The AI automation pitch for SMBs has gotten loud. Every SaaS tool has an 'AI' badge now, every consultant is promising productivity gains, and every founder feels pressure to have an AI strategy.
Here's the quieter truth: most SMBs don't need an AI strategy. They need two or three specific automation wins that save real hours and are cheap enough to pay back in under six months. Everything else is distraction.
How to find the right workflows
Start with your team's frustrations, not vendor demos. The workflows worth automating are the ones people dread — the repetitive, predictable, high-volume tasks that eat the same kind of time every day or every week.
Ask your team: what do you do on repeat that feels like it shouldn't require a person? The answers to that question are almost always where AI automation creates real value. They're rarely where the demos are.
Good candidates for SMB AI automation: first-line customer enquiry triage, invoice and document data extraction, meeting notes and action item summaries, social media content drafts from a brief, first-draft email responses based on context.
Poor candidates: anything requiring human judgement, relationship nuance, or accountability. AI should not be the one deciding whether to give a refund to a dissatisfied customer or handling a complaint that could escalate.
The tools that actually work
For most SMBs, the right AI stack is embarrassingly simple: an LLM API (OpenAI or Anthropic), Zapier or Make for integration, and a document processing tool like Reducto or Textract for anything involving PDFs or forms.
You don't need to build anything custom for the first two or three workflows. The majority of SMB AI automation wins can be implemented in a weekend using no-code tools.
Custom development makes sense when: the workflow is core to your competitive advantage, the volume is high enough to justify the ongoing maintenance cost, or the off-the-shelf tools can't handle the specific logic your process requires.
The ROI calculation
Before building anything, estimate the cost of the current workflow. How many hours per week? What's the loaded hourly cost of the person doing it? What's the error rate and what does an error cost?
Then estimate the automation cost: implementation time, ongoing API costs, and a maintenance factor (add 30% to whatever you think maintenance will cost — it's always higher).
If the payback period is under six months, it's probably worth doing. If it's over 12 months, you're likely solving the wrong problem or solving it in the wrong way.
What to ignore
AI-generated content at scale. The economics look great — replace expensive writers with cheap AI output — until you look at the quality, the SEO consequences, and the fact that your customers can tell. A few well-written pieces outperform a flood of mediocre AI content every time.
AI customer service as a cost-cutting measure. Reducing headcount by replacing human support with a chatbot is the kind of decision that generates short-term savings and long-term brand damage. Use AI to support your support team, not to replace them.
Anything that requires the AI to make decisions with real consequences without human review. AI is good at drafting, summarising, triaging, and generating options. Humans should still be making the calls that matter.
A practical starting point
Pick one workflow. The one with the highest weekly cost in human hours. Spend two weeks automating it with the simplest possible tool. Measure the result. Then decide whether to do a second one.
That's the whole strategy. Anyone selling you something more complicated than that is probably selling you the complexity, not the result.