Use risk-based review levels for AI-assisted work 

Not every piece of AI-assisted work deserves the same level of review.

If you treat everything as high risk, senior people will spend their lives checking low-value work. If you treat everything as low risk, something important will eventually slip through.

This is the basis of a risk-based review for AI-assisted accounting work.

Low consequence: internal summaries, formatting, meeting notes, first drafts, and routine administrative work. The preparer checks accuracy, context, and tone.

Moderate consequence: client-facing drafts, working paper summaries, management reports and advisory notes where assumptions or source information matter. These need source checking, assumption review, and manager-level review where appropriate.

High consequence: tax positions, technical advice, financial reporting judgements or material client recommendations. These require experienced human review, evidence checking, clear assumptions, and formal sign-off by the right person.

The purpose is to put human attention where the consequence of being wrong is highest. 

A practical AI review checklist for accounting firms

Before AI-assisted work moves to senior review or client delivery, the person preparing it should be able to answer these questions:

1. Inputs: Did I give the AI the right client information, period, source documents and context?
2. Completeness: Is anything important missing?
3. Assumptions: Has the AI filled any gaps or made something sound certain that is not?
4. Figures: Have material numbers, dates, and calculations been checked?
5. Sources: Can important claims be traced back to reliable source material?
6. Client impact: Could this affect a client's decision, tax position, financial report or advisory recommendation?
7. Professional judgement: Does any part of this need a more experienced person?
8. Sign-off: Is the right person approving the work based on risk and consequences?

That is a better accounting firm's QA process than sending everything upwards and hoping the reviewer catches what matters.

An AI review checklist does not replace judgement. It stops senior judgement from being wasted on problems that should have been caught earlier.

Trust should live in the workflow, not in one person's memory 

If quality depends on your memory, you do not have consistent quality assurance. If the owner or partner remains in the final safety net for everything, you do not have a scalable AI workflow for your accounting firm.

You have one experienced person holding the system together.

Standards need to be visible. The team needs to know what good looks like, what must be verified, what gets escalated, and who owns the final decision.

Take one workflow your team already uses. Define the standard. Decide where AI can assist. Decide what the preparer checks, what requires manager review and what needs formal sign-off.

Then make that process more repeatable. That is how trust moves out of one person's head and into the way the firm works.

The goal is not blind trust in AI 

I do not think the goal is to trust AI more. The goal is to build an AI review process your accounting firm can trust.

You should know where AI is useful, where professional judgement belongs, what the team checks before work moves forward, and why something needs senior review.

Without that system, faster AI-generated work can simply create a bigger pile of partner review. With it, AI can help create real capacity without lowering the standard.

That is what AI review should achieve.

This week, choose one AI-assisted workflow in your firm and ask your team: What has to be checked before we are willing to put our name on this?

If you get three different answers, you have found your first QA gap. 

PREVIOUS POST

Slowing Down to Speed Up: How Business Owners Can Reduce Rework