---
title: "Discussion on \"Human-in-the-loop AI in accounting: what the reviewer should actually see\" | Hashnode"
description: "Discussion on \"Human-in-the-loop AI in accounting: what the reviewer should actually see\". Several major accounting software vendors just announced AI agents, and they lead with a human review step. One is also selling help designing the workflows. The useful questions are more specific."
url: "https://hashnode.com/posts/human-in-the-loop-ai-in-accounting-what-the-reviewer-should-actually-see/6af74502812c6fd986f230c1"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Human-in-the-loop AI in accounting: what the reviewer should actually see](https://precisionaios.hashnode.dev/human-in-the-loop-ai-in-accounting-what-the-reviewer-should-actually-see?utm_source=hashnode&utm_medium=feed)

Several major accounting software vendors just announced AI agents, and they lead with a human review step. One is also selling help designing the workflows. The useful questions are more specific.

Nov 13 · 6 min read · precisionaios.hashnode.dev  
Read the article

-   [#ai-agents](https://hashnode.com/tag/ai-agents)
-   [#human-in-the-loop](https://hashnode.com/tag/human-in-the-loop)
-   [#accounting](https://hashnode.com/tag/accounting)
-   [#audit-log](https://hashnode.com/tag/audit-log)
-   [#approvals](https://hashnode.com/tag/approvals)

## Written by

[Precision AI OS](https://hashnode.com/@precisionaios)

@precisionaios

  
[View profile](https://hashnode.com/@precisionaios)

## Responses 1

Markdown supported · ⌘↵ to post

[Ascendra Ventures](https://hashnode.com/@ascendra-ventures)

[Nov 14](https://hashnode.com/posts/human-in-the-loop-ai-in-accounting-what-the-reviewer-should-actually-see/6af74502812c6fd986f230c1/comment/6af7ca8591cbad6052ab0e26)

The test you propose is the right one: can a careful person catch a material error from the screen they are given, in the time they actually have? A yes-or-no prompt on a summary passes every vendor's "human in the loop" claim and still fails that test.

The point about queues of similar-looking items is the one I would underline for buyers. Once reviewers start approving what looks familiar, the approval step becomes a log entry rather than a control.

We dealt with this in a regulated workflow at Ascendra Ventures when we built AML SoftServe for Australian firms newly covered by AUSTRAC's Tranche 2 rules. AI drafts the suspicious-matter reports for a person to review, red flags are raised automatically, and the rest of the work runs as guided steps for due diligence, KYC and risk assessment. Most of those users had no compliance officer, so the review had to work for someone who is not a specialist, which is a harder bar than an approval button.

Did any of the announcements you looked at say how they show the agent's flagged exceptions, or is that still mostly left to the firm?
