> ## Documentation Index
> Fetch the complete documentation index at: https://dialnexa.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# AI/ML API

> Connect DialNexa calls to AI/ML API for record, request, case, lookup, approval, workflow run, or operational task workflows.

AI/ML API provides a suite of AI models and solutions for various tasks, including text generation, image processing, and more.

<Note>
  Use AI/ML API with DialNexa when the call creates a specialized follow-up that needs owner, urgency, and clear operational context.
</Note>

## Where AI/ML API fits in a DialNexa workflow

AI/ML API should receive DialNexa output when the conversation affects a record, request, case, lookup, approval, workflow run, or operational task. The handoff should explain what the caller asked for, what DialNexa learned, which record or object is affected, and who owns the next step.

<CardGroup cols={2}>
  <Card title="Measure recurring issues" icon="check-circle">
    Tag repeated call reasons so operations can see where customers keep getting stuck.
  </Card>

  <Card title="Create structured handoffs" icon="check-circle">
    Capture caller identity, request, affected object, owner, urgency, and decision needed.
  </Card>

  <Card title="Route niche requests" icon="check-circle">
    Send specialized calls to the person who knows the system, product, policy, or customer context.
  </Card>

  <Card title="Build review queues" icon="check-circle">
    Hold unclear, sensitive, high-value, or low-confidence cases for human review.
  </Card>
</CardGroup>

## What DialNexa should capture for AI/ML API

* Caller identity, account, source, owner, category, urgency, and related object ID
* Call summary, requested outcome, missing information, blocker, and promised next step
* Status, priority, deadline, approval requirement, duplicate key, and review reason
* Transcript link, recording link, DialNexa call ID, CRM link, ticket link, and file links
* Sensitive-data flag and routing note for human review

## High-value AI/ML API workflows

<AccordionGroup>
  <Accordion title="Owner should be alerted quickly">
    For this workflow, DialNexa should send AI/ML API a concise, action-ready handoff: matched caller, affected record, reason for the update, urgency, owner, next step, and links to call evidence.
  </Accordion>

  <Accordion title="Caller creates an operational request">
    For this workflow, DialNexa should send AI/ML API a concise, action-ready handoff: matched caller, affected record, reason for the update, urgency, owner, next step, and links to call evidence.
  </Accordion>

  <Accordion title="Specialist review is required">
    For this workflow, DialNexa should send AI/ML API a concise, action-ready handoff: matched caller, affected record, reason for the update, urgency, owner, next step, and links to call evidence.
  </Accordion>

  <Accordion title="Missing information blocks progress">
    DialNexa should attach the relevant file or visual evidence, summarize what the caller says it proves, and mark the review owner in AI/ML API. Sensitive files should stay behind restricted links.
  </Accordion>

  <Accordion title="Approval is needed before action">
    For this workflow, DialNexa should send AI/ML API a concise, action-ready handoff: matched caller, affected record, reason for the update, urgency, owner, next step, and links to call evidence.
  </Accordion>

  <Accordion title="Recurring issue should be categorized">
    DialNexa should write the symptom, expected behavior, actual behavior, affected area, business impact, and evidence links into AI/ML API. A teammate should be able to triage the issue without replaying the call.
  </Accordion>

  <Accordion title="Use create message">
    Use create message only when DialNexa has a matched caller, a clear destination object, and enough call context to justify opening a new operational record. If the caller is unclear, route to review instead of creating noise.
  </Accordion>

  <Accordion title="Use update message">
    Use update message when the caller changes a field, status, owner, date, priority, note, consent choice, or next step on an existing AI/ML API record. Include the old value, new value, and reason from the call.
  </Accordion>
</AccordionGroup>

## Workflows that pair AI/ML API with other integrations

* [AI/ML API](/integrations/ai_ml_api) + [Gmail](/integrations/gmail): Gmail for approved customer follow-up.
* [AI/ML API](/integrations/ai_ml_api) + [Google Calendar](/integrations/googlecalendar): Google Calendar for scheduled callbacks.
* [AI/ML API](/integrations/ai_ml_api) + [HubSpot](/integrations/hubspot): HubSpot for customer context.
* [AI/ML API](/integrations/ai_ml_api) + [Slack](/integrations/slack): Slack for owner alerts.
* [AI/ML API](/integrations/ai_ml_api) + [Google Sheets](/integrations/googlesheets): Google Sheets for review queues.
* [AI/ML API](/integrations/ai_ml_api) + [Zendesk](/integrations/zendesk): Zendesk for support follow-up.
* [AI/ML API](/integrations/ai_ml_api) + [Google Docs](/integrations/googledocs): Google Docs for operational briefs.

## Implementation notes

* Use the DialNexa call ID as the idempotency key before running AI/ML API actions.
* Write a short operational summary into AI/ML API and link to the full transcript or recording for audit.
* Map required fields before launch: destination object, owner, status, urgency, next step, and record URL.
* Create review paths for low-confidence matches, sensitive requests, high-value customers, and actions that change money, access, legal terms, or customer commitments.

## FAQs

<AccordionGroup>
  <Accordion title="How should extraction errors be handled?">
    Store confidence and route missing or conflicting fields to review rather than silently updating downstream systems.
  </Accordion>

  <Accordion title="When should an AI answer fall back to a human?">
    When source context is missing, confidence is low, the caller disputes the answer, or the next step changes money, access, or legal commitments.
  </Accordion>

  <Accordion title="What should be measured over time?">
    Intent accuracy, extraction accuracy, fallback rate, review overrides, bad answers, and customer outcomes after AI-generated follow-up.
  </Accordion>

  <Accordion title="Should AI outputs act without review?">
    Only for low-risk workflows with clear confidence thresholds. Account changes, money, access, legal, and sensitive support cases need review.
  </Accordion>

  <Accordion title="What should be logged for AI decisions?">
    Prompt version, source context, model or workflow ID, confidence, output, call ID, and reviewer when applicable.
  </Accordion>

  <Accordion title="How do we reduce hallucinations?">
    Use approved knowledge sources, cite source records, set fallback behavior, and route low-confidence answers to humans.
  </Accordion>

  <Accordion title="Can DialNexa write follow-up messages with AI?">
    Yes, but drafts should use the actual call outcome, approved tone, and customer context. Sensitive messages should be reviewed.
  </Accordion>
</AccordionGroup>