
How to Deploy Auditor-Approved AI
A topic that often comes up in our conversations with Big 4 auditors and internal SOX teams is how AI fits within standard control frameworks. Specifically, how should AI agents and humans share the work of preparing accrual entries and routing them for approval?
Would one human approver be enough, or would the process need two?
Is there a way to avoid requiring duplicate approvals and still keep the separation of duties that SOX requires?
What guidance does the auditor provide in this scenario?
These types of questions are no longer edge cases. Finance teams deploying AI are grappling with these workflow design decisions every day. Until the standards are settled, we believe the right move is to pick the most defensible design available today and document why.
The dichotomy of AI
AI models today are inherently non-deterministic. Ask AI the same question twice and you likely won’t get the same reply twice. The wording, or even the answer, can change. However, financial controls require a system that demands the opposite: defined rules, repeatable results, and support that a third party can follow from start to finish.
As a finance team leader, you sit where those two worlds meet. You own the control environment, and your team is trying to use efficient AI teammates to unlock capacity. But these new teammates don’t quite work with the systems the controls were designed for.
One of the hardest parts of adopting AI in finance is fitting probabilistic systems into accounting frameworks built for certainty.
In our work with customers and SOX teams, we’ve seen the teams that get this right follow these three principles:
1. Reinforce segregation of duties
Segregating duties is more critical than ever.
The classic concept is simple: have two parties with independent agency, one prepares and one approves.
The same holds true with the introduction of AI, but the segregation of duties has to be established to give the agent autonomy from the human approver. If a human directly steers the AI, as with a chatbot, the AI has no independent agency. For AI to truly serve as a preparer, it has to be equipped with independent methodology that a human approver cannot override. Any change to its methodology needs a separate human reviewer.
Here’s how we designed the system in one customer deployment, a bank reconciliation agent.
In this deployment, the agent proposed new journal entries, and a human approved them. The approver had no way to change the agent’s logic, so the agent could act as the sole preparer.
Here we elevated the human to approver.
In the AI era, this should be your new framework for segregation of duties:
- Two distinct profiles, with AI-as-preparer counting as one of them.
- The same entity (whether human or AI) cannot propose and approve.
- Changes to the AI’s logic need their own preparer and approver to keep the agent independent.
- Admins and super users cannot approve and post their own entries.
- No user access in the AI platform exceeds their access in the ERP or data warehouse.
- Perform quarterly look-back reviews to flag any inappropriate transactions.
2. Reasoning transparency
Teams need stronger audit evidence and a visible reasoning trail.
In the traditional audit standard, you attached the invoice, and anyone reviewing the entry could follow the logic from there.
Expectations are higher for AI. When AI takes on the role of preparer, you need the backing data and the documented reasoning behind every decision, stored and retrievable long after the close.
To review an AI output, you need three things: reasoning on how it got the answer, what data it used, and who approved it. This addresses the black box uncertainty that auditors are wary of when it comes to AI, especially generative AI.
Take another customer example where we set up an AI agent to act as an independent preparer at a public company. We deployed a system where any change made to the agent’s logic was automatically recorded and logged in the customer’s IT project management system. Every change to the logic needed its own preparer and approver.
We also added a natural-language reasoning trail to each transaction; a controller or auditor can refer to it when tracing or vouching for it. The same record also showed who approved what and when, which replaces a manual checklist with evidence. SOX-compliant audit logs also showed who accessed each record, and when.
Here’s what your AI-era audit trails should look like:
- A written standard operating procedure (SOP) in place before testing, because the testing will be measured against it.
- A quarterly review of all in-scope changes, with tickets attached and a recorded sign-off.
- A log of every change to the AI’s reasoning, showing the date, a plain-language description, the before-and-after state, who or what made the change, and, most importantly, why the change was made.
3. Process flexibility
Build for an evolving rulebook.
In our experience, the Big 4 are willing to let finance teams push the boundaries on technology and efficiency with AI. The trade-off? You need to be ready for auditors to reassess guidelines as the industry adjusts to this new era where AI capabilities and processes can change quickly.
The rules for evaluating AI in finance are still being written, and the model of what good AI governance looks like is constantly evolving. This means the standard you are measured against may shift. A governance framework built only for today’s interpretation isn’t enough. It needs to be configurable as interpretations change.
Structure your system to change as AI guidelines evolve:
- Set rules at a granular level. Establish thresholds and approval requirements by entity, account or entry type. If auditors tighten expectations in one area, you simply need to change that area, not the entire process.
- Create roles that are easy to restructure. Segregation-of-duties expectations for AI are still forming, so your roles need room to shift with them.
- Capture the full context of every decision. The system records the full context at every point in time: the control structure, the agent’s methodology, the approval rules and the auditor guidance in effect when the decision was made. You must be able to show that each past action met the standard that applied when it was taken.
Your pre-deployment checklist
Before your AI system goes live, use these three questions to shape your team's decision framework:
- Are the system’s rules enforced regardless of whether the approver agrees with them, and is it structurally impossible for any one identity, human or AI, to both propose and approve the same entry?
- Can we reproduce this output and show the reasoning behind it a year from now?
- Can we change how this system works as interpretations evolve, and show that every change was tested and logged?
Teams that can answer yes to all three questions can move from pilots to production with confidence.


