Halfway through a podcast conversation about why his accounting firm has not adopted AI, Josh described the thing he would actually use. He was not pitching a product. He was thinking out loud.
This is the other half of a problem I wrote about separately: people are already using AI at work without approval, largely because no sanctioned alternative exists. What follows is what a sanctioned one would look like.
He wanted to hand a project to an AI. Let it do the whole thing. And then have a screen that says: your AI has finished this project, here's an audit log and the result, hit approve to commit the changes.
That is it. That is the whole design.
I've thought about it a lot since, because it quietly solves the problem that most AI tooling refuses to admit exists.
What it gets right
The AI does the full task. Not a suggestion, not a draft, not a copilot hovering over your shoulder. The actual work, start to finish. This matters because half-automation is often worse than none. If I still have to understand every step to supervise every step, I've not saved anything.
Nothing commits without a human. The work exists in a staging state until a person says go. The AI has authority to prepare and no authority to finalize.
The human sees the reasoning, not just the answer. This is the piece almost everyone drops. An audit log is not a receipt. It is the difference between "here's what I did" and "trust me."
Without the log, approval is theater. You are clicking a button on work you cannot evaluate, which is the same as no oversight but with worse psychology, because now you're on the hook for it.
Why so few products work this way
Because it's less impressive.
A demo where AI does the thing instantly is exciting. A demo where AI does the thing and then waits looks like the tool is not confident. Product teams optimize for the first demo. Professionals need the second product.
There is also a harder reason. Building a real audit log means your system has to be able to explain itself, which means it has to have been built in steps that can be explained. That is architectural work you cannot bolt on afterward.
Easier to ship the magic box and let the user decide whether to trust it. Which, for anyone handling other people's money or health or legal exposure, isn't a decision they're permitted to make.
The other half: scope
Josh paired the approve button with a second idea, and the two belong together.
He said the AI should not have access to closed periods. In accounting, a closed period is finished, reconciled, reported. Reaching back into one is how you get real damage.
So: let the tool work where mistakes are recoverable, and lock it out of where they aren't.
That is a general principle wearing accounting clothes. Every business has closed periods. Sent invoices. Signed contracts. Published content. Payroll that already ran. The question worth asking before you automate anything is: which parts of this can be undone, and does the automation have any business touching the parts that can't?
Most automation failures I've seen were not the AI being wrong. They were the AI being wrong somewhere it should never have been able to reach.
Approve and audit is not a compromise
There is a framing that treats human review as a temporary crutch, something we will drop once the models get good enough.
I do not think that's right, and not only for safety reasons.
Approval is where responsibility lives. When you click approve, you're saying this is mine now. That isn't friction to be optimized away. That is the thing that keeps a person accountable for work done in their name.
"The AI did it" has never been a defense and it is not going to become one. If the output goes out under your firm's name, it's your firm's output. The approve button just makes that honest, and gives you the information you need to mean it.
A system that removes the human from the loop hasn't removed the responsibility. It has only removed their ability to exercise it.
Building this yourself
You do not need a vendor. The pattern is:
- AI does the work against a copy or a staging area, never production.
- It writes down what it did, in a form a human can actually check.
- A person reviews and approves.
- Only then does anything commit.
- Scope the whole thing so it can't reach anything already closed.
None of that is exotic. Most of it is how careful engineers have handled risky changes for decades. AI didn't create the problem, it just made it easier to skip the safeguards.
Josh described this system as the thing he wants and does not have. If you're building tools for professionals, he has handed you the spec.
If you are buying them, that's your question: show me the audit log, and show me what this cannot touch.
From a conversation with Josh on J&C Unscripted.