The one-sentence answer
Artificial intelligence is software that has learned patterns from lots of examples, and uses those patterns to make a prediction, a decision or a piece of writing.
Traditional software follows rules a programmer wrote: “if the invoice total is over $5,000, send it to the owner.” AI software is different. Instead of being given every rule, it’s shown a very large number of examples and works out the patterns itself. Show a system enough emails marked “spam” and “not spam” and it learns to sort new ones it has never seen.
The U.S. Small Business Administration puts it simply: AI is “a set of systems that program computers to solve problems or work through tasks,” and those systems “adjust and adapt to new information.” [1]
The three kinds of AI you already meet
When people say “AI” they usually mean one of three things.
1. AI that predicts and sorts. This kind has been around for years, mostly out of sight. The spam filter in your inbox, the fraud check on your card payments and the “customers also bought” suggestions in online stores are all examples. They make one narrow decision, very fast, over and over.
2. Generative AI: assistants that write, summarise and answer. This is the kind that got everyone’s attention. You type a request in plain English and it writes a reply, a summary, a draft email or a first pass at a policy. Chat assistants went mainstream after OpenAI released ChatGPT to the public on November 30, 2022. [2] Many similar tools now exist.
3. AI built into tools you already pay for. Email, accounting, CRM, booking and document tools increasingly include AI features: “summarise this thread”, “draft a reply”, “suggest a category”. Often this is the cheapest place to start, because you’re already paying for the tool.
How a chat assistant works, without the maths
The tools behind chat assistants are called large language models, or LLMs. The SBA describes a language model as one that “uses math to predict the next word in a series of words”, a bigger, much more capable version of the autocomplete on your phone. [1]
An LLM has been trained on an enormous amount of text. From it, the model has learned how words, facts and ideas tend to fit together. When you ask a question, it doesn’t look up an answer in a database. It builds a reply one piece at a time, each time predicting what’s most likely to come next.
That explains both the magic and the problem:
- The magic: because it has seen so much writing, it can draft, rephrase, summarise, translate and explain almost anything in seconds.
- The problem: it’s predicting what sounds right, not checking what is right. When OpenAI introduced ChatGPT, it said plainly that the tool “sometimes writes plausible-sounding but incorrect or nonsensical answers.” [2] The industry calls these made-up answers “hallucinations.”
So a chat assistant is a very fast, very well-read first-draft writer. It doesn’t know your business and doesn’t check its own facts, and it is often confident when it shouldn’t be.
What AI is good at, and what it isn’t
Good at:
- Drafting: replies, quotes, job posts, product descriptions, policies, ready for someone to check.
- Summarising: long email threads, meeting notes, contracts, reports.
- Sorting and routing: reading incoming emails or forms and tagging them by type or urgency.
- Pulling details out of documents: names, dates, amounts and order numbers from emails, PDFs and forms.
- Answering repeat questions, as long as it’s working from information you’ve given it.
Not good at, at least not on its own:
- Facts it hasn’t been given. It doesn’t know your prices, your stock or your customers unless it’s connected to that information.
- Judgment calls. Whether to give a long-time customer a break, or how to handle an upset employee.
- Being accountable. If it’s wrong, someone in your business still owns the outcome.
- Anything it can’t see. A tool that can’t reach your files, inbox or systems can only guess.
Why “AI that knows your business” matters
A general chat assistant knows a lot about the world and nothing about your company. Most of the useful business applications fix that by giving the AI your own information, an approach often called grounding. Your price list, procedures, product sheets, FAQs or past orders become the material it answers from.
Grounded AI is more useful and easier to check. When a customer asks about your return policy, the AI answers from your return policy, and you can see which document it used. It still needs a person in charge of what goes out, but it’s no longer making things up from general knowledge.
This is also where the data questions start: which tool is reading your information, where it’s stored, and who can see it. We cover that in AI and your business data: privacy and safety basics.
A few more words you’ll hear
- Model: the trained AI system itself, the “engine” inside a tool.
- Prompt: what you type or send to the AI. The SBA defines it as “the information you feed into GenAI.” [1]
- Generative AI (GenAI): AI that creates new text, images, audio or video. [1]
- Machine learning: the broader field of building systems that learn from data instead of following hand-written rules. [1]
- Hallucination: a confident answer that’s wrong or made up.
- Grounding: connecting AI to your own documents and data so it answers from them.
- Automation: a workflow that runs on its own, for example “when a form arrives, read it, add it to the spreadsheet, email the right person.” AI can be one step in it.
- Agent: AI that takes a series of steps toward a goal, such as reading a request, looking something up and drafting a reply, rather than answering one question. Useful, but it still needs limits and checks.
Is anyone setting rules for this?
Yes, though most of it is guidance rather than law for small businesses. In the U.S., the National Institute of Standards and Technology (NIST) released its AI Risk Management Framework on January 26, 2023. It’s a voluntary framework for building and using AI responsibly. [3] The SBA’s own advice for small businesses is practical: have a person review what AI produces, keep sensitive information out of tools you don’t control, and consider telling customers how you use AI. [1]
What this means for you
You don’t need to understand the maths to get value from AI, any more than you need to understand payment networks to take card payments. You need three things:
- A clear, repeated task that eats real time: answering the same questions, drafting the same kinds of replies, typing the same details from one place into another.
- The right information in reach, so the AI works from your facts, not guesses.
- A person in charge of what goes out, at least until you trust it on that task.
Start small and pick one job. Measure the before and after. The rest is detail, and we’ve written that part up in How to start with AI without wasting money.