In short: Custom GPT can mean a configured assistant, retrieval over your documents, or a model called from your own software. Say which, which files are in scope, where a person reviews the answer, and whether your data may be used for training. Test on your examples. Checked 12 Nov 2026.
Chatbot package figures, where this site publishes them, stay on the AI chatbot cost guide. The chatbot calculator is a separate tool. This page is how to brief and accept a custom integration. It has no price table and no model ranking.
Ten points
1. Say which product you mean
A staff assistant over internal files. A customer-facing bot. Or a step inside a workflow that classifies, extracts or drafts. If someone is selling an “agent”, read AI agent versus workflow automation before you sign that word into a contract.
2. Do not start with a model brand
Accuracy on your examples, cost as volume grows, speed, and where data is processed matter more than a product name. Name a model only to describe what the vendor proposes. The contract questions are questions to ask an AI agency.
3. Retrieval is not “the model remembers us”
If answers must come from approved documents, say so, and ask how a changed document gets into the index. If the vendor wants to fine-tune, ask why retrieval is not enough and what data that needs. The data checks are data readiness.
4. Name the sources of truth
Which sites, folders and systems are in. Which are out. Whether the answer links back to a source document. A bot that cannot point at a file should not be the only source of a customer commitment.
5. Each channel is extra scope
A website widget, email, a CRM field and WhatsApp are different integrations. Do not let one quote assume all of them. Channel rules, including WhatsApp, belong with the vendor and the platform. This page does not invent them.
6. A person still handles the hard cases
Prices, legal advice, medical advice and promises to a customer. The pattern is human approval. The system should be able to refuse or escalate. That test is on the pilot checklist.
7. Evaluate on questions you held back
A demo on the vendor’s examples is not acceptance. Agree a set of real questions, including ones the system should refuse. Do not invent an accuracy target here. You and the vendor set it for that set.
8. Watch usage during the pilot
Questions and retrieved context can grow the bill. Record calls or tokens so a later quote is not a surprise. No unit price is published here. Cost drivers are the cost factors page.
9. Data use, logs and the PDPA
No training on your data unless that is written. No production personal data if a redacted sample will do. The PDPC guidelines of 1 March 2024 are advisory. The organisation list is SME AI governance. Checked 12 November 2026.
10. An internal policy assistant versus a public negotiator
Illustrative only. An assistant over a tidy internal knowledge base, with citations and a person sending anything external, is a bounded brief. A public bot expected to negotiate prices is a different product and a different risk. This page claims neither will be accurate.
Industry versions of the matching page include professional services, e-commerce and education.
WhatsApp +65 9071 3976 (chat).
Frequently asked questions
What is the difference between a custom GPT and an AI chatbot?
The labels overlap. Ask whether you are buying a configured assistant, a retrieval system over your files, or a customer-facing integration. The chatbot cost guide covers chatbot-shaped pricing. This page does not add a price range.
How much does LLM integration cost in Singapore?
There is no official fee in the sources for this page. Compare scope, usage and support. Chatbot-shaped work is priced on the chatbot cost guide. Consulting lines are on the cost-factors page.
Will our documents be used to train the model?
Only if the contract says so, and only with a lawful basis for any personal data. Do not assume a default of no. Ask, and get the answer in writing.
