Ten days ago, at its first DevDay, OpenAI announced GPTs: custom versions of ChatGPT that anyone can build without code, with their own instructions, their own uploaded files and optional tools. Since the announcement, many small business owners have expressed interest in having one. OpenAI even paused new ChatGPT Plus sign-ups this week because demand exceeded its capacity. A GPT Store is expected later this month.
This development builds on earlier steps. In July, OpenAI added custom instructions, which allow users to tell ChatGPT once who they are and how they want it to answer. Developers could already go further with the API and a system prompt and build their own assistant from scratch. GPTs make that capability available through a form anyone can fill in.
What is a custom AI assistant?
A custom GPT combines three elements: a general model, a detailed set of instructions defining how it should behave, and a collection of your own material that it can consult.
No additional training takes place and no new model is created. It is the same model everyone else uses, configured with your guidelines and your reference documents.
This may appear modest, but in practice it makes a significant difference. What usually limits a generic chatbot in a business setting is that it does not know your prices, your policies, your tone, your customers or your way of working. Once it has that information, the same model provides answers your business can use.
One practical limitation applies at the moment: to use a GPT, people need a paid ChatGPT account. A GPT is therefore well suited to internal use by your team, but not yet to a public help desk for customers.
We have been writing long instructions for language models since last fall, when we began using them as game masters for solo role-playing games with ChatGPT RPG. That work has been our most valuable source of knowledge about writing instructions. We are now gathering our notes on business GPTs under GPT Navigator Pro.
What should you give the assistant to work with?
The quality of a custom assistant is determined almost entirely before anyone asks it a question, by the documents and the instructions it receives.
The documents are important: your price list, your FAQ, your cancellation policy, your service descriptions and a few examples of emails that represent your business well. They should be current versions, not a 2019 file stored in a folder. If the files are disorganized or contradict one another, the assistant will reproduce that confusion in its answers, with the same apparent confidence.
The instructions are even more important, and this is the step most often rushed. Many people write two lines, such as "You are a helpful assistant for Bob's Marina. Be friendly.", and then find that the assistant does not behave as expected.
Good instructions read like a handover document for a new employee who is highly capable but knows nothing about your business. They specify:
- who the customers are and what they usually want
- what the assistant may answer, and what it must hand off to a person
- what it must never do: quote a price that is not in the file, promise dates, give medical or legal advice
- the tone, with one or two examples
- what to say when it does not know
The last point is the most important instruction of all. An assistant that says "I'm not sure, let me get someone to call you" is useful. An assistant that guesses creates a risk for the business.
In short, a custom GPT performs only as well as the instructions written for it.
How should you test a custom assistant before using it?
Test it with the real questions your team and your customers ask, not only the questions you expect. Assistants tested only with anticipated questions usually answer them well and are then launched.
Real users, however, phrase their questions differently: half in French and half in English, with typos, with three questions in a single message, or about a service you stopped offering two years ago.
We recommend the following steps before any deployment:
- Collect twenty or thirty real questions: those in your inbox, those staff hear on the phone, and the unusual ones.
- Run all of them and read every answer.
- When an answer is wrong, correct the instruction or the document that caused the error rather than the answer alone.
- Run the full set again.
This process takes time, and it is what distinguishes a dependable tool from a demonstration.
Repeat it each time you change the instructions, because a correction in one area can introduce an error in another.
What are the likely weak points of custom assistants?
The technology is still new. Based on our knowledge of these models, we expect the same few weak points to appear repeatedly.
Outdated information. Your prices change, but the assistant does not, because nobody updated the file. A specific person should be responsible for keeping it current.
Overreach. When a question falls slightly outside its material, the assistant tends to fill the gap with a plausible guess instead of stating that it does not know. Better instructions reduce this behaviour, but they do not eliminate it.
Premature trust. In the first week, everyone checks everything. By the third month, checks often stop, and that is when an occasional wrong answer can go unnoticed.
Unsuitable tasks. Some tasks should not be entrusted to an assistant. Any decision where an error could cost someone money, safety or dignity requires a person to make the call.
How do you keep control over what is sent to clients?
The arrangement we recommend to most small businesses is straightforward: the assistant drafts, and a person reviews and sends.
For internal work, the assistant can operate more freely: summarizing notes, drafting a plan, organizing ideas, or helping someone find a policy quickly. Errors in that context are caught by your own team.
Anything going outside the business, such as a quote, a reply to a complaint or a message to a supplier, should be reviewed by a person. That review takes about thirty seconds and prevents the kind of error that can cost you a customer.
If you are considering AI chatbots and assistants built on your own material, make sure they are tested thoroughly before anyone relies on them. Most of the work involved consists of writing.