What this page is for
AI features (the concierge on your booking website, AI agents, reply suggestions) need an AI service to answer. Here you connect one or more: your own subscription with Claude, OpenAI, Gemini, Azure OpenAI, Mistral, Groq, OpenRouter or DeepSeek, any OpenAI-compatible service, or a Local AI server (Ollama) that you run yourself. Without a connection, every feature still works with its simple fallback.
Who can use it
Set up AI connections and features (by default Admin). Usage and costs need See AI usage and costs.
How to add a connection
- Under Add a connection, give it a name you will recognise, for example "Claude for replies".
- Choose the Service. Open Where do I get these details? for where to create a key.
- Paste the API key. Local AI does not need one.
- Leave Model and Address blank to use the suggestions, or type your own. For Azure, enter your endpoint as the address, the deployment name as the model, and the API version.
- Optionally enter the prices per million tokens from the service price list, and a Monthly budget.
- Select Save connection, then Test. A working connection answers in a second or two.
How to change or delete a connection
Select Edit. Leave the API key blank to keep the saved one, or paste a new key to replace it. To delete, tick the confirmation and select Delete. Features that used it move to their next connection; the usage history stays.
Fields and columns
| Field | Meaning |
|---|---|
| Name | For you only. |
| Service | Which AI service this is. "Data stays with you" marks Local AI. |
| API key | Your secret key with the service. Stored encrypted and never shown again. |
| Model | Which model to use. Blank uses the suggestion for the service. |
| Address | The service address. Needed for Azure, compatible services and a Local AI server on another machine. |
| Organisation or project ID | Only for OpenAI accounts that use one. Stored encrypted. |
| API version | Azure OpenAI only. |
| Wait for an answer | Seconds before giving up and trying the next connection. |
| Prices per million tokens | Used for the cost estimate, in your base currency. Leave 0 for Local AI. |
| Monthly budget | Optional limit per calendar month in your timezone. |
| Order | Lower numbers are tried first by features that use every connection. |
| Active | Switch a connection off without deleting it. |
| This month | Estimated spend so far this month, and the budget. |
| Last test | Result of the last test and the last error, if any. |
Statuses and automation
- Every call is logged with tokens, estimated cost, time and any error (see AI usage).
- When a connection fails, the feature tries its next connection at once.
- At 80% of the monthly budget, people who manage AI get a notification (once a month). At 100% the connection is skipped until the next month.
- Before text goes to an outside service, email addresses and phone, card or account numbers are replaced (setting Remove contact details before sending text to outside AI, on by default). Local AI gets the full text.
Settings that change this page
Settings, AI connections: removing contact details before sending text to outside AI.
FAQs
Which service should I choose? For the best answers, Claude Sonnet or a similar model. For low cost, a small model such as Claude Haiku. For privacy and no bill per message, Local AI.
Can I use more than one? Yes. Add several and choose the order per feature under AI features, for example Local AI first and your subscription as backup.
Is my key safe? It is encrypted in the database, never shown again and never written to logs.
What is "Demo AI"? On demonstration sites only: short sample answers so you can try the features without a key.
Troubleshooting
| Problem | What to do |
|---|---|
| Test says "HTTP 401" | The key is wrong or was revoked. Paste a new key. |
| Test says "HTTP 404" | The model or address is wrong. For Azure, check the deployment name and API version. |
| Test says "HTTP 429" | The service is busy or your account is out of credit. Check your account with the service. |
| Local AI does not answer | Check the Ollama server is running and the address is reachable from this server, and that the model was downloaded. |
| A feature stopped using AI | The budget may be used up. See AI usage, raise the budget or wait for next month. |