Skip to main content
Documentation

Reviewing AI Classifications and Corrections

When the AI's classification looks wrong, here's how to correct it — and what your correction does for future accuracy.

Reviewing AI Classifications and Corrections

The AI gets classification right most of the time. When it doesn't, your correction is what makes it get it right next time. This page explains the review flow.

When AI gets it wrong

Look at the document detail panel. The classification appears as a chip near the title — e.g., “Invoice” with a confidence indicator.

If the classification is wrong:

  1. Click the chip
  2. Choose the correct category from the dropdown
  3. Optionally add a note explaining
  4. Confirm

The correction is captured immediately. The document moves to its correct type; downstream automations (retention policy, default folder, workflow) re-evaluate.

The review queue (for admins)

Admins can review documents with low classification confidence (below the tenant's threshold, default 0.70) in a dedicated queue:

/admin/classification-review

Each entry shows:

  • The document preview
  • The AI's predicted category and confidence
  • The top 2-3 alternative categories the AI considered
  • A quick-action panel (Confirm / Reclassify / Skip)

Bulk-confirm or bulk-reclassify when many documents share the same correction pattern.

What corrections do

Every correction is logged with:

  • The original AI prediction (kept in the document's audit trail)
  • The corrected category
  • The user who corrected
  • The timestamp

Corrections feed into the next quarterly retraining cycle. Over time, the AI learns the patterns your tenant cares about — including custom categories your admins have defined.

Custom categories

Beyond the 25+ built-in categories, your admins can define tenant-specific ones (Business+ plans). Examples:

  • “Loan Restructuring Memo” (for a SACCO)
  • “Donor Disbursement Report” (for an NGO)
  • “Tea Plucker Daily Log” (for an agribusiness)

For the AI to classify into these reliably, the tenant typically needs to label 50+ examples. Below that, the category exists but classification confidence stays low (review queue handles them).

Confidence threshold

The tenant-wide confidence threshold (default 0.70) determines:

  • Below the threshold: AI classification is advisory; the review queue surfaces the document
  • Above the threshold: AI classification is accepted silently; still revisable by any user

Tenant Admins can raise or lower the threshold under Admin → Classification Settings.

  • Raise it (e.g., 0.85): more documents go to review; higher overall confidence in classifications
  • Lower it (e.g., 0.55): fewer documents need review; trust the AI more

What happens to incorrect classifications you don't catch

They sit with the wrong classification, possibly trigger wrong workflows, may apply wrong retention. The data still works — the document is searchable, the audit log still records who did what — but the category-driven automations may be off.

Periodic review of the lowest-confidence classifications is a good Records Officer habit.

Accuracy benchmarks

We publish quarterly per-category accuracy. Look at Admin → AI Governance → Accuracy to see how your tenant compares to the platform-wide baseline. Categories where your tenant's accuracy is materially lower are candidates for additional training data (more corrections + custom rules).

Rejoining the server...

Rejoin failed... trying again in seconds.

Failed to rejoin.
Please retry or reload the page.

The session has been paused by the server.

Failed to resume the session.
Please retry or reload the page.