The most practical way to use AI in your business is to start with work you already do repeatedly, that follows a pattern, and that doesn't require a judgement call only you can make. Think drafting, summarising, classifying, extracting, researching and formatting. Not strategy. Not creativity. Not anything where being wrong silently would cost you. Get those narrower, well-defined tasks working reliably first, and the broader picture becomes much clearer from there.
Why Most AI Advice Misses the Mark for Small Business
Most guides to using AI in business are written by people who aren't running one. They're written by marketers, journalists, or consultancies with something to sell. The advice tends to be either wildly abstract ("use AI to scale your operations") or weirdly specific to enterprises with IT departments and six-figure tool budgets. Neither is much use to a founder covering multiple roles or an operator trying to reclaim a few hours a week.
The second problem is that the advice gets written around tools rather than problems. "Here are 10 AI tools you should try" is the format. That gets things backwards. Tools are only useful once you've identified a specific, recurring friction in your work. Without that, you're collecting software, not solving anything.
Step 1: Audit What You Actually Do Each Week
Before touching any tool, spend a week noting every task that you either repeat frequently, dread doing, or do just well enough to get by. Be specific. Not "admin" but "writing follow-up emails after client calls" or "summarising contractor proposals before a decision meeting." The more granular the list, the more useful the next step becomes.
- Which tasks do you do more than once a week that follow a similar structure each time?
- Where do you spend time pulling information from one place and reformatting it somewhere else?
- What do you write from scratch that you suspect could be 80% drafted for you?
- Where are you the bottleneck, not because of your judgement, but because of your availability?
A useful filter: if someone new could follow a written set of instructions to do the task, it's a strong candidate for AI. If the task requires implicit knowledge built up over years, it probably isn't — at least not yet.
Step 2: Match Tasks to the Right Type of AI Capability
Not all AI is doing the same thing. The category that's genuinely accessible to small businesses right now is large language models (LLMs): tools like ChatGPT, Claude, and Gemini that can read, write, summarise, classify and reason about text. That's a broad surface area, but it's still a specific one. They're not databases, they're not automating your calendar, and they're not replacing your accountant.
| Task Type | Good Fit for LLMs? | Notes |
|---|---|---|
| Drafting first versions of emails, proposals, briefs | Yes | Give it context and a clear brief; edit the output |
| Summarising long documents or meeting notes | Yes | Works well; paste the text directly into the prompt |
| Classifying inbound enquiries or support tickets | Yes | Useful for routing; needs testing on your real data |
| Extracting structured data from unstructured text | Yes | E.g. pulling line items from a PDF or email |
| Making strategic decisions | No | It can help you think, but it doesn't know your business |
| Replacing a specialist (accountant, lawyer, etc.) | No | Too much risk; use it to prep for those conversations instead |
| Automating multi-step workflows | Partially | Tools like n8n or Make can stitch LLMs into workflows, but needs setup |
Step 3: Run a Real Experiment, Not a Demo
Pick one task from your audit. Take a real example of that task, something you actually need done this week, and try doing it with an LLM. Don't test it on a made-up scenario. Real inputs surface real failure modes: the model hedges when you need directness, or it produces something plausible-sounding but wrong about your industry.
- Write out exactly what you'd tell a capable colleague if you were handing this task to them. That's your starting prompt.
- Include relevant context: who the audience is, what the output should look like, what to avoid.
- Run the prompt. Read the output critically, not hopefully.
- Note specifically where it falls short, then adjust the prompt and try again.
- After three or four iterations, judge whether the output is genuinely saving you time or just creating editing work.
If after a few iterations the output still needs substantial rewriting, that task may not be the right starting point. Move on to another candidate from your audit rather than forcing it.
Step 4: Build a Small Prompt Library, Not a Tool Stack
The temptation once something works is to go and sign up for five more tools. Resist it. What actually compounds in value is a small collection of prompts that work reliably for your specific business, your tone, your clients, your context. A well-crafted prompt for your most common use case is worth more than a dozen subscriptions you haven't figured out yet.
