This article sits alongside the Stage 1 slide deck. If you were in the room for one of the live sessions, or you’ve been handed the slides without the two hours of talking that went with them, this fills in the reasoning behind each section. If you weren’t in the room at all, it stands on its own.
Start with realistic expectations, not the tool
The pitch you’ll see online is always some version of “you’ll never need to do this again.” Outsource it, automate it, let AI handle it. Anyone telling you that your tools are going to do everything for you is making themselves rich, not you. That’s worth saying plainly before anything else, because it sets the frame for everything that follows: AI changes how you do things. It doesn’t remove the need to think about them.
The more useful starting question isn’t which tool to pick. It’s what your business actually looks like right now. A business with documented processes, clear procedures and a defined way of communicating can plug AI in and feel the benefit almost immediately. A business where most of the operational knowledge lives in one person’s head, which describes a lot of small and solo operations, has nothing to plug AI into yet. That’s not a criticism. It’s just the honest starting point, and it’s also, incidentally, one of the more secure states a business can be in: nobody can hack what was never written down.
So the first piece of foundation work isn’t a tool decision. It’s an audit. What actually happens in your business, in order, on a normal day. Write it down. Only once that exists does a tool decision mean anything.
Someone needs to own this
Most businesses already have “the guy” for hardware and software, and another for data and cloud. AI tools and the framework around them need the same thing: someone whose job it is to pay attention to what’s changing, what’s connected to what, and why. It can be you, wearing a different hat. It doesn’t need to be a hire. But it needs to be a named responsibility, not something everyone assumes someone else is watching.
This matters more than it sounds like it should, because AI tools don’t arrive with edges. They connect to email, to calendars, to customer platforms, to each other. Without one person (or a clear, written agreement between the people involved) deciding how instructions get updated and who’s allowed to change them, you end up with colleagues quietly overriding each other’s settings. That’s also, not coincidentally, when AI output starts drifting into territory nobody asked for.
The tool landscape, and why picking one is the wrong first move
Claude, ChatGPT, Gemini, Microsoft Copilot: none of them is the single right answer, and anyone claiming otherwise is selling something. What matters more than which one you choose is that you plan for change. The tool that’s best in class today may not be in twelve months, and businesses that have engineered flexibility into how they work (a single source of truth for their information, separate from any one AI platform) can move to a better tool with minimal disruption when one appears. Businesses that have built their entire operation inside one platform’s walls can’t.
A single source of truth just means: your information, your training materials, your skills and instructions, live somewhere that isn’t only inside one AI tool. That might be Notion, Google Drive, a shared file system, or something simpler. The point isn’t the platform. It’s that if a tool disappeared tomorrow, your business knowledge wouldn’t disappear with it.
On the specific tools: start with what you already have access to. If you’re on Google Workspace, you already have Gemini included. If you’re on Microsoft 365, you already have Copilot. If neither, Claude and ChatGPT both have workable free tiers to start with, though a paid tier is worth it quickly once you’re using it for real work, since free tiers train on your input and cap you fast. For anything genuinely sensitive, where the information can’t leave your own machine, a locally-run a locally-run tool such as Ollama is worth knowing exists. It costs nothing to run, works offline, and nothing you put into it touches the internet. It’s less polished than the big platforms, but for certain kinds of business, that trade-off is the right one.
On tokens, briefly, because almost everyone asks: tokens are the currency your subscription buys. Every request uses some amount, and how much depends on the platform, the model, and how the request is phrased. Broad, vague instructions burn through tokens fast, because the tool has to do more work to figure out what you actually want. Specific instructions are cheaper and get you a better result in one pass. It’s entirely possible to burn through a large top-up in one afternoon by giving a tool too much room to run with an underspecified request. Ask it to check in with you before it executes anything substantial, and you keep control of both the cost and the outcome.
Prompting is a skill, not a search bar
Everyone knows roughly what prompting is. Fewer people get much further than “I typed something, it gave me something, I tweaked it until it sounded more like me.” That works for small tasks, but it plateaus quickly. The output only ever matches the quality of the input.
The most useful comparison here is instructing a new team member, or explaining something to a small child. Telling a child “don’t jump on the sofa” only tells them what not to do. They’ll happily run across it instead, technically compliant. “Put your bottom on the sofa and your feet on the floor” tells them what good actually looks like. AI works the same way. Define the role you want it to take, the information it has to work with, the tools or sources it should draw on, and the actual goal. That’s the difference between a mediocre first draft and something close to usable.
Two habits are worth building early:
Tell it not to execute until you say go. For anything that will take real time or resources to produce, get the tool to ask clarifying questions first, then wait for explicit confirmation. Skipping this step is exactly how you end up with three overcooked drafts of something you needed a simple version of, and a token bill to match.
