Practical AI for small businesses: what it's good at, where it goes wrong, and how to start
Setting aside the hype and the fear, AI tools can save a small business real time on specific jobs, and cause real problems on others. Here's a plain guide to both, and a careful way to begin.
Depending on who you listen to, artificial intelligence is either about to run your whole business or about to ruin it. Neither is a useful place to start.
A more practical view: today’s AI tools are very good at some specific jobs, unreliable at others, and risky when used carelessly. Used well, they can take hours of routine work off a small team each week. Used badly, they can send a customer wrong information with complete confidence.
This guide is about the practical middle. What these tools do well, where they go wrong, how to protect your customers’ information, and how to start in a way that’s safe and actually useful. It’s written for business owners, not engineers.
What we mean by “AI” here
“AI” covers a lot of things. In this guide, we mostly mean the tools that have become widespread in the last few years: large language models, the technology behind assistants like ChatGPT, Claude and Gemini. You type or speak, and they respond in fluent language. Many can also read documents and images, and some can take actions in other software.
They’re built by learning patterns from enormous amounts of text. That explains both their strengths and their weaknesses. They’re remarkably good at producing language that reads well and fits the situation. But producing language that sounds right is not the same as knowing what is true.
Where AI tools genuinely help
These are the kinds of tasks where AI tools tend to be useful in small businesses. What they have in common: a person checks the result, and a mistake is easy to spot and cheap to fix.
Drafting
A first draft of an email, a product description, a job advert, a social media post, a proposal or a policy. The AI gets you from a blank page to something to edit in seconds. You still edit it: for accuracy, for your voice, and to remove the generic phrases these tools are fond of.
Rewriting and adjusting
Making a message shorter, clearer, friendlier or more formal. Turning rough notes into a tidy summary. Translating a message into another language (with a fluent speaker checking anything that matters).
Summarising
Condensing a long document, a meeting transcript or a thread of emails into the key points and actions. Useful, but check anything you’re going to rely on, because summaries can leave out or misstate details.
Pulling information out of documents
Reading invoices, receipts, forms or delivery notes and extracting the details into a structured list, ready to check and import. This can save a lot of typing, as long as someone checks the results, especially the numbers.
Sorting and routing
Reading incoming enquiries and suggesting a category (“new quote”, “complaint”, “delivery question”) and who should handle them. A person can easily correct a wrong suggestion, and the time saved adds up.
Answering questions from your own information
An assistant that answers questions using your own documents, such as your price list, policies or product information, rather than general knowledge. This is more reliable than a general chatbot, because it’s working from your material, but it still needs testing and limits (more on that below).
Thinking aloud
Talking through a problem, getting a list of things to consider, or having your plan questioned. The AI isn’t an expert, but it can be a useful sounding board, especially for spotting things you haven’t thought of.
Getting better results
How you ask makes a big difference to what you get. A few habits help:
Give it context. Tell it who you are, who the text is for and what it’s for. “Write a reply to a customer” gets a generic answer. “I run a small bakery. A customer ordered a birthday cake for Saturday and wants to change the flavour. We can do it if they confirm by Thursday. Write a short, friendly WhatsApp reply” gets something you can nearly send.
Give it your material. Paste in the facts it should use, such as your price list, the policy or the order details (with personal details removed), and ask it to use only that.
Show an example. If you have a past message or product description you like, include it and ask for the same style.
Say what you don’t want. “No exclamation marks”, “under 80 words”, “don’t mention discounts”.
Ask it to flag uncertainty. “If anything isn’t covered by the information I’ve given you, say so instead of guessing.” It won’t catch every mistake, but it helps.
Keep what works. When you find instructions that give good results for a regular task, save them and share them with your team, so everyone gets consistent results.
Where it goes wrong
It can be confidently wrong
This is the most important thing to understand. Language models can produce statements that are false, including invented facts, figures, quotations, references and policies, written in exactly the same confident tone as true ones. This is often called “hallucination”. It isn’t a rare glitch that has been fixed; it’s a consequence of how the technology works. It can be reduced, but not eliminated.
In practice, that means:
- Never publish or send AI-written facts without checking them. Especially numbers, prices, dates, legal points, health information and anything about a real person or business.
- Don’t let an AI make promises to customers on your behalf unless it’s limited to information you’ve given it and tested.
- Be especially careful with anything that sounds authoritative, like a citation, a regulation or a statistic. Check that it exists.
It doesn’t know your business
A general AI tool doesn’t know your prices, your policies, your stock or your customers, unless you give it that information. Asked about them, it may guess, and the guess may sound entirely reasonable.
