Getting Started with AI

Why this matters now, even if you've ignored it so far

You have probably opened ChatGPT, Claude, or Gemini in a browser at some point in the last two years, asked it a question, watched it produce a confident reply, and closed the tab. Maybe you went back a few times. Maybe you used it to draft an email, or to summarise a long PDF you didn't fancy reading. Maybe you decided it was either useful but unreliable, or impressive but irrelevant to the actual running of your business.

That tab-and-close pattern was a sensible response to the first generation of these tools. They were toys with adult-shaped output. The cost of building something real with them was high, the reliability was patchy, and the integration paths to the systems you actually run your business on (your accounting software, your e-commerce backend, your inventory tracker, your email platform) were basically nonexistent.

That stopped being the picture somewhere in the last twelve months. The models got materially better at staying on-task. The cost of running them dropped sharply: as a back-of-envelope from our own Anthropic API and OpenAI API usage between January 2025 and January 2026, the per-token cost of the cheaper tiers we lean on (Haiku 4.5, GPT-5-mini) fell by roughly an order of magnitude for comparable work. Your mileage will vary by workload, but the direction of travel is clear if you check the public pricing pages directly. The connection points between these models and the SaaS tools you already pay for (Shopify, Klaviyo, Notion, Xero, your CRM) became standard. And a meaningful number of SME operators stopped using AI as a "browser tab thing" and started using it as part of how they actually run their day.

The reason this matters specifically for owner-operators of small businesses is that you are the person closest to every load-bearing decision in your company. You read every customer complaint that escalates. You sign off on the wholesale prices. You decide what gets ordered next month. The cost of your attention is the highest cost in the business, and it is the bottleneck on everything that grows. AI is, for the first time, a tool that can take real work off your plate without requiring you to hire someone to manage it. That changes the maths of how a small business operates.

This guide is the first rung on the ladder. We will not show you "20 ChatGPT prompts that will change your business." We will give you a working mental model, the three jobs AI is actually good at right now, the three jobs to keep it well clear of, and a single 30-minute exercise you can run on a real task in your business today. By the end you should know what AI is for, what it isn't for, and what to try next.

What AI actually is, at the practical level

The term "AI" covers a lot of ground. For the purposes of running an SME, the bit that matters is the large language model, or LLM. Everything else you're seeing in 2026 (image generators, voice tools, agents that book your meetings) is either built on top of an LLM or interacting with one. Get the mental model of the LLM right and the rest fits in around it.

An LLM is a piece of software that has been trained on an enormous amount of text. Its job, mechanically, is to predict the next chunk of text given everything that came before. That sounds underwhelming until you realise that "predict the next chunk of text given everything that came before" turns out to cover an extraordinary range of real-world tasks. Drafting an email is predicting the next chunk of text. Summarising a contract is predicting the next chunk of text. Translating French into English is predicting the next chunk of text. Writing code, suggesting product descriptions, classifying a customer enquiry as urgent or routine, all of these are versions of the same underlying mechanism.

The three things that fall out of this mechanism that you need to internalise are:

• It is excellent at fluency. Whatever it produces will read well. The grammar will be right. The tone will be plausible. This is the source of most of its value and most of its risk.
• It has no model of the world separate from text. It knows what people have written about something, not what the something actually is. It can tell you what a wine-tasting note typically sounds like; it has never tasted wine.
• It does not know what it doesn't know. When it has no information, it will produce confident-sounding output anyway. We call this "hallucinating." It is not the model lying. It is the model doing its job (predict the next chunk of text) in a situation where there isn't enough signal to do that job well.

The practical implications of all three are the same: the LLM is a brilliant first-draft producer and a terrible final-authority. When you treat it as the former, you get enormous leverage. When you treat it as the latter, you make expensive mistakes.

The models you'll actually encounter

The market in 2026 has three serious general-purpose LLM providers and a few specialist ones. The serious three are Claude (Anthropic), ChatGPT (OpenAI), and Gemini (Google). As of writing the flagship models are Sonnet 4.6 and Opus 4.7 (Anthropic), GPT-5 (OpenAI), and Gemini 2.5 Pro (Google), with cheaper sibling tiers (Haiku 4.5, GPT-5-mini, Gemini 2.5 Flash) that are good enough for most SME workloads. There are open-weights options (Llama, Mistral, Qwen) which matter if you're building infrastructure, but for an owner-operator getting started, you'll be interacting with one of the three above through a browser tab or a paid tier (Claude Pro, ChatGPT Plus, or Gemini Advanced).

All three are good. They have small differences in personality, in what they refuse to do, and in which integrations they ship with. The right starting move is to pick one and use it for a fortnight. The wrong move is to spend two weeks comparing them. The differences between models matter much less than the difference between using one well and not using one at all.

The three jobs AI is good at right now

We have been running these tools in production for our own businesses and our clients for the last two years. The pattern of what works is consistent. There are three job shapes where you can hand the work over with a high success rate, provided you keep a verification step at the end.

