AI models and chat
Open AI Models & Chat to configure the AI service the portal calls. A provider is the service receiving the request; a model is the AI selected within that service.
This is different from MCP, which lets an external assistant call portal tools. A provider key does not grant MCP access, and an MCP key does not pay for AI responses.
Choose the right settings tab
| Tab | What it controls |
|---|---|
| Provider & customer model | Cloud provider, its key and customer model, customer-chat switch, greeting, and information-sharing level |
| Admin model | A separate model for the admin assistant and supported admin AI features |
| Quote drafting model | The model that turns typed or dictated notes into quote drafts |
| Company context | Business information and guidance used in AI requests |
| Usage limits | Request limits for customer support chat |
| Test customer chat | Try the saved customer-support setup against a selected customer |
| Quality & usage | Your recent usage, feedback, and sandbox protocol evaluations |
The provider/model configuration, model discovery, and model-response test actions require an account owner. Customer tests, evaluations, and other AI actions also enforce their current write/AI permissions. A read-only Viewer does not gain those permissions by opening this page.
Connect and test a model
- Select Anthropic, OpenAI, Google Gemini, or xAI. Each cloud provider keeps its own saved key and model.
- Enter that provider's API key. A blank key field keeps the existing key for that provider; it does not delete it.
- Choose a suggested model, use Refresh models, or expand Enter a custom model ID.
- Use Test selected model for a short response test, then Save chat settings.
- Open Test customer chat to check the saved setup with customer context.
Refresh models checks the list available through the chosen connection. It does not prove that a model can answer, use tools, read images, or draft a quote. Test selected model sends a short test without customer data, can use unsaved connection details, and does not save changes. Provider usage charges may apply.
Suggested model lists are starting points, not a promise of availability, accuracy, speed, or cost. Saved IDs are preserved even when absent from the list. Check the provider links in settings for current availability and pricing.
Which model wins?
The customer model belongs to the selected provider. A blank Admin model override uses the customer model. A blank Quote drafting model override uses the effective admin model.
Changing the quote model does not switch the provider or key. Save each tab's changes before leaving it. Test the admin assistant and quote drafting separately: the customer test does not exercise those workflows.
Customer support versus the admin assistant
Enable customer chat controls the customer-portal widget, not the admin assistant. It is off by default.
Customer support answers from that customer's permitted portal context and business guidance; it is not the admin tool interface. The information-sharing choices are:
- Summary: project names, statuses, and counts, with structured financial amounts removed where possible.
- Full detail: additional context such as hours, invoices, and budgets.
Summary is not a guarantee that free-text notes or business guidance contain no financial information. Review the information you supply to your AI provider.
The admin assistant can look up records and use supported tools within your permissions. It can return tables, record links, clarification questions, action previews, and follow-up suggestions. Suggestions are not completed actions: check the approval card and final receipt.
Approvals and interrupted actions
For actions that require approval, review the exact recipients, records, amounts, and message before confirming. Creating an invoice draft is separate from approving its send. Expired approvals require a fresh preview.
If an action times out, open Recent work & saved chats and check its status before trying again. A completed receipt means the recorded action has already completed. An uncertain result requires checking the affected record or delivery history; it is not permission to repeat a send or payment.
Not every change has an approval card. For example, supported internal note or routine time-tracking actions can apply directly. State the intended client, project, and action clearly.
Company context
Use this for factual business information: services offered, operating hours, support boundaries, and communication preferences. You can edit the profile, scrape a site/page, or append a supported file. Import actions can append and save context immediately; inspect the resulting text, remove unwanted material, and use Save Company Context to save corrections. Do not import passwords, API keys, or instructions that encourage invented prices or promises.
Context guides wording; it does not replace saved billing terms, live records, or review of generated work. Your selected provider processes the context included in its requests.
Usage limits and quality checks
Customer support defaults are 10 requests per user per minute, 60 per user per hour, and 100 per organization per minute. The Usage limits form applies to customer support messages, not every AI feature. Setting a limit to 0 disables that particular cap. Other features and the provider can still impose limits.
Quality & usage shows recent request counts, available token usage, and your feedback. Actual dollar cost is not available there; use the provider account for billing.
Its evaluation uses made-up records and simulated tools. It is a protocol check, not certification that every production integration, permission, or model behavior works. Evaluation reports are retained for up to seven days without raw model responses.
Saved conversations and working memory
In the admin assistant, Recent work & saved chats separates action status from conversation text:
- Action status is retained privately for up to seven days.
- Saving a conversation is an explicit choice with a consent checkbox. It saves the recent 20 messages, encrypted, for up to seven days for your account in the current organization.
- Resumed messages are historical context. The assistant must refresh live records and obtain fresh approval for sends or financial actions. Approval tokens are not saved with the conversation.
- Use the saved-chat controls to resume or delete a saved conversation.
A saved conversation is not permanent, organization-wide memory. Business context and supported private client notes are separate sources of ongoing guidance. Field Mode has its own temporary quote-recovery workflow.
Self-hosted AI
Self-Hosted AI configures an outbound gateway instead of a cloud provider: its endpoint, gateway token, and supported model ID. The page supports OpenClaw, Hermes, and a custom compatible gateway.
The page has Setup, Permissions, Workflows, and Installation help tabs. An account owner saves the gateway configuration. The portal must be able to reach that endpoint. A connection/model-list check is not a full inference test. Switching back to cloud requires choosing the cloud provider and its intended credential; the gateway token is not interchangeable with a cloud key.
If an older shared credential is marked as having unknown ownership, follow the page's re-entry prompt. Do not copy a gateway token into a cloud provider's key field.
When an answer is wrong
Check the selected provider/model, saved context, record dates, and the source records first. Ask the assistant to identify missing information instead of filling gaps. Review generated prices, quantities, recurring charges, and recipient details before saving or sending. Use the feedback controls or Report a Problem for a reproducible failure.