You are going to get pitched a lot this year. Some of it will be good. Most of it will be a wrapper around the same underlying technology, sold at a markup, to someone who cannot yet tell the difference.
You do not need to understand how any of it works to evaluate it. You need six questions.
1. What happens the first time it is wrong?
Not is it accurate. Everything is accurate in the demo. Ask what the failure looks like.
There are two kinds. Loud failures stop and tell you something went wrong. Quiet failures produce a confident, plausible, incorrect result and hand it to a customer.
Quiet failures are the dangerous ones, and they are the default behavior of this technology unless someone deliberately designed against it.
So ask: when this is wrong, who finds out, and how fast? If the honest answer is the customer finds out, you want a human checkpoint before it goes out the door. That is not a flaw. That is just what this thing is, and anyone who tells you otherwise is selling.
2. Can I see it work on my stuff?
Not a demo. Not a sandbox with clean sample data. Your customer list, your invoices, your messy job history with the duplicate entries and the guy who typed everything in caps for two years.
Every one of these tools looks brilliant on prepared data. The entire question is what it does with yours.
If a vendor will not run it on a slice of your real data before you commit, there is a reason, and the reason is usually that setup is much harder than the pitch implied.
3. What does it need from me every week?
Every tool has an ongoing input cost. Someone has to check it, correct it, feed it, or approve things.
Get that number before you buy, in hours, and be suspicious of zero. Zero means either nobody has run this in production or nobody is checking whether it still works.
Then compare it honestly against the hours you are supposedly saving. A tool that saves five hours and costs two hours of supervision saves three hours. That can still be a good deal. It is just a different deal than the one on the website.
4. If I stop paying, what do I keep?
This one separates a tool from a dependency.
If you cancel, do you keep your data in a form you can actually use? Your customer records, your history, the things it wrote? Or does it all live inside their system and leave with them?
The answer determines your negotiating position at every renewal for the rest of the relationship. Ask it while you are still a prospect, which is the only time you will get a straight answer.
5. Who is accountable when it matters?
There is a difference between software you rent and a person who is responsible.
Software has a support email and a status page. When something breaks at 4pm on a Friday and a customer is angry, a status page is not much company.
Neither model is wrong. But know which one you are buying, and price accordingly. A cheap tool with no accountability is genuinely cheaper, and it is also genuinely on you.
6. What is the smallest version of this I could try first?
The best answer to almost every AI pitch is let me do the tiny version by hand for two weeks.
If a tool claims it will save you hours chasing invoices, spend two weeks chasing them manually on a schedule. You will learn three things: whether the problem is as big as you think, what the work actually involves, and whether the promised output is the thing you needed.
Most purchases do not survive this. That is the point. The ones that do, you buy with confidence and a baseline number, which means you will actually be able to tell whether it worked.
The tell
Here is the shortcut, if you only remember one thing.
Ask the vendor what their tool is bad at.
Someone who has actually run this in real businesses will have an immediate, specific, slightly annoyed answer. They have been burned. They know the edges. They will tell you which situations it handles poorly, because they would rather tell you now than deal with you in month three.
Someone who is reselling a wrapper will say it depends on your use case, or pivot to a feature, or tell you it keeps getting better.
That question has cost us business and it has saved clients considerably more. It is the fastest way to find out whether the person across the table has built anything.
None of these six questions require you to know anything about how AI works. They are the same questions you would ask about hiring a subcontractor: what happens when it goes wrong, can I see your work, what do you need from me, what do I keep, who is responsible, and can we start small.
You already know how to make this decision. It is just wearing an unfamiliar costume.