I went down a research rabbit hole on where the money actually is for small businesses adopting AI, and the answer surprised me — so I’m sharing it because it runs against what the hype tells you.
The hype says the value is in the smartest model. The data says otherwise. Right now, 57% of organizations report having AI agents running in production. But when you ask what’s blocking them, the top answer — a third of them — is quality: the AI giving accurate, relevant, consistent answers. Second is speed. Cost barely registers anymore, because model prices keep dropping.
Read that again if you’re a small business deciding where to spend. The fanciest model is not your bottleneck. The layer around the model is.
“Infrastructure” is an intimidating word, so let me translate it into the three things it actually means for a business:
Orchestration — does the work survive a hiccup? An agent that silently dies on step 3 of 5 isn’t a minor bug. It’s the moment your team stops trusting it and goes back to doing it by hand.
Tool access — can the AI actually do anything? An AI that can only talk is a chatbot. An AI connected to your files, your database, your email is a tool. There’s now a common standard for these connections (it’s called MCP) so you’re not locked into one vendor’s way of doing it.
Observability — can you see what it’s doing and what it’s costing? Most teams can see token usage; barely half measure whether the output is any good. That gap is how a cheap pilot becomes an $18,000 surprise.
The cost math makes this concrete. A smart setup routes easy questions to a cheap model and only escalates to the expensive one when confidence drops. Done right, the cost per task stops being an interesting comparison to human labor and starts being an obvious one.
The practical takeaway: this isn’t a six-month transformation project. The realistic path is about 30 days — pick one high-volume, repetitive task, connect the tools, add visibility, route by confidence, run it in parallel with the human for a while, then switch. The businesses that fail at this don’t fail because they picked the wrong model. They fail because nobody built the layer around it.