Which AI Model Should You Connect to Your Website Chatbot? A BYOK Guide

With a bring-your-own-key chatbot, you connect your own account at an AI provider and pay them directly for what the assistant uses. That gives you control over cost and quality, and it also hands you a decision the platform would otherwise make for you: which provider and which model? The choices look intimidating, because there are many providers, dozens of models and prices that change every few months. The good news is that a customer-facing assistant that answers from your own documents needs less from the model than most people expect. This guide explains what actually matters, compares the main provider families, shows a simple way to test models on your own questions and gives sensible starting points for a small business.

Providers change their models, prices and policies often. This guide describes categories and trade-offs and avoids quoting prices or model names as facts. Check each provider's current documentation before you decide.

What the model does in a grounded assistant

In an assistant that answers from your documents, the search step finds the right passages, and the model's job is to read them and write a clear, faithful reply in the right tone. That is a reading and writing task, not a knowledge task. The model should follow instructions, stay within the passages, admit when they do not contain the answer and write naturally. Very large, expensive models are not required for that, and a smaller model that follows instructions well is often the best value. See how an AI website assistant works for the whole pipeline.

Five criteria for choosing

Multilingual ability is a sixth criterion if you serve several languages; see multilingual chatbots.

The main provider families

Crescelo supports several providers, and you paste a key for the one you choose. The list below describes what each family is generally known for, without prices.

A free engine is also available in Crescelo to start and test before you connect your own key. See the provider guides for OpenAI, Gemini and the others.

Small, medium or large model?

Start small, test on your real questions and step up only if needed. Reasoning-style models that think before answering can be slower and more expensive per reply, and are rarely needed for customer questions from a knowledge base.

A one-hour test on your own questions

1. Write 25 questions: fifteen your documents answer, five they do not, three that tempt the assistant to promise something and two in another language if you serve one.
2. Connect the first candidate model and run all questions in the Testing Studio.
3. Score each answer: correct, partly correct, wrong or honestly unknown.
4. Note speed and how natural the tone feels.
5. Switch to the next candidate and repeat.
6. Estimate the cost per conversation for each from the provider's pricing.
7. Choose the cheapest model that scores well on the honest-unknown and no-promise questions.

The honest-unknown score matters more than the correct-answer score, because a model that guesses is dangerous however fluent it sounds. See how to stop a chatbot from making things up.

Privacy and data handling

• Read each provider's terms for API usage: whether prompts and outputs are stored, for how long and whether they are used to train models.
• Check where data is processed if your customers are in regulated regions.
• Prefer providers that offer clear business terms and a data processing agreement.
• For the strictest cases, a self-hosted model keeps data on your own systems, at the cost of running it.
• Tell visitors they are talking to an AI and link to your privacy policy.

The GDPR checklist for website chatbots explains what to document when a provider processes personal data for you. This is general information, not legal advice.

Keys, limits and safety

1. Create a dedicated API key for this assistant, so you can revoke it without affecting anything else.
2. Set a monthly spending limit or alert at the provider.
3. Paste the key only into the assistant's AI settings, where it is stored encrypted. Crescelo stores provider keys with AES-256-GCM encryption.
4. Never put a key in your website code or share it in a chat or email.
5. Rotate the key if a team member with access leaves.

See how a bring-your-own-key vault works for what good key storage looks like.

When to switch models

• Answers are wrong or ignore your documents even after content fixes.
• Replies are too slow for visitors.
• Costs are higher than the value warrants.
• A provider changes prices, terms or retires a model.
• A new model in your test set clearly outperforms the current one.

Because you own the key, switching is a settings change. Re-run your 25-question test after any switch, and keep the results so you can compare.

Fallbacks and reliability

Providers have outages and rate limits. A robust assistant handles them gracefully: it retries, falls back to another provider when configured, and never shows a raw error to a visitor. Crescelo's free engine cascades across free providers for testing, and with your own key you should watch your provider's status and limits. For production, consider keeping a second provider account ready. See fail-closed pipelines and chat reliability.

Sensible starting points

Mistakes to avoid

• Choosing the biggest model by default.
• Choosing on price alone without testing honesty.
• Sharing one key across many projects.
• Forgetting spending limits.
• Never retesting after a model or price change.
• Ignoring the provider's data terms.

Frequently asked questions

Try the free engine on your own documents in the playground, read about the AI assistant and start a free trial to connect your own key.

Further reading

• OpenAI API documentation: Models, pricing and limits for the OpenAI API.
• Anthropic Claude API documentation: Models and usage of the Claude API.