1. Summarising messy inputs into clean outputs

This is the job where AI delivers the most immediate value to a small business. You have a long email thread, a meeting transcript, a customer complaint, a wholesale enquiry that bounced between three people, a PDF supplier contract, and you need to know: what's actually going on here, what does the other party want, and what do I need to do.

The LLM is excellent at this. You paste in the messy input, ask for a structured summary (decision points, action items, open questions, dates), and you get a result you can act on in 30 seconds instead of 15 minutes. We use this pattern across our businesses for: monthly bookkeeping pre-reads, supplier email triage, customer-service escalations, internal meeting notes, and reading the actually-important parts of long legal documents.

2. Drafting first-pass content from a brief

The second job AI is genuinely good at is producing a first draft of something you'd otherwise have to start from a blank page. Product descriptions, email replies, press release announcements, social copy, internal SOPs, training documents, FAQ answers. The pattern is: you give the model the brief (audience, tone, key points, constraints), the model produces a draft, you edit it into the final version.

The reason this works is that first drafts are mechanically the hardest part of writing for most people, and they are mechanically the easiest part for an LLM. The model has read several million product descriptions; producing one more is trivial. Editing one is where your judgement and your brand voice live. Trying to skip the editing step is where SME founders get into trouble.

We do this every day. The draft is the AI's job. The voice, the precision, the choice of what to keep and what to cut, is ours.

3. Extracting structure from unstructured material

The third job is the quiet, unglamorous one that delivers a surprising amount of value over time: taking a pile of unstructured text and turning it into structured data you can actually work with.

You have 200 customer reviews on Shopify. You want to know the top five complaints. You have a year of support emails. You want to know which product issues come up most often. You have a stack of supplier quotes in different formats. You want them in a single spreadsheet with the same columns. You have a transcript of every sales call from the last quarter. You want to know what objections came up most.

All of these are versions of the same job: read unstructured text, extract specific fields, return a tidy table. The model is excellent at it. The output goes into Notion, a spreadsheet, your CRM, wherever your real work lives. The leverage compounds, because each time you do this you build a clearer picture of your business from data you already had but couldn't see.

The three jobs to avoid right now

The flip side. There are job shapes where the LLM will look like it's doing the work but is actually setting up a problem. These three are the most common ways small businesses lose money or trust trying to use AI.

1. Anything requiring real-time accuracy without a verification step

If the answer to a question depends on the current state of the world (today's exchange rate, your current Shopify stock level, the latest legal position on something, the actual content of an attached invoice), and there is no programmatic check that the model has the right answer, do not let the model be the source of truth. The model will produce a confident answer regardless. It might be right. It might be from 18 months ago. It might be confidently wrong in a way you won't catch.

The pattern that works: the model proposes, a programmatic check confirms, the human approves. The pattern that fails: the model is asked, the answer goes straight into customer-facing output, you find out a week later it was wrong.

2. Anything where being wrong has legal or financial liability

Tax advice. Employment law. Contract clauses. Medical questions. Anything where the cost of "the AI was confidently wrong" is a real liability you would carry personally as the business owner, do not delegate to an AI without a qualified human in the loop. The model will draft something that sounds professional. It will not flag the part it's unsure about. That is the part you can't afford it to get wrong.

This applies in shapes that aren't obvious. Drafting a returns policy. Answering a customer query about allergens. Writing the small print on a wholesale agreement. All of these have legal weight you might not feel as you produce them. The fluency of the model lulls you into trusting the output more than you would if a confident-sounding intern had handed it to you.

3. Anything where small errors compound across many steps

This is the failure mode we see most often in over-ambitious early AI builds. The setup looks elegant: a chain of AI steps where each one feeds into the next. Step one reads a customer email and extracts the request. Step two looks up the customer's order. Step three drafts a reply. Step four sends it. Each step is 95% accurate, which sounds fine. Run the back-of-envelope: if each step really is independently right 95% of the time, four steps in a chain delivers 0.95 × 0.95 × 0.95 × 0.95 ≈ 81% end-to-end. Ten steps in a chain at 95% each delivers roughly 60%. This is a modelled estimate (real per-step error is rarely fully independent, and a well-designed retry can claw some of it back), but the shape holds in practice, and it's the same compounding-error pattern Andrej Karpathy has flagged repeatedly in his writing on LLM agents and chained reasoning.

When you build a workflow that depends on AI getting many things right in a row without a human checkpoint, you are building a system that will be wrong far more often than the per-step accuracy would suggest. The fix is not better models. The fix is fewer steps between human checkpoints. Until you have measured your real per-step accuracy on your real workload, design for human-in-the-loop at every meaningful decision.

The mental model that actually works

The mental model we keep coming back to, after a few years of building these systems, is this: treat the AI as a fast, tireless junior with no judgement and no memory.

