Spend an afternoon on LinkedIn and you’ll see a lot of small-business owners proudly posting that they “added AI” to their company. Look closer at what they actually did and the picture darkens fast. They bought a ChatGPT Team subscription. They installed an AI scheduler. They had ChatGPT write their next email newsletter. They are, by their own measure, AI-enabled.
They are also, by every measure that matters, no further along than they were six months ago.
I’ve spent the last year inside the SMB and lower-mid-market AI-adoption conversation, as a builder, as a consultant, and as a long-time product leader who has watched a lot of “transformation” cycles come and go. The same five mistakes keep showing up. The companies making them don’t know they’re making them, because the consultants advising them are usually making the same mistakes one rung up.
Mistake 1: Treating AI as a tool category instead of a capability shift
Most SMBs treat “AI” the way they treated “the cloud” in 2014 a list of products to evaluate, license, and deploy. They buy a CRM with AI features bolted on. They buy an AI receptionist. They buy a “ChatGPT for accounting.” Each purchase is justified separately, integrated badly, and never tied back to a coherent operating model.
The companies getting real lift treat AI as a capability shift instead. They ask: “Where in our business is the bottleneck a human deciding something repeatable, or a human writing something repeatable, or a human routing something repeatable?” Those are the places to apply AI. The vendor list is downstream.
The test: if you can’t describe what got faster, what got cheaper, or what got more accurate after your AI purchase, and back that with a number, you bought a tool category. You did not adopt a capability.
Mistake 2: Letting the implementation lead the strategy
AI vendors are good at sales. They will happily tell you their product is the right solution for any problem you describe, because that is the job description for a salesperson. SMBs without an internal product or strategy function tend to take vendor framing as ground truth, and end up with an architecture defined by which vendors they bought from rather than by what their business actually needs.
The right sequence: define the bottleneck, define the success metric, define the data shape required, then evaluate vendors. The wrong sequence: evaluate vendors, pick a winner, retrofit a problem statement.
I see the wrong sequence five times a week. The right sequence almost never.
Mistake 3: Generic AI for differentiated work
Everyone has access to the same large language models. Your competitors run the same Claude, the same GPT, the same Gemini, against the same general prompts. If your AI usage is “I asked ChatGPT to write our marketing emails,” you have not built a competitive advantage. You have built parity with everyone else who watched the same LinkedIn video you did.
The actual competitive moat is in what you put around the model: your proprietary data, your structured prompts derived from how your specific business operates, your evaluation harness that catches when the model is wrong, and the workflow integration that makes the AI output land in the right human’s queue at the right time. None of that comes in the box.
This is the work most SMBs skip. It’s also the only work that produces durable advantage.
Mistake 4: No evaluation discipline
When you ship a software feature, you test it. You write unit tests, you have QA review it, you run it in staging, you watch error rates after deploy. Most SMB AI adoption skips all of this. The AI receptionist starts taking calls; nobody’s checking what it tells customers. The AI sales-email tool sends thousands of emails; nobody’s reading the outputs to verify they’re not making promises the company can’t keep.
Then a customer screenshots the AI-generated promise and tweets it. Then the AI hallucinates an inventory item that doesn’t exist. Then the company’s brand voice, the voice the founder spent ten years carefully cultivating, gets diluted in the model’s slightly-off paraphrase, and customers notice.
You need an evaluation discipline before you scale AI usage, not after. That means: a sample of AI outputs reviewed weekly by an actual human. A red-team check before any AI talks to a customer in your name. A clear escalation path when the AI gets it wrong, because it will. A measurable bar, accuracy, helpfulness, brand-fit, that you grade against.
This is unglamorous and almost no one does it.
Mistake 5: Hiring an “AI consultant” who has never shipped anything
This one is uncomfortable for me to write, because it sits inside my own profession.
The consulting market is currently flooded with people who have completed two coursera courses and rebranded themselves as AI experts. They speak the vocabulary. They can hold a meeting. They cannot tell you what happens when a Claude API call returns a malformed JSON response in the middle of a multi-step workflow, because they have never had to handle that case. They have never shipped anything that has to keep running while they sleep.
If you are an SMB owner choosing an AI consultant, ask three questions:
- What have you shipped? Not “advised on.” Shipped. With users. Still running.
- What’s the worst production failure you’ve debugged? A real practitioner has scars. A pretender has slogans.
- Show me one prompt you’ve written that’s running in production right now. Not the marketing version. The actual prompt, with the structure, the system message, the temperature setting, the JSON schema for the output. If they can’t show you, they don’t have one.
If those three questions get you slogans instead of artifacts, walk away.
What to do instead
Stop buying AI. Start applying it. Pick one high-leverage workflow in your business, the one where a human is doing repeatable cognitive work you wish you didn’t need them to do, and design an AI-augmented version of it. Build it small. Evaluate it weekly. Decide whether to expand based on data.
Then do it again, somewhere else.
This is not glamorous. It’s not the deck your board wants. But it’s the only path I’ve seen actually produce returns.
If you want help finding the right first workflow and building the evaluation discipline around it, that’s exactly what I do at Local Value Marketing. Two case studies of production AI systems I’ve shipped, for Local.Pet, Local.Dog, Vetsnear.me and LocalValue.co, show what the right approach looks like in practice.
Rob Lewis is the founder of Local Value Marketing and an AI Product Consultant for SMB and mid-market businesses. He has 25+ years of senior product leadership experience and ships production AI applications. Reach him at [email protected].
