Custom GPT vs RAG Chatbot vs AI Agent: A Buyer's Guide to Cost and Ownership
Usama · November 6, 2026
Pick a ChatGPT-hosted assistant when a handful of your own staff will use it, a RAG chatbot when customers or many employees need answers from a large, changing document set, and an AI agent only when the assistant must take actions in your other systems. What decides it is not which technology is cleverest. It is who will use the thing, what data it touches, and who will still be looking after it a year from now.
This article is about the purchase. If you need the definitions first, read our guides on training ChatGPT on your own data, RAG versus fine-tuning and AI agents versus chatbots versus automation.
Which one fits your situation?
| If your situation is... | Lean towards | Why |
|---|---|---|
| Three to ten staff need quick answers from a few dozen documents | ChatGPT-hosted assistant (a Project today) | Cheapest test, no build, easy to abandon |
| Customers will ask questions on your website or WhatsApp | RAG chatbot | You control the data, the wording, the channels and the escalation path |
| Hundreds of documents, edited weekly, different access for different teams | RAG chatbot with permission filtering | File caps and manual uploads stop scaling |
| You need answers that link back to a page or clause | RAG chatbot | Source display is part of the design rather than a request in a prompt |
| The assistant must create tickets, update a CRM or book appointments | AI agent, or a workflow with an AI step | Actions need tools, permissions and approval rules |
| The task is the same every time and you can write the steps | None of the three: use a workflow | Cheaper, more predictable, easier to audit |
The three options compared on ownership
| ChatGPT-hosted assistant | RAG chatbot | AI agent | |
|---|---|---|---|
| Build effort | Hours to days; mostly gathering and cleaning documents | Weeks for a first useful version; documents, search, prompt, interface, tests | Weeks to months; tools, permissions, limits, approval steps, test scenarios |
| Who keeps it running | Whoever owns the account; files replaced by hand | You or a supplier: re-indexing, monitoring, failed-question review | You or a supplier: all of the above plus tool changes, cost caps, incident response |
| Where your data sits | Inside the vendor's product, under that plan's data terms | In storage you choose; passes to a model provider only for each question | Same as RAG, plus whatever the connected systems expose to it |
| Biggest risk | Product changes or retirement; data-setting mistakes | Stale or poorly organised documents give bad answers | Wrong actions; permissions that are too wide; runaway cost |
| Ongoing cost shape | Seat subscriptions | Model usage per question, hosting, upkeep time | Model usage per run (usually several calls), hosting, monitoring, review time |
| Example uses | Internal policy helper; sales team FAQ | Support assistant; product advisor; internal knowledge search with citations | Intake handler that reads an email, checks the CRM and drafts or books |
Last verified November 2026. Cost shapes are described, not priced, because vendor pricing changes; we do not quote figures for any provider here.
Why a Custom GPT is a risky foundation in late 2026
Custom GPTs were the default answer for a small business wanting its own ChatGPT. OpenAI's help centre now says the company is planning to retire them in favour of plugins, with standard retirement scheduled for 11 December 2026, and personal accounts can no longer create or publish new ones. Knowledge files are copied across when a GPT is migrated, but connections built with custom actions do not carry over. The details are in OpenAI's retirement and migration FAQ.
The lesson is bigger than one product. Anything you build inside someone else's interface is on their roadmap, not yours. Keep your documents, your prompts and your test questions in files you own, so that moving from a Project to a plugin to your own system takes days rather than a rewrite. For the current set-up options, our walk-through of training ChatGPT on your own data covers Projects and file limits.
Last verified November 2026. OpenAI describes the dates as planned and says some workspaces differ, so read the notice for your own account.
What does ownership cost beyond the build?
Buyers compare build quotes and forget the lines that appear afterwards. Ask for each of these, in writing, from any supplier including us.
- Content upkeep. Someone has to retire old documents and add new ones. This is the most underestimated cost, and it never goes away. An assistant is only as current as its sources.