Unpack each piece of that:

• Fast. It produces output in seconds. This is its core advantage. The tasks that take you 30 minutes take it about 30 seconds.
• Tireless. It will do the 200th product description with the same energy as the first. It does not get bored, distracted, or resentful.
• Junior. It can do real work but it does not have the seniority to make calls. It does not know what your business is for. It does not know which customer you cannot afford to upset. It does not know that the supplier you're emailing has been with you ten years and gets a softer tone than the one you've been having issues with.
• No judgement. Whatever you ask it to do, it will attempt. There is no internal voice saying "this seems like a bad idea" or "are you sure you want to send this to the customer?" The judgement is yours.
• No memory. Unless you're using a tool that has been specifically wired with memory (and most aren't, by default), each conversation starts from scratch. The model does not remember the brief you gave it yesterday, or that you decided last week not to use the word "premium" any more, or that this customer is the third one this month with the same complaint.

The implication of this mental model is that you remain the senior. You are the editor, the decision-maker, the holder of context, the keeper of brand voice. The AI is doing the work you would otherwise have to do yourself or hire someone for. You are not "using AI." You are managing a fast, tireless junior. That managerial framing is the one that gets you the most out of these tools.

The corollary worth saying out loud: if you treat AI as a senior (a thing you trust to make decisions you would otherwise make), it will let you down in expensive ways. If you treat it as a junior (a thing whose output you check before it leaves your business), it will earn its keep many times over.

The three habits that separate signal from noise

If you take nothing else from this piece, take these three habits:

1. Always give it a brief. A two-sentence brief produces a generic output. A two-paragraph brief that captures the audience, tone, constraints, and what success looks like produces a usable draft. The brief is your job; the draft is the AI's. Skipping the brief is the most common reason people decide "AI doesn't work for my business."
2. Always read what it produced. If the output is going anywhere with consequences (a customer, a supplier, your accounting software, your social channels), a human reads it before it leaves. Not a skim. A read. The fluency of the output makes this feel unnecessary; that's exactly when it's most necessary.
3. Always keep score. Track the cases where it got it right and the cases where it got it wrong. Patterns will emerge. You will learn which categories of work you can hand over with confidence and which you can't. Without this feedback loop, you are guessing.

The 30-minute first-use exercise

The point of this guide is that you do something with it before closing the tab. Here is a concrete exercise you can run in 30 minutes today on a real task in your business.

That's 30 minutes. By the end of it, you'll have done one of three things. You'll have produced something genuinely useful, which gives you the first concrete evidence that this is worth investing more time in. You'll have produced something close to useful but not quite, which tells you the gap is mostly about the brief, the iteration, or the choice of task. Or you'll have produced something genuinely poor, which tells you something specific about either the task shape (some tasks really are wrong for AI) or the brief quality, both of which are fixable.

In any of the three cases you will have learned something specific about your own business that no generic guide can teach you. That's the point.

Three things to try this week

If the 30-minute exercise produced something useful, here are the three follow-on moves that compound the value:

1. Pick one recurring task and template it. If you do something every week (a Monday morning summary of last week's orders, a draft reply to a category of customer enquiry, a weekly social post), build the brief once and reuse it. The brief becomes a template; the model becomes the engine. Five minutes a week saved adds up fast.
2. Try the summarisation pattern on a backlog. Find a pile of unread material (customer reviews you haven't read, a folder of contracts, a year of support emails) and try the extraction job. Ask for the top five themes. Ask for a structured table. The output usually surprises you, because the patterns in the pile aren't the ones you'd have guessed.
3. Pick one task to deliberately NOT use AI for. This is the most useful habit nobody mentions. Look at your week and pick one task where you decide, in advance, that you'll do it without AI: usually the highest-judgement piece (the conversation with a key supplier, the decision about a problem hire, the strategy call). The discipline of choosing what to keep human is how you avoid the failure mode of letting AI seep into work it shouldn't be near.

Where to go next in this library

This piece is the first rung. The Absolution Labs resources library is built as a ladder: each piece assumes the previous and goes one step deeper.

The natural next reads, depending on which direction you want to climb:

• If you want to understand the risks before going further: read Safety & Security Considerations for SMEs. It covers the six surfaces of AI risk every business takes on the moment it wires AI into a real workflow.
• If you want to think about who owns AI in your business: the piece on AI tool drift and the owner's job (in this library) is the right next read. The short version: AI tools drift the moment you stop paying attention to them, and the only durable fix is owning the relationship rather than delegating it.
• If you want to think about specific use cases for FMCG or drinks brands: the case studies in this library walk through real deployments at SME scale, with the numbers attached.

The ladder is deliberately gradual. Two years of the AI conversation has been dominated by hype, by enterprise-scale advice that doesn't translate, and by "10 prompts to transform your business" content that has never been near a real business. The library is the antidote: practitioner-written, FMCG-aware, and aimed at the actual operating reality of running a small business in 2026.