It sounds like everyone else
AI-written text has recognisable habits: certain words and phrases, a smooth but empty tone, and lists of three. Customers increasingly notice it. If everything you write sounds generated, it can make your business feel less personal and less trustworthy. Use the draft as a starting point and make it sound like you.
It can reflect bias
Because these tools learn from human writing, they can reflect human biases. That matters if you use them for anything involving decisions about people, such as screening job applicants. For those uses, be very cautious, keep a person making the decision, and check what the law says where you operate.
Automating too much
The temptation is to connect an AI to everything and let it run. The risk is that mistakes happen at scale, without anyone noticing. A person replying to a customer might make one mistake; an automated system making the same mistake might send it to every customer who asks that question.
Protecting your customers’ information
Before you paste anything into an AI tool, think about where it goes.
Read the terms. AI providers have different policies on how they use what you send them, including whether it may be used to train future models. Consumer and business versions of the same tool can have different terms. Business plans often promise not to train on your data, but check the actual terms for the plan you use, not what you assume.
Don’t paste sensitive information into tools you haven’t checked. Customer details, staff records, health information, financial records and anything covered by a confidentiality agreement.
Remove details you don’t need. Often the AI doesn’t need the customer’s name or phone number to help with the task. Take them out.
Know your obligations. If you hold personal information about customers, data protection law where you operate probably says something about how it can be shared and processed. Using an AI service can count as sharing it with a third party.
Make a simple policy. Even a one-page note for your team helps: which tools are approved, what can and can’t be put into them, and that AI-written content is always checked before it’s sent.
Customer-facing assistants: do it carefully
An AI assistant answering customers, on your website or on WhatsApp, can be genuinely useful: instant answers at any hour to common questions. It’s also where the risks are most visible. If you go down this road:
Give it your information, and limit it to that. The assistant should answer from your own approved material (your prices, policies, hours, product details), and say “I don’t know” or hand over to a person when the answer isn’t there.
Decide what it must never do. For example: never promise refunds or discounts, never give medical or legal advice, never make up an order status.
Always offer a person. Every conversation needs a clear, easy way to reach a human, and the handover should pass on the conversation so the customer doesn’t have to repeat themselves.
Be open about it. Tell customers they’re talking to an automated assistant.
Test it properly before launch. Collect real questions customers have asked, including awkward and unusual ones, and check the answers. Try to trick it. Keep that list and re-run it whenever you change the assistant.
Keep watching after launch. Read a sample of conversations regularly. Look for wrong answers, frustrated customers and questions it can’t handle.
Know the running cost. AI services usually charge by use. A busy assistant can cost more than expected, so set limits and monitor spending.
Building AI into your own systems
Beyond off-the-shelf tools, AI can be built into your own website, app or internal systems. For example, reading uploaded documents, suggesting replies to your team, or searching your records in plain language.
Done well, this follows the same principles as everything above, with a few more:
- Start from a specific job, not from “we should use AI”. What task takes too long, and how will you know it’s better?
- Measure quality before and after. Build a set of real examples with known correct answers, and check the system against them before launch and after every change.
- Design for mistakes. Show people what the AI suggested and let them correct it easily. Keep a record of what it did.
- Have a fallback. If the AI service is slow, down or unsure, the system should still work, just without the AI help.
- Avoid lock-in. AI models and providers change quickly. A system designed so that the model can be swapped is easier to keep current and to keep affordable.
How to start
The best way to start is small, specific and measured.
- Pick one routine task that takes someone real time each week, where a person can check the result. Replying to common enquiries, drafting product descriptions, or turning meeting notes into action lists are good candidates.
- Choose an approved tool, having checked its terms.
- Try it for a few weeks, with the person who does the task now.
- Keep a person checking every result.
- Notice what changes: time saved, mistakes caught, and whether the result is actually as good.
- Decide honestly. Keep it, adjust it, or drop it. Not every task benefits.
Then pick the next task. Over time, you’ll build a clear picture of where AI helps your business and where it doesn’t. That’s worth more than any grand plan.
Questions to ask anyone selling you an AI solution
- What specific task does it do, and how will we measure whether it’s working?
- Where does our data go, and is it used to train anything?
- What happens when it’s wrong? How will we know?
- What stops it from making things up or promising things we don’t offer?
- How do customers reach a person?
- What does it cost to run each month as usage grows?
- If the AI service changes its prices or shuts down, what happens to our system?
If the answers are mostly about how impressive the technology is and not about these questions, be cautious.
How we approach it
At Variance, we use AI where it genuinely helps and say so plainly when it doesn’t. When we build it into a client’s website or systems, we start from one specific job, keep a person in control of anything that matters, test against real examples, and design it so the rest of the system keeps working without it.
If you’re wondering whether AI could save your team time, tell us about the task. We’ll give you a straight answer, including if the answer is “not yet”.