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

# Reviewing AI-detected denials

> What the AI catches, what it misses, and how to handle low-confidence cases.

<Info>
  After upload, every denial lands in the **Denials inbox** with an AI-extracted classification, deadline, and strategy recommendation. Your job is to review, adjust, and approve — not re-enter the data.
</Info>

## The inbox view

The inbox shows one row per detected denial with these columns:

| Column            | What it means                                                                    |
| ----------------- | -------------------------------------------------------------------------------- |
| **Patient**       | From the EOB or claim. PHI columns are masked in list views (e.g. `J***n D***`). |
| **Claim ID**      | As extracted from the document.                                                  |
| **Payer**         | Matched to your payer list.                                                      |
| **Denied amount** | From the claim.                                                                  |
| **Denial type**   | Classified into one of 10 categories (see below).                                |
| **Deadline**      | Days until the appeal deadline. Red if \< 7 days.                                |
| **Confidence**    | AI confidence in the classification.                                             |
| **Status**        | New, Reviewing, Appealed, Outcome.                                               |

## Denial types

| Type                       | Examples                                                          |
| -------------------------- | ----------------------------------------------------------------- |
| `medical_necessity`        | "Service not medically necessary", "experimental/investigational" |
| `prior_authorization`      | "No prior auth on file", "auth denied"                            |
| `coding_error`             | "Invalid CPT/HCPCS", "bundling edit", "unbundling"                |
| `timely_filing`            | "Claim submitted after filing deadline"                           |
| `out_of_network`           | "Provider out of network"                                         |
| `coverage_exclusion`       | "Service not a covered benefit"                                   |
| `coordination_of_benefits` | "COB information needed"                                          |
| `patient_responsibility`   | "Applied to deductible", "patient responsibility"                 |
| `duplicate`                | "Duplicate claim"                                                 |
| `other`                    | Anything we couldn't confidently bucket                           |

## Confidence score

| Score      | What we do                                                                                       |
| ---------- | ------------------------------------------------------------------------------------------------ |
| **≥ 85%**  | Auto-route to your inbox as "Ready for review".                                                  |
| **70–84%** | Flagged "Check — medium confidence". The AI is reasonably sure, but you'll see an extra warning. |
| **\< 70%** | Routed to **Human-in-the-loop review** (see below). Not surfaced to your inbox until reviewed.   |

## Human-in-the-loop review

Denials below 70% confidence — or where the AI detects conflicting signals — are automatically queued for review by a trained Denialbase reviewer.

<Steps>
  <Step title="Triage">
    The reviewer reads the document, confirms or corrects the classification, and adjusts the deadline if needed.
  </Step>

  <Step title="Notes">
    The reviewer adds any context (e.g. "payer used an unusual CARC code; reclassifying as prior-auth based on attached letter").
  </Step>

  <Step title="Release">
    The denial is released to your inbox with a "Reviewed" badge and the reviewer's notes attached.
  </Step>
</Steps>

Typical HITL turnaround is under 1 hour during business hours.

## Adjusting a classification

You can override the AI on any denial:

* Click the **Type** dropdown on the detail view to reclassify.
* Edit the **Deadline** if the AI picked it up incorrectly.
* Add **Clinical notes** that won't change the appeal draft but will be visible to reviewers and to the AI on re-run.

Every override is captured and fed back into our training process — this is how the classifier gets better for your practice over time.

## Bulk actions

<CardGroup cols={2}>
  <Card title="Assign" icon="user-plus">
    Select multiple denials → **Assign to** → pick a teammate. Assignees get an email and in-app notification.
  </Card>

  <Card title="Prioritize" icon="flag">
    Mark as **Urgent**, **Normal**, or **Low**. Urgent denials are pinned to the top of everyone's inbox.
  </Card>

  <Card title="Snooze" icon="clock">
    Hide until a specific date. Denials unhide automatically a week before the appeal deadline.
  </Card>

  <Card title="Archive" icon="box-archive">
    Archive if you've decided not to appeal. Archived denials are included in reporting but out of the main inbox.
  </Card>
</CardGroup>

## Reading the strategy recommendation

For each denial, you'll see a **Strategy** box with:

* **Recommended path**: peer-to-peer, formal written appeal, external/IRO review, or "do not appeal" (e.g. duplicate, valid patient responsibility).
* **Citations**: specific plan language, state regulations, or federal rules (ERISA, ACA) supporting the appeal.
* **Estimated overturn probability**: based on historical outcomes for similar denials in your practice and the broader Denialbase dataset.
* **Prerequisites**: e.g. "attach operative report and medical necessity letter before filing".

<Info>
  The strategy is advisory. You're always in control of what to appeal and how — the AI's job is to show you the path with the evidence, not make the decision.
</Info>