How SMBs actually adopt LLMs (without the hype)
We talk to a lot of small and mid-size companies about AI, and the pattern has stayed the same for a while. The companies that adopt well are not the ones with the biggest models or the biggest budgets. They're the ones that started with a boring question.
The boring question wins
"Which answer do we repeat forty times a week, and why?" That's the whole method. It beats every question about which model is smartest, because it starts from a cost you actually pay: someone's time, spent re-typing the same answer, pulling the same report, hunting the same policy document.
What adoption actually looks like
In our experience, a working adoption has a shape:
- One workflow, end to end. Not "AI across the company", but one concrete job: support tickets, inbound quotes, order status, done better than before.
- The documents first. If the answer should come from a policy, a catalog or a manual, then an answer from a model that hasn't read it is a liability, not a feature.
- A human in the loop, on purpose. Early on, someone reviews the model's output. Not because the model is bad, but because you're learning which outputs to trust.
- A way to say no. Most of the time, the honest answer is "that's not worth it yet", or "a cheap tool already does this". Adoption is as much about not building as about building.
What slows it down
- Buying a platform before naming the workflow.
- Treating prompt quality as a one-day task. It's a maintenance job, like any software.
- Expecting the first version to be safe to ship unattended. It isn't. Nobody's is.
The bar is lower than the hype
The gap between a pilot and a production assistant for an SMB is usually not a research problem. It's an operations problem: where the data lives, who reviews edge cases, what "good enough" means for this workflow. Those are questions we can answer in a week or two of scoping, and they're the reason a working prototype in weeks is realistic, when the alternative is a deck about where the industry is "headed".