Skip to content

Black Friday:

60% off

Cliedesk Pro

Every plan, every site count. Ends Cyber Monday.

00 days 00 hrs 00 min 00 sec
Claim 60% off

AI

AI settings

This guide covers a feature of Cliedesk Pro. The free plugin is a complete help desk; see what Pro adds.

AI is part of Cliedesk Pro. You bring your own key: you pay your AI provider for what you use. Cliedesk charges no AI fee per resolution, per outcome or per seat. People with the Manage AI permission (Admins) set up AI in Settings → AI. Supervisors can see costs and traces.

Connect a provider

  1. Open Settings → AI → Providers.
  2. Under Add a provider, choose the provider: OpenAI, Anthropic Claude, Google Gemini, Mistral AI, OpenRouter, Revoye, Voyage AI (embeddings) or OpenAI-compatible (presets for Ollama, LM Studio, vLLM, xAI and DeepSeek, or your own endpoint address).
  3. Paste the API key. Cliedesk stores it encrypted and never shows it again (only •••• and the last 4 characters).
  4. For OpenAI-compatible, you can type the Model id (for example llama3.1 or grok-4). Cliedesk checks the key with it, adds it to the model list and uses it for the AI tasks that have no model yet. Leave it empty to use the models the endpoint lists.
  5. Select Check key to test it before saving, then Connect. Accept the AI processing notice for each new provider.

Revoye answers in 3 to 90 seconds and has no tools, images or embeddings. Use it for drafts, summaries, rewrites and translations. The AI agent and live chat need another provider. Cliedesk sets Revoye’s model revoye/auto (Revoye chooses the model) for the tasks it can do. Check key reads Revoye’s model list instead of waiting for an answer.

Test connection shows whether chat, structured output, tool calling and embeddings work, and how long the answer took. To change a key, paste the new key and select Save key; an empty field keeps the current key.

DeepSeek stores data in China: Cliedesk shows a warning. Free Gemini keys allow Google to train on your data; use a paid-tier key.

Enterprise providers

You can also connect Azure OpenAI / Microsoft Foundry, Amazon Bedrock and Google Vertex AI:

  • Azure OpenAI / Microsoft Foundry: enter the resource address (for example https://my-resource.openai.azure.com) and paste the resource key. To sign in with Microsoft Entra ID instead, paste the app as JSON: {"tenantId": "…", "clientId": "…", "clientSecret": "…"}. As Region, choose the deployment type (for example Data Zone EU or Sweden Central (regional)). Azure does not list deployments: after connecting, type each deployment name under Add a deployment name.
  • Amazon Bedrock: choose the region, then paste a Bedrock API key, or AWS access keys as JSON: {"accessKeyId": "…", "secretAccessKey": "…"} (add "sessionToken" for temporary keys). Use a model id or an inference profile; eu. profiles stay in the EU, global. and us. profiles do not. Bedrock answers appear when the AI has finished (no word-by-word streaming yet).
  • Google Vertex AI: choose the location (for example Netherlands (europe-west4) or EU multi-region) and paste the service account key JSON. The service account needs the role Vertex AI User.

Keys, access keys and service account keys are stored encrypted, like every other key.

EU compliance mode

Settings → AI → EU compliance has a switch: EU compliance mode: only use AI that processes data in the EU. When it is on:

  • only providers and models that process data in the EU can answer: OpenAI with the EU region, Mistral, OpenRouter EU, Azure Data Zone EU or EU regional deployments, Bedrock EU regions (Frankfurt, Ireland, Paris, Stockholm, Milan, Spain) with in-region models or eu. profiles, Vertex AI EU locations, and models on your own server;
  • every other provider or model is skipped, for every AI task, embeddings and speech-to-text. When no EU model is left, the feature shows “Blocked by EU compliance mode”;
  • you cannot choose a non-EU model in Tasks or as the embedding model;
  • masking of customer details stays on, and traces keep only hashes.

The page lists each provider as EU or Not EU, and how many task models are blocked. For Azure, the address must be an Azure address; when the address does not show the region, the page says the EU status is “as you declared” (the deployment type you chose). Test connection does not test a model outside the EU. The Evidence report lists, for a period, every AI call by provider, region, model and task, what kind of data the task sends and which details were masked. Select Download CSV to keep it for your records. Turning the mode on or off is recorded in the audit log.

Keep AI on your own server

To use a model on your server (Ollama, LM Studio or vLLM), turn on Allow local AI endpoints under Privacy, then add a Custom OpenAI-compatible provider with its preset. Turn on No cloud AI to call only these local models. Features that the local model cannot do show “Unavailable in local mode”.

Before AI calls, Cliedesk replaces email addresses, phone numbers, IBANs and street addresses with placeholders, and removes card numbers. The answer gets the real values back.

Choose models per task

In Settings → AI → Tasks, choose the model for each task (classify, draft, summarise, AI agent and more) and up to 2 fallbacks. If a model fails or is too slow (2 retries by default), the next one answers. The trace shows the switch. Choose a product in Settings for to use other models for that product. Pick the Embedding model for knowledge search.

Switchboard and “AI off”

Settings → AI → Switchboard lists every AI feature with its model, this month’s cost and its last error. Switch a feature off, or set its autonomy. Background drafting for assigned conversations is off until you turn it on.

To stop all AI at once, turn on Turn every AI feature off at once. Everyone sees an “AI is off” banner. The inbox keeps working with keyword rules. A developer can also run wp cliedesk ai off or add define( 'CLIEDESK_AI_OFF', true ); to wp-config.php (then the switch cannot turn AI on).

Budgets

In Settings → AI → Budgets, set the Monthly site budget (USD) (suggested: 50) or choose No monthly budget. Admins get a notice and an email at 50 %, 80 % and 100 %. At 100 %:

  • background AI (triage, summaries, quality checks) pauses, and triage uses your keyword rules;
  • copilot works on request only, up to the Allowed overage after the cap;
  • no conversation is dropped.

Test runs (the AI test lab and previews) count toward the site budget too, and they stop at $5 a month.

Per conversation limits the AI cost of one conversation (0 = no cap). Under When the budget runs out, choose a rule with the trigger The monthly AI budget runs out. It runs once. Rules can also use the condition AI budget reached.

Usage and traces

Settings → AI → Usage and cost shows tokens and cost by feature, model, brand, channel or day, the cost per conversation and a forecast. Calls to models without a price show as “unpriced”, never $0. Select Download CSV.

Settings → AI → Traces shows why the AI answered as it did: model, sources, retries and fallbacks, tokens, cost and time. Customer details stay masked. Select Share link for a link that expires (after 72 hours by default), or Export JSON.