There is a moment, about six weeks into most companies' AI enthusiasm, where the mood turns.
It's not dramatic. Nobody cancels anything. It's just that the owner, who was talking about this constantly in March, has stopped bringing it up by May. And if you ask him why, you get some version of: it's fine, it's just not what I thought.
What he thought was that he was buying agency. What he bought was assistance. Those are different products, they have wildly different value, and almost nobody in this market is drawing the line clearly — because the line is where the sale gets harder.
The two things people mean by "AI"
Assistance is a machine that makes a human faster at something the human is still doing. It drafts the email; you send it. It summarizes the thread; you decide. It suggests the code; you commit it. The human remains in the loop on every single output, and the value is a percentage improvement on their throughput.
Assistance is real, it's useful, and it's worth paying for. It is also, at a twenty-person company, worth a lot less than people expect — because the bottleneck in most small businesses is not how fast individuals type. Making everyone 15% faster at drafting emails does not make the company 15% better. It mostly makes people slightly less tired.
Agency is a machine that completes a unit of work end to end and hands you an outcome, not a draft. The invoice gets coded, entered, and routed without anyone looking at it, except the 6% that got flagged. The lead gets qualified and booked while you sleep. The document gets abstracted and lands in the system with a confidence score. Nobody is in the loop on the happy path.
Agency is where the money is. It's also five times harder to build, and that's the whole reason the market is flooded with assistance products dressed in agency language.
How to tell which one you're being sold
Ask one question: when this runs, does a human have to look at every output before anything happens?
If yes, you're buying assistance. That's fine — just price it as assistance. The return is a productivity percentage, and it's usually modest, and it depends entirely on people changing their habits, which they mostly won't.
If no, you're buying agency. Now ask the follow-up questions that actually matter, and watch how fast the conversation gets vague: What's the accuracy rate on my data, not on your demo data? What happens on the ones it gets wrong — how do I find out, and when? What's the confidence threshold and who sets it? What does the audit trail look like when a regulator or a client asks? Who's on the hook when it acts and it's wrong?
A vendor with a real agency product has crisp answers to all five, because they've had to. A vendor with an assistance product wearing a costume will start talking about the roadmap.
The uncomfortable middle
Most valuable real-world systems are neither. They're agency with a supervised exception path, and this is the design that actually works at SMB scale.
The machine handles 85% of cases end to end, without a human. It routes the other 15% to a person, with everything already gathered, the ambiguity explained, and a recommendation attached. The human spends their time exclusively on the hard cases — which is, not coincidentally, the only part of the work that was ever interesting.
This design has two properties that pure agency doesn't. It's honest about uncertainty, which is what makes people trust it. And it degrades gracefully — when something changes upstream and the machine gets confused, the exception rate climbs and a human notices, rather than the whole thing silently producing garbage for three weeks.
If someone sells you 100% automation of a real-world business process, they are either lying or they haven't run into your exceptions yet. The right question isn't "does it handle everything." It's "what does it do when it can't."
What this means for your expectations
Set them at the workflow level, not the technology level.
Don't ask "what can AI do for my business." That question has no useful answer and it's how you end up with six pilots and no results. Ask instead: for this specific workflow, which happens 300 times a month and costs me $40,000 a year — can a machine complete it end to end, at what accuracy, with what exception rate, and what happens to the exceptions?
That's a question with a real answer. Sometimes the answer is yes, and you should build it. Sometimes the answer is "it can get 70% of the way and a human has to check every one," and then you're buying assistance and the return is thin and you should probably pass.
Knowing the difference is worth more than any tool you'll buy this year. It's the difference between a company that got real leverage out of this technology and a company with a subscription it forgot about.
The machines are extraordinary now. They are not, however, magic — and the ones that pay for themselves are the ones pointed at a workflow you priced first, doing work end to end, and telling you honestly when they're not sure.