AI providers and model jobs

Agents configure models on Assistant → Models by job. A provider supplies external models or a compatible model endpoint; an assistant supplies identity, function and permissions. The model connection alone does not create a new team member.

Connect Chat

1. Find Chat provider and select OpenAI, Claude, Gemini, Azure OpenAI or OpenAI-compatible.
2. Fill the provider's displayed fields. Standard cloud providers need your organization's API key. Azure needs its endpoint and deployment details. A compatible server needs its Base URL and any key required by that server. Use values from the provider account you control.
3. Select Load models and choose Model. Refresh reloads the provider's list. Show all models includes choices outside the default chat filter; Show dated versions exposes dated variants. A model listed as not a chat model or lacking tool support is not suitable merely because it appears in the list. Azure uses its deployment name rather than a public model name.
4. Read What will be sent, and to whom. The destination and information-sharing statement change with the provider and privacy settings.
5. Review Also tell the assistant the person's department, off by default. Set organization-specific ID-redaction patterns, one regular expression per line, if required; matching text becomes [id] before sending. Redaction does not clean image pixels.
6. Select Test. The Chat test sends only a request to reply with OK, not a real ticket. Read the actual answer or error. This establishes that the test request worked, not that all future responses are correct.
7. Tick the information-sharing agreement when required and select Save and use. For an existing active connection the button is Save changes.
8. Check the active Chat card, connection light, model and Recent activity. Keep a copy of the key in your own password manager; Empeiro only shows a saved-key hint.

A provider test can contact a real external service and use its quota. The test's sample content is distinct from invented employees or directory imports; no real tenant should be replaced with simulated records.

Change or remove a saved provider

Select the provider's saved connection to review it. Replace the key opens a new key field; Keep the key on file cancels replacement. Changing the model or endpoint requires Save changes. To activate another saved provider, review its agreement and select Use this one.

Disconnect stops using the active Chat connection. Forget removes the selected saved connection after confirmation; read any warning about jobs using it. These actions do not cancel a provider subscription or revoke the key at the provider. Rotate compromised credentials in the provider account too.

Configure each job

Expand What it receives, returns and may ask on a job card before choosing its provider. The card explains its inputs, outputs, permitted cross-job requests and where its results appear.

1. In the chosen card, select its provider. Key on file reuses that provider's saved key; otherwise enter its API key and any endpoint or Base URL requested.
2. Load and choose the job's Model, or enter Azure's Deployment name. Show all models can expose unsuitable models marked not for this job; select a compatible one.
3. Read the job-specific sharing sentence and tick I understand, and my organization agrees to this when required.
4. Select Test and inspect Sent and the answer/error. These tests use sample questions, text or an image, rather than ticket content.
5. Select Save and use or Save changes. Verify its status and selected model.

Decision confidence

Under Decision's Act above this confidence, otherwise “unsure”, review each judgment's label and Unsure fallback. Set the threshold from 0 to 1 and select Save thresholds. The interface shows defaults beside each judgment. A higher threshold generally requires stronger confidence to act; an unsure fallback can leave work for a person rather than making the decision. Inspect routing reasons, function logs and Activity to understand individual outcomes.

Index status and rebuild

The Index card shows Up to date, Building, Stopped or Text-only search, with indexed/total items, last run, tokens and any error. After selecting or changing the embedding configuration, review the status and use Rebuild to regenerate searchable representations when needed. The action can consume provider resources; wait for completion before relying on the refreshed related results.

Ask Decision whether a resolution is a real fix before indexing it optionally filters resolutions through Decision. Set its daily limit and select the adjacent Save. That allowance protects this classification cost and is separate from an assistant's background budget. With no index model, assistants retain text-only search.

Permit pictures

Pictures requires a ready vision configuration before Let the assistants look at pictures people attach can be enabled. The organization switch is off by default and saves when toggled. Even when enabled, each image is sent only when the person ticks Let the assistant look at this on that attachment. Text redaction cannot remove names or secrets visible inside an image: crop those before sharing.

If the switch is unavailable, read the reason below it and fix the Pictures provider/model first. An attached image is not automatically consent to send it to an AI service.

Usage and troubleshooting

Cards show Today and This month, with calls or items processed and input/output tokens. Review Reports for longer comparisons and cost estimates; the provider account remains authoritative for billing, quotas and rate limits.

Configure team behavior separately in Assistant functions and permissions.