Should a Custom GPT Answer Your Customers? Limits and Safer Options

No. A custom GPT shouldn't answer your customers: OpenAI is retiring custom GPTs on 11 December 2026, and they were never built for it. Customers must sign in to ChatGPT to use one, you can't see their conversations, and there's no handover to your staff. Use a website or messaging chatbot with human handover instead.

The better options are cheap and made for the job. Shopify Inbox now includes a free AI agent for shops. Intercom's Fin charges $0.99 per resolved outcome, Tidio's Lyro starts at $39 a month for 50 conversations, and Meta's Business Agent answers WhatsApp, Instagram and Messenger messages for roughly 4-5 cents a message. HubSpot's Customer Agent suits businesses already on its Service Hub Professional plan. Each gives you a log of what was said and a way to pass the customer to a person.

Follow me on Instagram@sagnikteaches

A custom GPT can still be the right idea behind the scenes, as a helper your staff use to draft replies, but even there it's being replaced: use a ChatGPT Project for that instead. For anything a customer types into, the tool needs features a GPT simply doesn't have.

Connect on LinkedInSagnik Bhattacharya

Why a custom GPT was never a customer channel

Four limits rule it out, retirement aside:

Subscribe on YouTube@codingliquids

• Customers have to sign in to ChatGPT. OpenAI's help pages say people can view a public GPT's page but must sign in before chatting. Many customers won't create an account to ask about your opening hours, and some shouldn't need to.
• You can't read the conversations. GPT builders can't see individual users' chats. If it tells a customer the wrong price, you find out when they quote it at you.
• There's no handover. A GPT has no staff inbox, no "talk to a person" button and no way to pass a conversation to you with its history.
• There are no customer-service guardrails. Its instructions are the only control. There's no escalation rule, no topic block, no business-hours logic and no report on what it couldn't answer.

Customer chat tools exist precisely to solve those four problems. That's the gap, not the quality of the AI underneath.

What a customer-facing assistant needs that a GPT lacks

The driving school's GPT experiment, and what replaced it

Here's an illustrative driving school: an owner who also teaches, four instructors and a part-time office manager, with about 250 enquiries a month from learners and parents about prices, availability, manual or automatic cars, test booking and the 48-hour cancellation rule. In spring, the owner built a custom GPT with the price list and FAQ and linked it from the website as "Ask our assistant".

Six weeks later the website's own analytics showed 210 clicks on the link. The owner had no way to know how many turned into conversations, because GPT builders can't see them, but the office manager noticed three things. Several parents rang to say they'd been "asked to make an account" and gave up. Two learners booked quoting "$35 an hour, the chatbot said", a price the school had raised in March; the owner had updated the website but not the GPT's file. And one parent asked it about a refund for a cancelled block of lessons, got a general answer about consumer rights, and escalated by email in some irritation. None of this would have been visible if the office manager hadn't been listening.

The replacement was a website chat tool with AI and handover, set to answer only five topics (prices, availability, car types, test booking steps and the cancellation policy) and to pass everything else to the office inbox. The first month's log showed what the GPT had hidden: 41% of chats arrived outside office hours, "can I pay in instalments?" was the fifth most common question and wasn't in any FAQ, and nine chats were about changing instructors, which the owner wanted to handle personally anyway. Sector detail on this kind of enquiry handling is in answering pupil and parent enquiries with AI at a driving school.

Here is the same refund question, handled by each (both illustrative):

GPT: "Under consumer protection law you may be entitled to a refund for services not provided. I'd recommend contacting the driving school directly to discuss your options."

Chat tool with handover: "Refunds for lesson blocks are handled by our office team, so I've passed your message to them with this chat attached. They reply within one working day, usually sooner. Is the best contact the email you're chatting from?"

The first is generic and slightly alarming; the second follows the school's process, sets an expectation and keeps the customer. The rules for when to hand over are set out in when an AI chatbot should hand over to a human.

Five safer options and what each costs

List prices in USD as of September 2026. Resolution-based prices count differently from vendor to vendor, so read each definition before comparing, and recheck each vendor's page on the day you sign up.

