answer looks like.

• An overloaded knowledge base. More files aren't automatically better. Documents that are too large or contradictory tend to confuse the GPT rather than help it.

• No clear escalation rule. Without a defined handover point to a human, the GPT will eventually answer questions it shouldn't.

• Forgetting maintenance. A GPT whose knowledge base nobody updates will give outdated answers with complete confidence. Budget for ongoing maintenance, refining and updating; realistically, that's a few hours a month (digital M., 2026).

My Take

For me, this is the actual point of a customer service GPT: it's not about removing the human from the interaction, but about taking the time burden off the repetitive 40 percent of enquiries, so there's time again for the tricky, emotional, or sales-relevant cases. The AI takes on what's repeatable; the human stays responsible for everything that needs judgement and genuine contact.

What You Can Do Right Now

1. Start with a tightly scoped use case, not "a GPT for everything."

2. Write concrete instructions: role, tone, clear boundaries, an escalation rule to a human.

3. Keep the knowledge base lean and current, and budget for ongoing maintenance from the start.

4. Clarify transparency obligations and data storage before customers come into contact with the GPT.

5. Test with realistic questions before you take the GPT live.

About the Author

Harriet Moser is an AI expert specialising in branding and the founder of Ask Harriet, an AI training consultancy based in Switzerland. As part of her master's thesis, she developed the AI Brand Strategy Framework for integrating artificial intelligence across five levels of brand management. She specialises in helping businesses and individuals use AI systematically, ethically, and effectively.

For more information: www.askharriet.ch | LinkedIn: linkedin.com/in/harriet-moser

References

AXA Schweiz. (2025, October 8). AXA KMU-Arbeitsmarktstudie: Künstliche Intelligenz erobert Schweizer KMU. https://www.axa.ch/de/ueber-axa/medien/medienmitteilungen/aktuelle-medienmitteilungen/2025/20251008-kmu-arbeitsmarktstudie-2025-ki.html

digital M. (2026, July). KI-Chatbots für KMU: Anbieter, Kosten und was sie wirklich können. https://digitalm.ch/wissen/guides/ki-chatbots-kmu/

Massive Mind. (2026, March 18). KI im Kundensupport für KMU: Realistischer Einstieg. https://massivemind.ch/en/blog/ki-kundensupport-kmu

OpenAI. (2023, November 6). Introducing GPTs. https://openai.com/index/introducing-gpts/

OpenAI. (2024, January 10). Introducing the GPT Store. https://openai.com/index/introducing-the-gpt-store/

OpenAI. (2025, December 8). The state of enterprise AI | 2025 report. https://openai.com/business/guides-and-resources/the-state-of-enterprise-ai-2025-report/

OpenAI. (n.d.-a). Creating and editing GPTs. OpenAI Help Center. Retrieved July 12, 2026, from https://help.openai.com/en/articles/8554397-creating-and-editing-gpts

OpenAI. (n.d.-b). Key guidelines for writing instructions for custom GPTs. OpenAI Help Center. Retrieved July 12, 2026, from https://help.openai.com/en/articles/9358033-key-guidelines-for-writing-instructions-for-custom-gpts

OpenAI. (n.d.-c). Using custom GPTs. OpenAI Academy. Retrieved July 12, 2026, from https://openai.com/academy/custom-gpts/