Review Queue: Decide on AI Suggestions

Work the Review Queue — where below-threshold AI category and merchant-name suggestions wait for you to accept, edit, or reject, with every decision teaching Memory.

2 min read

On this page

How the queue is organized
What accept and reject actually do
Bulk accept
Housekeeping
Related guides

The Review Queue is where AI suggestions that were not confident enough to apply automatically wait for a human decision. When the AI Categorizer or Merchant Cleanup scores a suggestion below the confidence bar, the field is left empty and the suggestion lands here instead — the AI never guesses quietly into your books.

Open it from the Enrichments page — the Review Queue link, at …/enrichments/review.

How the queue is organized#

Suggestions are grouped by merchant pattern, because that is the unit a decision actually covers: accepting "starbucks → Dining" resolves the pattern, and every future record matching it fills automatically. Each row shows:

the normalized merchant pattern the suggestion is about,
the target field (for example Category or Merchant Name),
the suggested value, editable inline before you accept,
the model's confidence as a percentage, color-coded,
Accept and Reject actions.

The Review Queue: a header reading 4 suggestions waiting, accepting teaches Memory for future records, with an Accept all at or above 90 percent button, above suggestions grouped by merchant pattern — sq blue bottle cof with Meals and Entertainment at 93 percent and Blue Bottle Coffee at 97 percent, amzn mktp us with Office Supplies at 62 percent, and gusto pay 872231 with Payroll at 78 percent — each row with an editable suggested value, a color-coded confidence, and Accept and Reject buttons. — Suggestions grouped by merchant pattern, each with its editable value, confidence, and Accept/Reject — plus the bulk accept.

What accept and reject actually do#

The value (or your inline edit of it) is taught to Memory with full user-edit credit — it applies to matching records immediately, starting with the next sync. One accept covers every future record with that pattern.

Reject

The pattern is suppressed for that field: the AI stops suggesting it, and nothing is ever auto-filled for it. Rejection is how you say "leave this one to me".

Both decisions teach Memory — the queue is a teaching surface, not just a to-do list. If your workspace does not have a Memory enrichment yet, your first accept creates it automatically: resolving a suggestion is a request for learning.

Decisions apply to future records only

Accepting a suggestion never rewrites rows already delivered to your connected sheet, database, or table — it teaches BankSync what to write from now on. If you want past rows updated too, edit them in the destination; Memory learns from those edits as well.

Bulk accept#

The Accept all ≥90% button resolves every open suggestion at or above 90% confidence in one pass — the fastest way to clear a backlog after a big first sync, leaving only the genuinely uncertain rows for your eyes. Inline edits you have made are honored by the bulk accept.

Housekeeping#

You do not need to keep the queue at zero for correctness — unresolved suggestions simply mean the field stays empty:

Open entries expire after 90 days, and the queue is capped at 2,000 open entries per workspace — it cannot grow without bound.
A resolved entry never reappears: once accepted or rejected, the same suggestion is not re-queued even if the AI encounters the pattern again.
The queue count appears in your Digest, so a growing backlog surfaces on its own.

The queue shrinks itself over time

Every accept teaches Memory, and Memory fills before AI runs — so patterns you have decided on never come back, and the AI spends its confidence on genuinely new merchants. An active workspace converges toward an empty queue.

AI Categorizer: the confidence threshold that decides what lands here.
Merchant Cleanup: the other source of suggestions in the queue.
Memory: auto-learn: where accepted decisions live and how learning works.