For the driving school's 250 enquiries a month, a quick sum shows how close the realistic options sit. Intercom with one Essential seat on monthly billing ($39) and Fin resolving 100 out-of-hours chats ($99) comes to about $138. Tidio with the Growth plan ($59 for 250 human conversations) and a 100-conversation Lyro block limited to evenings and weekends ($75) comes to $134. HubSpot would cost more in onboarding alone than a year of either, so it only makes sense for a business already paying for it. The detailed comparison of the first two is in Tidio vs Intercom for a small business.

If your enquiries arrive mostly by email rather than chat, a helpdesk with AI answers is the closer fit. Help Scout, for example, sells AI Answers as a $0.75-per-resolution add-on on its paid plans, which start at $25 a user a month, and doesn't bill a resolution when the customer escalates, searches the help docs, asks again or says they need more help. That definition is kinder to small businesses than an "assumed resolution" after silence, so it's worth reading every vendor's version before comparing prices.

Checking the replacement works in its first month

The point of moving off the GPT was visibility, so use it. The driving school tracked four numbers from the chat tool's reports and its Friday review of 20 AI-closed chats (illustrative first-month figures):

The remaining 15% were people who left before finishing, a number worth watching but not over-reading. None of these figures existed while the GPT was live, which is the strongest argument for the switch.

It depends on where your customers already message you

The best choice follows your customers' habits, not the most capable AI. Four questions settle it:

1. Where do enquiries arrive? If more than half come through WhatsApp or Instagram, start there; a website widget won't change habits. If most come through your website or online shop, start with web chat.
2. How many a month? Under about 50, a well-written FAQ page and fast human replies often beat any bot. Between 50 and a few hundred, a chat tool with AI on routine questions pays back. Above that, compare per-outcome and per-conversation prices on your real resolution rate.
3. How many of them are routine? If fewer than a third could be answered from a written policy, AI will mostly hand over, and you're paying for a receptionist that forwards messages.
4. What must never be automated? Refunds, complaints, anything involving health or safety, and anything about a named member of staff. List them before choosing, because every tool needs to be set up to pass them on.

Picks for an online bakery and a WhatsApp-first groomer

Two more illustrations show how the answer changes with the business. An online bakery selling cake boxes through Shopify gets most questions about delivery dates, ingredients and gift messages on its product pages. Its first move is Shopify Inbox's free AI agent, tested on its own policies before going live, with one firm rule: allergen questions always go to a person, because an answer pulled from the web about "typical" ingredients could harm someone. Because the agent can use web search as a secondary source, the bakery checked specifically that it never answered an allergen question from anywhere but a person.

A dog groomer, by contrast, gets 80% of messages on WhatsApp from regulars rebooking every six to eight weeks. A website chat would sit unused. The groomer uses the WhatsApp Business app with Meta's Business Agent set to answer opening hours, prices by breed and the booking link, and to hand over anything about a dog's health or behaviour. At roughly 4-5 cents a message, 150 conversations of five AI messages each would cost around $30-$38 a month. The trade-offs of letting AI reply on WhatsApp are covered in WhatsApp customer service with AI.

Retiring a GPT you've already shared with customers

If a custom GPT is already linked from your business, don't wait for 11 December. Links to it hide in more places than the website: the driving school found its GPT link on the FAQ page, in two email signatures, in an Instagram bio and in the automatic reply sent to new enquiries. Search for the link, replace each one with your new chat or messaging option, and put a short note on the old FAQ entry: "Questions? Use the chat button on this page or message us on WhatsApp; a person replies within one working day." Then keep the GPT's instructions and files: they're a ready-made first draft of the knowledge your new tool needs, as long as you check every price and policy in them first. The school's GPT file still had the old $35 lesson price, which is exactly the kind of error that would otherwise have followed it into the new tool.

Rules to write before any bot talks to a customer

Whichever tool you choose, write these down first. The driving school's version, filled in:

OPENING MESSAGE
"Hi, I'm the [school name] automated assistant. I can help with
prices, availability, car types, booking your test and our
cancellation policy. Type 'person' at any time to reach the office."

