ChatGPT for Customer Service (2026): What Works, What Breaks, and What It Costs a Small Business

ChatGPT for customer service: the short answer

As of September 2026, ChatGPT is a good customer service writer and a bad customer service system. It will draft replies in your voice, answer policy and pricing questions from documents you give it, triage an inbox, and translate, and on ChatGPT Business that costs $20 a seat a month billed annually or $25 monthly with a two-seat minimum. It will not look up an order, check a booking, issue a refund, or remember a customer from last month unless the system that holds that data is connected, and even then a human should click send. Every number here comes from OpenAI's pricing and help pages, read on September 15, 2026.

I install this stuff for owner-led businesses. The pattern that works is boring: ChatGPT drafts, a person approves, and nothing reaches a customer without a human deciding it should. The pattern that fails is "let it answer." Here's where the line sits and what to do on each side of it.

What ChatGPT can actually do for customer service

The capability line, honestly drawn:

Two things on that table deserve a footnote. OpenAI's August 25, 2026 release notes say ChatGPT Work's browser can now complete tasks on websites that require you to sign in, and it "will always ask for confirmation before consequential actions." That's available on Plus and Pro. It's a way to check an order portal, not a way to run one. And OpenAI's own guidance for its Company Knowledge plugin says to "review the answer and any source links provided before relying on the information." If OpenAI says review it, review it.

What it costs: the plan you need

The column that decides it is the training one. OpenAI's Plus help page says "OpenAI may use conversations to improve model performance and safety" with an opt-out; the Business overview says "OpenAI won't train on your workspace's data." Support email is customer data. Two Standard seats billed annually is $40 a month, which is the real entry price for doing this properly. Premium buys 5x the ChatGPT Work and Codex usage with no 5-hour limit; you don't need it for a support inbox unless one person is running heavy automations all day. The seat math in detail is on ChatGPT Business pricing.

Two things Business doesn't include, per OpenAI: a BAA (that's a contracted offering, so healthcare support is out of scope for the self-serve plan), and data export from the workspace.

A realistic setup, step by step

1. Open a Business workspace with two seats. Owner plus whoever answers email. Domain and SSO setup is covered in the ChatGPT Business setup guide; you don't need it on day one.
2. Create a shared project called Support. Upload the files that answer 80% of questions: return and refund policy, shipping or scheduling rules, price sheet, warranty terms, and five of your best real replies. OpenAI allows 40 files on Business and 10 uploads at a time. Shared projects switch to project-only memory automatically, which is what you want: the project can't see anyone's personal chats, and personal chats can't see it.
3. Write the instructions like a job description. Voice, length, sign-off, what to never promise. One line I put in every support project: "If the answer is not in the files, say you need to check and draft a question for the owner instead of guessing."
4. Give the team chat access, not edit. Business lets you invite by individual, group, or workspace link and set chat or edit at invite time. The owner edits the files; everyone else chats.
5. Add the inbox trigger. Connect Gmail under Settings, then Apps, and create an event-triggered task: new message filtered by subject or sender, classify it, draft a reply from the Support project's rules, put the draft in a place a person will see. OpenAI says actions that send a message may require approval and the task pauses until you review it. Leave it that way.
6. Review weekly. Read the drafts that got edited before sending. Every edit is a missing file or a missing instruction. Fix the project, not the reply.

One gotcha from OpenAI's help center that bit a client: a scheduled task created inside a project "cannot access uploaded files or files stored in that project." Create the task outside the project and paste the rules you need into its instructions, or point it at the connected source app.

The failure modes

• Hallucinated policy. The big one. Ask about a scenario your files don't cover and you get a fluent answer that reads like your policy and isn't. The instruction in step 3 plus a human on send is the fix. There's no setting that makes this go away.
• No order lookup. "Where's my order" is the most common support question and ChatGPT can't answer it without a connected system. If your order data lives in a tool with an app in ChatGPT's directory, connect it. If not, the draft says "let me check" and a person checks.
• Stale files. You changed the shipping cutoff in March and the project still has the January PDF. Put a date in every file name and a monthly reminder on the calendar.
• Tone drift on Instant. Paid plans run Instant on GPT-5.6 Sol, and Business can pick Medium, High, Extra High, or Pro thinking. Replies to complaints are worth a higher setting; sorting the inbox isn't. Business Standard includes 15 Pro messages a month, so spend them on the angry ones.
• Building it as a custom GPT. OpenAI's September 11, 2026 release note announced a planned retirement of custom GPTs across plans with a migration to plugins. Anything you build as a GPT today is a migration project later. Use a project.
• Usage walls. Chat is unlimited on paid plans, but a Gmail-triggered task runs in ChatGPT Work, which is metered. Standard seats get roughly 10 to 100 GPT-5.6 Sol messages per five-hour window there. A busy inbox will find that limit; route the sorting to Luna, which OpenAI describes as built for "routing, classification, extraction, support," and save Sol for the drafts.

When to move to a proper agent

ChatGPT with a human on send is the right tool until the human becomes the bottleneck. The signals I watch for:

• Support volume where reviewing drafts takes longer than writing them used to.
• Questions that need a system answer (order status, booking changes, account balance) more than half the time.
• After-hours or phone coverage, which ChatGPT doesn't do at all.
• A need to log every conversation into a ticketing tool or CRM automatically.

There's a middle step before you buy a dedicated product. Business and Enterprise workspaces include workspace agents: OpenAI describes them as agents for repeatable workflows that you can connect to apps and tools, test before publishing, share with the team, run on a schedule, use in Slack, or trigger through an API. A "triage and draft" agent that the whole team shares is a real upgrade over each person prompting on their own, and it's included in the seat price.

Past that, you're shopping for a system with integrations, an escalation path, and transcripts you can read, and the feature checklist for that decision is in AI agents for customer service. The phone side in particular is a different product category. Don't try to make ChatGPT do it.

What I'd do this week

Pull your last 50 support emails. Sort them by hand into "answerable from a document" and "needs a system." If the first pile is bigger, two Business seats and a Support project will pay for themselves in the first month. If the second pile is bigger, ChatGPT is still worth it for drafting, but the real win is connecting the system, and that's a different project.

Either way, the rule doesn't change: ChatGPT drafts, a person sends. The broader picture of what's worth installing first is in ChatGPT for small business. Customer service is usually second on that list, right after the owner's own inbox.