I was reviewing a set of financials recently that had been prepared with AI. Something didn’t look right, so I started digging.
There was a small discrepancy in the numbers. Mistakes happen. A human could’ve made the exact same error. But it made me think about the processes that we have to rework when we introduce AI into our firms.
You can’t just add AI to the process you already have and expect everything to work the same way. And I think a lot of firms are going to start learning that the hard way.
When a Human Makes a Mistake, We Ask Why
Think about what happens when a person on your team makes a mistake. You don’t just fix it and move on. You ask what happened. Was something missed during the close? Were the accounts reconciled? What was reviewed before the work went out the door? Most importantly, what needs to change so you don’t have the same problem next month?
That’s basic accounting. But something weird happens when AI enters the equation. We start thinking that because the agent did the work, the work is done. It’s not!
If you’re signing your name to something, you still need to know it’s right. That means someone needs to understand what the AI did, what it didn’t do and how its work was checked before it got to you. And that’s bigger than adding “review AI output” to your existing checklist.
AI Changes the Assumptions Behind the Workflow
Our accounting processes were built around humans. Someone prepares the work. Someone reviews it. We know what each person is responsible for, what they should check and when something needs to be escalated. Over time, we train people to develop enough judgment to recognize when something just doesn’t look right.
Now we’re putting AI agents into those workflows, and that changes the assumptions the workflows were built on. So the process has to change, too.
Firms need to start asking:
Where does the AI belong?
What should it be responsible for?
Where does a human need to step in?
What does that person need to review?
How do they know when the AI is wrong?
If something goes wrong, how do we figure out whether the problem came from the AI, the human or the process itself?
Those aren’t technology questions. They’re workflow and control questions.
Using AI Is Not the Same as Reviewing AI
There’s also a training issue here that I don’t think firms are talking about enough. You can’t teach a young accountant how to use an AI agent and assume you’ve trained them to review its work. Those are completely different skills.
If AI is doing more of the preparation, we have to be even more intentional about teaching people what good work looks like. They need to understand why something should tie out, what should be reconciled and what questions to ask the AI agent when the numbers don’t make sense.
Otherwise, how are they ever going to know when the AI is wrong?
Don’t Drop AI Into Step Four
This is why I think AI implementation is going to require firms to rethink a lot more than their technology stack. We need to rethink our workflows, our review procedures, our training and even what we consider a finished work product.
Don’t take a 20-year-old workflow, drop an AI agent into step four and call yourself an AI-enabled firm. Start over. Look at what the technology can do, where it needs human judgment and what controls need to exist between the two. Then build the process around that reality.
Because whether a mistake comes from a human or an AI almost doesn’t matter. If bad work can make it all the way through your workflow without being caught, you don’t have an AI problem. You have a process problem.
AI didn’t create it. It just made it impossible to ignore.