ANSWER ONLY FROM
Current price list, availability rules, cancellation policy,
test booking steps, car types page. Nothing else.

ALWAYS HAND OVER
Refunds and complaints; changing instructor; anything about a
learner's health or medical conditions; anything about an
instructor's conduct; questions it can't answer from the files.

NEVER
Promise a test date or pass rate; quote a price not on the list;
ask for licence numbers or dates of birth in chat.

REVIEW
Office manager reads 20 AI-closed chats every Friday and updates
the files for anything answered wrongly or not at all.

The opening message does two jobs: it tells people they're talking to an automated assistant, which the EU AI Act's transparency duties require if you sell to customers in the EU, and it gives them the exit before they need it. Wording for that first message is covered in what to tell customers at the start of a chat.

A five-minute way to choose

Run it in order and stop at the first yes. Selling through Shopify? Try Shopify Inbox's AI agent first; it's free. Already on HubSpot Service Hub Professional? Use the Customer Agent you're paying for. Most messages on WhatsApp, Instagram or Messenger? Use Meta's Business Agent, kept to your business tasks. Otherwise, choose a website chat tool with handover, Intercom if one or two people answer chat and Tidio if several do. And keep ChatGPT, through a project rather than a GPT, for the staff side: drafting the replies a person sends.

Questions owners ask about GPTs and customers

Can I embed a custom GPT on my website?

Not as a proper website chat. A GPT lives inside ChatGPT, and customers have to sign in to ChatGPT before they can talk to it, so a link on your site sends them away to someone else's product. With custom GPTs retiring on 11 December 2026, any link to one will stop working. Website chat tools are built to sit on your own pages instead.

Could ChatGPT's newer plugins or projects replace a customer GPT?

No. Projects and the plugins that replace GPTs are for people working inside ChatGPT, such as your staff. They don't give you a customer chat window on your site, a shared inbox for handovers, or a log of customer conversations. For customers, use a chat or messaging tool designed for them and keep ChatGPT for internal drafting.

Do I have to tell customers they're talking to a bot?

It's good practice everywhere, and if you sell to customers in the EU, the EU AI Act's transparency duties, in force since 2 August 2026, require that people are told when they're interacting with an AI system unless it's obvious. Say it in the first message and explain how to reach a person, in plain words.

Further reads

• Customer-Facing or Back-Office: Where Should AI Go First? — Why back-office AI is usually the safer first project.
• How Much Does a Website AI Chatbot Cost, and Will It Pay Off? — Work out whether a website chatbot pays for itself.
• Meta AI in WhatsApp: What It Means for Your Business Chats — How Meta's own AI changes business chats on WhatsApp.
• How to Monitor AI That Talks to Customers: Hand-Offs and Errors — Watch hand-offs and errors once a bot goes live.
• How to Test a Customer Chatbot Before It Goes Live — The tests to run before customers see any chatbot.
• Custom GPT vs Claude Project vs Gemini Gem: Which Should You Use? — Where custom GPTs are going, for your internal uses.
• How to Create a Custom GPT for Your Business Step by Step — Custom GPTs are retiring. Build the same assistant as a ChatGPT Project, skill or plugin instead, and move any GPT you already have before 11 December 2026.
• 12 Custom GPT Examples for Small Businesses, With Setup Notes — Twelve assistants small businesses built as custom GPTs, rebuilt as Projects and Gems after the GPT retirement, each with instructions, files and sharing.
• Are Custom GPTs Private? What Happens to the Files You Upload — Who can see a custom GPT's chats and knowledge files, how consumer and business plans differ, and what the December 2026 retirement means for your uploads.
• What If Your AI Vendor Shuts Down? Checks Before You Commit — Real AI shutdowns and retirements, seven checks before you sign, a vendor-independent copy routine, and a ten-day runbook for exit day.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI's custom GPT retirement notice and help pages on using and building GPTs (via search results); Intercom and Tidio pricing pages; HubSpot, Meta Business Agent, Shopify Inbox and EU AI Act details from the verified fact sheet. All checked September 2026.