AI Engine uses two model settings in the Org Profile: a main model for most of the work, and an optional fast model for web searches and the in-app help assistant. This article explains how each one is used and how to choose them.
How the two models are used#
| Setting | Used for |
|---|---|
| AI Model (main) | Chat answers and the Salesforce queries behind them, record actions such as Next Steps and Draft Email, Meet Prep briefs, scoring, custom prompts and PDF text |
| Fast model (internal steps) | Web searches: Web Check, Company News, chat questions that need a web lookup, Magic Snap research, and the optional web steps of Meet Prep and AI scoring. The in-app help assistant also uses it when you pick a fast model; with Provider default the help assistant uses the main model. |
Most of the quality your users notice comes from the main model. The fast model mainly affects the cost and speed of web searches.
Choose the main model#
- Open Ai Agent Admin Setup → Org Profile and edit the profile.
- For OpenAI or Anthropic, pick a model from the AI Model list. Each entry has a short tag. One model per provider is tagged recommended, and that is a good starting point. Entries tagged experimental are newer models that have not yet been confirmed with every AI Engine feature. Avoid them for production users.
- For Azure OpenAI, type the deployment name from your Azure resource, not the model family name.
- Check that Named Credential still matches the Named Credential for that provider.
- Click ✓ Save.
Choose the fast model#
- Leave Fast model (internal steps) on Provider default unless you have a reason to change it. For OpenAI and Anthropic, the provider default is a smaller, cheaper model from the same provider. The list shows which model that is.
- The fast model must come from the same provider as the main model, because both go through the same Named Credential. The list only offers models from the selected provider. If you change the provider after picking a fast model, the old choice is not saved, so pick the fast model again.
- For Azure OpenAI, the field is a text box. You can enter the name of a second, cheaper deployment for the help assistant. If you leave it blank, the main deployment is used.
Web searches with Azure OpenAI#
Azure OpenAI deployments are not used for web searches. With Azure OpenAI selected, Web Check, Company News, web lookups from the chat, Magic Snap research and the web steps of Meet Prep and AI scoring send their search request to OpenAI instead, through a Named Credential named exactly OpenAI, using the gpt-4.1-mini model. Those requests are processed by OpenAI, not in your Azure resource.
- If you use these features, also create an OpenAI External Credential and a Named Credential named exactly
OpenAI, with the OpenAI URL and header, and make sure your OpenAI key can usegpt-4.1-mini. See Connect your AI provider. - If your data must stay in Azure, do not use these features. See Choosing a provider for compliance.
What to consider#
- Answer quality versus cost. Larger models write better analysis and follow long instructions more reliably. They also cost more per call. The Activity Monitor shows the estimated cost per feature, and the model and estimated cost of each call in its recent log. See Activity Monitor.
- Structured output. Several features ask the model for a structured answer that AI Engine turns into cards: Meet Prep, AI scoring, Next Steps, Analyze Email and Company News. Models differ in how reliably they return that structure. When a model does not return it, the result appears as plain text or as an empty or incomplete card rather than as an error.
- Answer length. Some models use part of the output budget for reasoning before they answer. If record actions, Meet Prep briefs or PDF text become shorter or stop mid-sentence after a model change, set Max Output Tokens above 3,500. Chat answers have a fixed limit that this setting does not change. See Org Profile settings.
- Compliance. Where your data may be processed often decides the provider before the model. See Choosing a provider for compliance.
After changing a model#
Test on a record with plenty of data and on a record with very little:
- Ask a list question and a chart question in the chat.
- Run Next Steps and Meet Prep.
- Score one Lead or Opportunity from the AI Score card.
- Generate one PDF from a template, if you use them.
If one of these returns an empty card, switch back to the previous model and contact Cloud Connected with the feature name and the model.
Good to know#
- The model list in the admin panel is a curated catalogue. The full list of supported providers is in Supported providers and models.
- If you keep a tested Named Credential for a second provider, you can switch during a provider outage by changing the provider, model and Named Credential in Org Profile.