I asked Josh what it would take for his accounting firm to let AI touch their general ledger. I expected a technical answer. Something about integrations, or accuracy, or the software not being there yet.
His answer was one sentence, and it reframed the whole conversation for me:
"It's pretty much a trust problem right now."
Then he added the part that matters. The functionality is there. He knows it's there. He could wire it up himself.
He hasn't, and won't, because he doesn't trust it yet. Not the accuracy. The custody.
Two different questions
There are two questions people conflate constantly, and separating them explains most of the confusion in this space.
Can it do the work? Increasingly, yes. Josh described a task he had that week: take a vendor bill, split the employee list by business line, get it into QuickBooks correctly. His words were that he knew exactly how to do it and could write out the instructions. An AI should be doing this.
Can I trust it with the work? Separate question. Different answer.
The first question is about the model. The second is about everything around the model: where the data goes, who keeps it, what happens when it's wrong, and whether you would be able to tell.
Almost all of the actual adoption friction lives in the second question. Almost all of the marketing addresses the first.
What trust actually decomposes into
When Josh explained what would move him, it was not vague. It was two specific things.
It can't be greedy with data. His phrasing was that he can redact all day, but there is going to be one or two times he uploads a file and forgets to redact something. He isn't describing a hypothetical failure. He is describing a certainty and asking the system to be safe anyway.
That is the correct engineering posture, by the way. Any control that depends on a human being perfect every single time isn't a control. It is a hope.
He has to be able to see what it did. Not trust that it worked. See it. Which is a different requirement than accuracy, and most tools don't offer it.
Notice that neither of those is "make the model smarter." A better model does not fix either one.
Why vendors keep missing this
Josh had a complaint I think is fair and underrated. He hasn't seen a vendor make the case for their AI in his specific field. Plenty of people talking about what you could build, or how you might use an enterprise plan. Very little of "here is the out-of-the-box tool that does this exact accounting thing."
His description of the current state was that it still feels like playing with sticks and rocks.
That is a man who wants to use this technology, who has the skill to wire it up, telling you the tooling hasn't met him where he works.
The gap isn't enthusiasm, and it isn't capability either. The gap is that most AI products are still sold as general-purpose potential, and potential isn't something a busy professional can act on. They need the specific thing, for their specific workflow, with the data question answered before they ask it.
What this means if you're building
Stop leading with what the model can do. Everyone has the same models.
Lead with the boring parts. Where does the data live. How long is it kept. What can this thing reach and what can it never reach. What does it show me when it's done. What happens when it's wrong.
Those answers are less exciting than a demo. They are also the only thing standing between a curious professional and an actual purchase.
What this means if you are buying
You are allowed to ask. You are allowed to not adopt something until the custody question has a real answer rather than a reassuring one.
And it is worth naming that waiting for trustworthy tools isn't the same as being a late adopter. Josh is not a late adopter. He is a careful one. He said he would start experimenting today if he had a local model where he could confirm the data was going nowhere.
There is a version of caution that's really just avoidance. And there's a version that's doing the job properly, which is refusing to gamble with something that was handed to you to look after.
Most of the people I meet who are dragging their feet on AI are doing the second one. They are usually right to.
The question worth asking your team isn't whether you trust AI. It is what specifically would have to be true before you did, and whether anyone has ever written that down.
From a conversation with Josh on J&C Unscripted.