AI Strategy for Small Business: Where to Actually Start

Let me save you some time: an AI strategy is not a glossy PDF that sits in a shared drive. It's not “let's get the team using ChatGPT.” And it's definitely not a list of AI tools you saw recommended on Twitter.

A useful AI strategy answers three questions: Where are our biggest operational bottlenecks? Which one should we automate first? And how do we keep improving from there?

If those three questions aren't answered with specific processes, dollar figures, and timelines, you've got a wish list, not a strategy. I've seen businesses spend $20K on consulting firms to produce a 40-page strategy document. Twelve months later, nothing has been built. The document was the deliverable. That's backwards.

The best AI strategies for small businesses fit on one page. They don't require you to become an AI company. They require you to know your own business well enough to spot where AI will make a measurable difference.

Forget the big-picture stuff for now. I need you to think about what your team physically does every day. Not the strategy. The grunt work.

Who spends two hours every morning copying data from emails into your CRM? Who manually chases up leads that came in yesterday? Who puts together that weekly report by pulling numbers from three different systems and pasting them into a spreadsheet?

Those are your candidates. The processes that eat the most hours across lead handling, client onboarding, document processing, reporting, and customer support.

Pick the three worst offenders. For each one, write down what triggers it, what steps happen, who touches it, how long it takes, and — this is the key one — what goes wrong. The pain points are where AI creates the most value. A process that takes four hours but runs flawlessly might be lower priority than one that takes one hour but produces errors that cost you clients.

Not everything manual can or should be automated. But AI has massively shifted the boundary of what's possible. Two years ago, reading an unstructured email and understanding intent was a human-only task. Today, AI does it reliably.

The best candidates share three traits: repetitive, high-volume, and they involve turning data from one form into another. Reading an email and extracting key details. Reviewing a document and scoring it. Sorting enquiries and routing them to the right person. Generating a report from raw data.

The bad candidates are one-off creative decisions, high-stakes negotiations, and anything where a wrong answer causes serious harm with no human review step. A client once asked me to automate their entire hiring decision process. That's a bad candidate. Automating the resume screening and shortlisting step? That's a great one.

This is where most businesses get it wrong. They want the chatbot. They want the AI assistant. They want the thing they can show off at a networking event. Meanwhile, their admin team is spending 12 hours a week on data entry that an automation could handle in minutes.

The highest-ROI projects are almost always the boring ones. Data entry. Report generation. Lead follow-up. Invoice processing. Document classification. These aren't sexy. They're profitable.

Here's the formula I use: hours per week × people involved × hourly cost = your baseline cost. Compare that against the build and monthly running cost. If it pays for itself within six months and keeps saving after that, you've got a winner.

This is something I feel strongly about. When you invest in AI automation, you need to ask: who owns this when it's done?

We build on open-source tools wherever possible. n8n for workflow orchestration and Supabase for databases. No platform lock-in. You can self-host for data sovereignty. You can switch providers without rebuilding everything from scratch.

That doesn't mean avoiding all third-party services. AI models from Anthropic, Google, and OpenAI are third-party and they're essential. But those are easily swappable. If Claude gets too expensive or GPT releases something better for your use case, you switch the model. Takes an afternoon. But your data, your workflow logic, your integration layer — those are expensive to migrate. Own them.

The question to ask: if this vendor disappeared tomorrow, could I keep running? If the answer is no, you've got a risk that needs addressing.

This is the step that separates the businesses that get lasting value from the ones that build something, celebrate, and then watch it slowly degrade.

AI moves fast. Models improve, costs drop, new capabilities appear quarterly. A system you build today will need updating. That's not a flaw — it's an opportunity, as long as you plan for it.

Build monitoring into every system from day one. Track accuracy, speed, volume, and error rates. I've seen systems fail silently for weeks because nobody was watching the logs. By the time someone noticed, three weeks of leads had been misrouted.

Review your AI systems every quarter. Is there a better model that cuts your API costs in half? Has a process changed enough to need an update? Something that was impossible six months ago might be straightforward now. Last year, reliable document processing required expensive vision models. Now smaller, cheaper models handle it fine.

Work with a partner who stays current. Keeping up with AI developments is a full-time job. A good agency will tell you when to upgrade, when to wait, and when something is hype. That ongoing relationship is where the compounding value comes from.

The businesses that win with AI treat it as a capability they keep developing, not a project that ends. Your first AI system is the starting point. The tenth one is where things get really interesting.