Skip to main content

Recommendation types

The AIMI Tier 1 Agent stores every recommendation as a record of the MAI_AIMI_Response_Log__c object. The record type identifies the kind of recommendation, and a MAI_Review_Outcome__c field tracks whether a specialist has Accepted or Rejected it. This page is a reference for each recommendation type, what the agent produces, and what gets created on accept.

The AIMI Response Log object

MAI_AIMI_Response_Log__c (label AIMI Response Log, name format ARL-{0000}) is the single home for the agent's output. Field history tracking and reporting are enabled, so agent activity can be audited and reported on.

Common fields across recommendation types:

FieldPurpose
MAI_Interaction__cThe Interaction (Case) the recommendation belongs to
MAI_Review_Outcome__cAccepted or Rejected, set when a specialist acts on the recommendation
MAI_Recommendation__cThe agent's verdict for AE and PQC recommendations (Recommended or Not Recommended)
MAI_Justification__cThe justification the agent generates for its PQC or AE recommendation
MAI_Rejected_Rationale__cThe reason a recommendation was rejected (required when rejecting in the review panel)
MAI_Parent_Response_Log__cLinks related response logs together
MAI_Country__cCountry context used for the recommendation and for resolving the suspect product
MAI_Request__c, MAI_Adverse_Event__c, MAI_Product_Quality_Complaint__c, MAI_Fulfillment__cOn accept, the log's lookup to the record it created, powering Open Record and reporting

The object has five active record types, described below in the order they appear in the review panel.

note

Reporting on MAI_AIMI_Response_Log__c is enabled, so agent activity can be monitored through your own reports and dashboards.

Interaction Summary

Record type: MAI_Interaction_Summary

The agent generates an analyst-ready summary of the Interaction (requester, products, delivery preference, context, and a factual recap of any evident event), stored in MAI_Interaction_Summary__c. The summary recaps rather than classifies. It does not label or assess a potential AE or PQC; the AE and PQC recommendations handle that.

On accept: there is no separate record to create; the summary itself lives on the response log. Use Add to Notepad to copy it into the Interaction notepad; the response log is marked Accepted. See AIMI Recommends.

Adverse Event

Record type: MAI_Adverse_Event

The agent evaluates the Interaction for a potential adverse event and produces:

  • MAI_Recommendation__c, Recommended or Not Recommended
  • MAI_Justification__c, the agent's justification for the recommendation
  • MAI_Narrative__c, a de-identified clinical case narrative, generated when the recommendation is Recommended. The agent captures clinical substance (dose, route, symptoms, the event) while removing identifying detail (exact names, dates, and locations) for a safety reviewer to restore downstream.
  • MAI_Product__c, the suspect product, when one is resolvable

On accept: when the recommendation is Recommended, a MED_Adverse_Event__c record is created. When a product is also resolvable, a related MED_AE_Drug_Information__c record is created for its matching MED_Product__c record. The product is matched by name or trade name for the Interaction's country (MED_Country__c), falling back to a country-agnostic base product when no country-specific match exists.

Product Quality Complaint (PQC)

Record type: MAI_PQC

The agent evaluates the Interaction for a potential product quality complaint and produces:

  • MAI_Recommendation__c, Recommended or Not Recommended
  • MAI_Justification__c, the agent's justification for the recommendation
  • MAI_Complaint__c, the complaint description, generated when the recommendation is Recommended
  • MAI_Product__c, the product, when resolvable (matched by name or trade name for the Interaction's country MED_Country__c, falling back to a country-agnostic base product)

On accept: the MED_Product_Quality_Complaint__c record is created, with the resolved product set directly on it.

Request

Record type: MAI_Request

The agent recommends the Request(s) to create for the Interaction. The card surfaces:

  • MAI_Request_Type__c, MAI_Product__c, MAI_Question__c
  • MAI_Category__c and MAI_Sub_Category__c (the suggested category/sub-category, with picklist dependencies respected)
  • MAI_Suggested_Documents__c, documents the agent recommends to fulfill the Request

On accept: the Request is created and the selected documents are attached. When a document has a content-version reference (its MCM document version), it is attached through the standard MIC content path; otherwise a direct MED_Request_Document__c insert is used. The chosen document keys are recorded in MAI_Accepted_Documents__c.

note

On accept, a picklist value the agent proposed (Type, Category, Sub-Category, Country) is applied only if it matches an entry in the org's picklist, by API value or label. A value the org's picklist does not contain is left blank. Make sure the org's picklists include the values the agent can produce.

Fulfillment

Record type: MAI_Fulfillment

A record type for the agentic fulfillment workflow, used to coordinate document fulfillment for the Interaction.

On accept: a MED_Fulfillment__c record is created and linked in the response log's MAI_Fulfillment__c field, and the review panel handles the card like the other types.

In the current release no subagent generates a Fulfillment recommendation, so these logs are not produced yet; the record type and its accept handling are in place for a future release.

How recommendations are generated

Each recommendation type is produced by a GenAI prompt template that a subagent runs. The subagent takes the structured JSON the template returns, and its flows map that JSON onto the response-log fields. The order the subagents run in is driven by MIC's prompt-option configuration. For the orchestration details, see the AIMI Tier 1 Agent.

note

The agent skips re-running on an Interaction that already has response logs, so changing context (such as Country) on an already-processed Interaction does not duplicate recommendations.