- Usage fees. Billed by the model provider, and rising with traffic. Ask for a usage estimate at your expected volume, and what a spike would cost.
- Hosting and search. A small fixed monthly cost for most RAG set-ups, but it exists.
- Testing time. A question set re-run after every change, and a weekly look at real conversations.
- Model changes. Providers retire models and change behaviour. Budget for re-testing when that happens.
- Security review. Access rules, logs, retention, and the right to delete.
- Exit costs. What you would need to move to another supplier or bring it in-house.
Who sees your data under each option?
Put these questions to every option. The answers are different, and so is the paperwork. This is general information, not legal advice, so have a lawyer or your data protection lead confirm the position for your sector and country.
- Where are the documents stored, and in which country?
- Does the vendor or model provider use inputs to improve its models? OpenAI says it does not by default for its API and for Business, Enterprise and Edu workspaces, while individual plans have a setting you must check. See the data-use page.
- Who at your supplier can read conversations, and for how long are logs kept?
- Can different users be limited to different documents?
- Can you delete everything, including copies in indexes and logs, on request?
- Which sub-processors are involved, and is there a data processing agreement?
Our article on AI privacy and compliance goes further on the regulatory side.
Who should own the finished assistant?
Whatever you buy, insist on a short list of things being yours from day one:
- The accounts with the model provider and the hosting company, opened in your company name.
- The source documents and the cleaned versions, in your storage.
- The prompts, instructions and settings, delivered as readable text.
- The test questions and the latest score sheet.
- A written handover note explaining how to update content, rotate keys and switch off the assistant.
A supplier that keeps these on their own accounts is renting you something you are told you own.
A staged route that limits regret
Stage one: prove the questions
Use a Project and write the question set. Run it for two or three weeks with real users. You spend almost nothing, and you learn which documents are missing and what people really ask.
Stage two: build retrieval where it pays
Move to a RAG assistant when you hit a concrete limit: file caps, public-facing use, access rules, or the need for citations and logs. Our custom AI assistant service works this way: scoped around your own questions, with sources shown under each answer. When the assistant lives on your website or in a messaging app, the chat interface is connected through chatbot integration.
Stage three: add one action with a gate
Once answers are reliable, add a single action such as creating a support ticket, with a person approving it. If that works, add the next. This is how agentic builds stay safe: autonomy is earned in small steps.
Red flags in a proposal
- "Trained on your data" with no explanation of whether that means retrieval or fine-tuning.
- A promise of 100 percent accurate answers. No language-model system can promise that.
- No test plan, or a test plan the supplier will write after launch.
- An agent proposed for a job you could describe as five fixed steps.
- Vague answers about where data is stored and who can see it.
- Pricing that covers the build but says nothing about content upkeep and usage fees.
- Source code, prompts or accounts that stay with the supplier.
Comparing quotes for an AI assistant?
Send us the proposals you have and we will point out what each one leaves out.
Review my options →Frequently asked questions
Is a custom GPT or a RAG chatbot better for a small business?
For internal use by a few people, a ChatGPT Project is quicker and cheaper to try. For customers, larger document sets, permissions or citations, a RAG chatbot is the better fit. Custom GPTs themselves are scheduled for retirement, per OpenAI.
Do I need an AI agent for customer support?
Usually not at first. A retrieval-based assistant answers questions; an agent is only needed when it must act, for example issuing a refund or editing an order. Start with answers and add actions one at a time, with approval.
Who maintains a RAG chatbot after launch?
Someone must keep documents current, watch real conversations and re-test after changes. That can be your staff, the supplier on a support plan, or both. Decide before signing, not after.
How much of my data does the AI provider see?
In a RAG set-up, only the passages sent with each question, plus the question itself. Your full document store stays in storage you control. Check the provider's data terms for how those requests are handled.
Can I move from one option to another later?
Yes, if you kept documents, prompts and test questions in your own files. That portability is the main reason to keep them outside any single vendor's tool.