Keep these prompts somewhere obvious: a Notion doc, a plain text file, a pinned message in Slack. The format should be: what the task is, the prompt itself, any example inputs and outputs, and a note on where it tends to go wrong. That last bit is the most underrated part. Knowing the failure modes of a prompt is what lets you use it confidently at speed.
Step 5: Know What Not to Automate
There's a category of work where using AI is genuinely counterproductive, not because the model can't produce something, but because the act of doing the work yourself is the point. Writing a proposal where the thinking process surfaces what you actually believe about a client's problem. Having a difficult conversation rather than generating a script for it. Making a call on a hire where your read of someone matters more than any summary.
AI doesn't have skin in the game. You do. The work that carries your judgement, your relationships, and your accountability is the work worth protecting from automation, at least at this stage. The goal is to shed the work that doesn't require you, so you can do more of the work that does.
A Realistic Picture of What to Expect
Many founders and operators who use AI practically report reclaiming meaningful hours each week on specific, well-scoped tasks. What almost nobody reports is a transformation of their whole operation overnight. The honest shape of it is: a few tasks get meaningfully faster, a few don't work as expected, you learn something about your own processes in the process, and you build from there. That's not a small thing. It compounds. But it doesn't happen by downloading apps.
Watch out for "AI washing" in tools you already pay for. Many SaaS products are adding AI features that are thin wrappers around existing models. Before paying a premium for an AI add-on, check whether you could get the same result by pasting the relevant text directly into ChatGPT or Claude.
Practical AI for Founders: Where to Actually Start
If you want a concrete starting point rather than another framework, here are the use cases that consistently prove their value for lean teams and solo operators, based on what shows up again and again in practice:
- Turning rough meeting notes into structured summaries or action lists
- Drafting outreach emails with a clear brief and a specific angle
- Researching an unfamiliar topic quickly to get to the right questions faster
- Rewriting something you've drafted until it's the right length and tone for the audience
- Generating a first pass of a FAQ, brief, or spec document from bullet points
- Classifying and triaging inbound messages when volume spikes
None of these require a new tool. They all work in a standard LLM interface today. Start there, prove value on something real, then decide whether a more integrated setup is worth the overhead.
Do I need to pay for a premium AI tool to get real value?
Not necessarily. The free tiers of tools like ChatGPT and Claude are capable enough to prove whether a use case works before you commit to a paid plan. The paid versions are worth it if you're hitting usage limits on tasks you've already validated. Don't pay first and figure out the use case later.
Is it safe to put business information into AI tools?
It depends on the tool and the data. Most major LLM providers offer settings to opt out of using your inputs for model training, but you should check the terms for the specific tool. For sensitive client data, financial records, or anything covered by a confidentiality agreement, either anonymise the information before using it in a prompt, or use an enterprise plan with appropriate data handling terms. The default consumer interfaces are not the right place for genuinely confidential information.
How long does it realistically take to see results from using AI in a small business?
For a single well-chosen use case, most people find a working prompt within a few hours of serious experimentation. Broader, consistent time savings typically take a few weeks to establish as you build up a reliable prompt library and integrate new habits into your actual workflow. If you're not seeing any value within a fortnight of genuine daily use, the issue is usually the choice of task rather than the technology.
What's the difference between using a general LLM and a specialist AI tool for my industry?
Specialist tools often add a curated interface, pre-built prompts, or integrations with other software in your stack. They can be worth it if the interface genuinely saves you setup time and the pricing is fair. The underlying model is usually the same or similar to what you'd get directly. The real question is whether you're paying for genuine convenience or just branding. For most early-stage use cases, starting with a general-purpose LLM directly gives you more flexibility to learn what actually works.
Should I be building AI into my products, or just using it in my operations?
These are separate questions with different risk profiles. Using AI in your internal operations is low-stakes and reversible: if a prompt produces bad output, you catch it before it ships. Building AI into a product that customers rely on requires much more rigour: clear failure modes, fallbacks, and honesty with users about what the system is and isn't doing. Start with internal use to build genuine intuition for how these models behave, then apply that knowledge to product decisions with much more care.