Can GPT-3 help a small business answer customer email?

For a small business, customer email is a steady drain on time. Most messages ask the same handful of things: opening hours, delivery delays, returns, a booking change. The answers are known, but each reply still has to be written, and written kindly, often at the end of a long day.

GPT-3, the language model OpenAI offers through its API and its Playground (a web page where you can type prompts and see results), is well suited to the writing part. Since January, OpenAI's default models are the InstructGPT versions, which are trained with human feedback to follow plain written instructions. That makes it practical to say "write a reply to this customer" and get a usable draft. The part GPT-3 cannot do on its own is know your policies, check your records or take responsibility for what is promised. That is where you come in.

What GPT-3 is good at, and what it is not

GPT-3 writes fluent, polite text quickly, adapts its tone on request, and can turn a short note of facts into a complete, friendly message. It is good at the shape of a reply.

It does not know your business. If you ask it about your return policy without supplying the policy, it will produce something that sounds like a typical return policy, which may promise a refund you do not offer. OpenAI says plainly that even its improved InstructGPT models can still make up facts and produce biased or toxic text, though less often than the original GPT-3. A draft is only as accurate as the information you put in the prompt.

A worked example: a draft-reply prompt

Here is a prompt for a small bakery that ships gift boxes. It gives the model three things: the role, the relevant policy text copied from the business's own documents, and the customer's message with personal details removed.

You write email replies for a small family bakery that ships gift boxes. Write a warm, short, professional reply to the customer message below. Use only the policy text provided. If the policy does not answer the question, say that a team member will follow up, and do not invent details. End with "[BAKERY NAME] team".

Policy:
"""
- Gift boxes ship Monday to Wednesday so they arrive fresh.
- Orders placed after Wednesday noon ship the following Monday.
- If a box arrives damaged, send a photo within 48 hours for a free replacement.
"""

Customer message:
"""
Hi, I ordered a gift box on Thursday for my mother's birthday on Saturday and it hasn't shipped. Also, last month's box arrived crushed. Can you help? - [CUSTOMER]
"""

Reply:

Notice the details. "Use only the policy text provided" and "do not invent details" steer the model away from guessing. The triple quotes mark where the policy and the message begin and end. The placeholder [CUSTOMER] replaces the real name. And the instruction covers what to do when the policy has no answer, which is where invented promises usually come from.

A good draft from this prompt would explain that Thursday orders ship the following Monday, apologize that the box will miss Saturday, and explain how to get a replacement for a damaged box. It is not yet ready to send. Is last month's damage still within the 48-hour window? Probably not, and only a person with the order history can decide whether to make an exception.

A privacy check before you paste

Customer emails contain personal information: names, addresses, phone numbers, order numbers, sometimes health or payment details. Before pasting a message into GPT-3, replace those with placeholders such as [CUSTOMER] or [ORDER], and add them back when you send.

There is a concrete reason. OpenAI's API terms, updated in January 2022, give the company a license to use and store the content submitted to the API in order to provide and improve its services, and they require developers to follow privacy laws such as Europe's GDPR and California's CCPA and to have the rights to the content they send. Health information covered by the US HIPAA law is excluded without a separate written agreement. Your own local privacy laws apply too. The simple habit of removing personal details covers most of these concerns at once.

Where should a person stay in the loop?

At the send button, every time. "Human in the loop" means a person reviews the system's output before it has any effect. For customer email, that means GPT-3 drafts and a person approves. Read each reply against your records: dates, prices, order details and anything that sounds like a promise. OpenAI's own publication policy, written for people sharing text made with its models, makes a similar point: a human must take ultimate responsibility for what is published.

A person should also handle certain messages from the start. Complaints from upset customers, refund disputes, anything legal or medical, and messages from someone in distress deserve a reply written by a human, perhaps with GPT-3's help for wording but not for judgment.

Finally, a person should own the policy text itself. The model's drafts can only be as good as the rules you give it, so keep that short policy file up to date. Used this way, GPT-3 handles the typing, and you keep the decisions and the relationship with your customers.