Know your four common exceptions, because AI defaults toward them unless told otherwise:
- It sounds like AI. Give it a humaniser, a set of instructions built from your own writing, your terminology, and the phrases you never want to see. Claude and similar tools let you build this once as a reusable skill rather than repeating the instruction every time.
- Format death. Left to its own devices, AI defaults to bullet points and headers whether or not the task calls for them. Say explicitly when you don’t want that.
- Boundaries for tone. Different audiences, different brands, different clients need different registers. This is worth setting per task or per client account, not left to guesswork.
- Forcing a creative breakthrough. This one is harder to engineer. Real creative leaps tend to happen by accident, in the moment something gets dragged to the wrong place and looks better there. AI, used well, can open a door to a different angle. It rarely replaces the accident itself.
None of this needs to be rebuilt every time. Once it’s set up as a saved skill or a standing instruction, it runs in the background of every future request.
AI will agree with you. Build in a check for that.
This is worth taking seriously rather than treating as a quirk. AI models are trained on human feedback, and people respond well to validation, so the models learn to give it, especially on decisions that matter. It isn’t a flaw in one tool over another. It’s close to universal.
The practical fix is simple: ask it to argue against you. Give it a plan you’re attached to and say, directly, “challenge this, tell me what’s wrong with it, what am I missing.” The pattern that holds up across every use case, from a quick social caption to a full proposal, is: you direct, AI drafts, you add the specifics only you know, AI scales that across the rest of the task, and you review before anything goes out. The review step doesn’t get to be optional just because the first four steps got faster.
What this is actually changing, beyond the obvious
A few shifts are worth naming plainly, because they explain a lot of what feels different about doing business right now, beyond just “there’s a new tool.”
Search is becoming answering. People increasingly don’t click through ten blue links, they read the AI-generated summary and stop there. That’s reshaping how content gets valued: a well-optimised blog post used to earn traffic and attention directly. Now it earns a mention inside someone else’s summary, or it earns nothing. The practical response is to write content structured around the actual questions people ask, phrased as questions in your headings, not just as topic statements. That’s the shift from search engine optimisation to answer engine optimisation, and it’s worth checking: if someone asked an AI tool to recommend a business like yours, in your town, right now, would it mention you at all?
Authenticity is becoming more valuable, not less. As AI-generated content becomes cheap and common, the imperfect, obviously human version starts to stand out and perform better, not worse. Over-polished AI content that isn’t pretending to be anything else can work fine. AI content trying to pass as real and not quite managing it actively damages trust. The safer failure mode, if you’re unsure, is the unfiltered photo taken on a phone, not the synthetic image that’s almost, but not quite, convincing.
Small no longer looks small. A tiny, regionally based provider can now put together a proposal, a set of visuals, or a communications package that looks as considered as one from a much larger competitor. That’s a genuine opportunity. It also means the basis for a buying decision shifts back toward things that can’t be faked at scale: reputation, word of mouth, verifiable track record. The paperwork stopped being the differentiator some time ago.
Governance isn’t optional once more than one person is involved
If you’re the only person in your business using AI tools, you still need a policy. Write it on a post-it note if that’s what it takes, but write it down. The moment a second person is using AI tools connected to your business, informal understanding isn’t enough, because instructions get overridden, sensitive information ends up in the wrong tier of tool, and nobody notices until something’s already gone out the door.
A workable starting policy is short: free tools for general tasks only. Client or customer specifics never go into a free tier, because free tools train on what you give them and the data-handling terms are different from paid accounts. And a human reviews everything before it’s sent, without exception, regardless of how good the draft looked.
If you want a formal reference point beyond your own post-it note, the National AI Centre’s Guidance for AI Adoption sets out six practices worth checking your own habits against: deciding who’s accountable, understanding likely impacts before you roll something out, measuring and managing risk on an ongoing basis, being transparent about where AI is being used, testing and monitoring outputs rather than assuming they stay reliable, and keeping a human in control throughout. None of it requires a compliance department. It requires someone being able to answer, in one sentence, who checks the output before it goes out.
Where to actually start
Not with a tool. With an audit.
- Audit before you buy. Write down what actually happens in your business, in order. Where does the friction sit. That’s the only reliable place to start, because it tells you what problem you’re solving rather than which tool sounds impressive.
- Implement one thing, small. Not a full system. One task, run properly, tested, and refined. Businesses that try to build something enormous on day one (a full custom process, a from-scratch tool) tend to spend far more in tokens and time than the off-the-shelf option would have cost, and end up with something nobody’s checked for security.
- Train it properly, on its own time. Don’t try to build a skill while you’re also trying to get the underlying task done. The two compete for the same attention, and the result is usually a rushed instruction set that doesn’t hold up, followed by more time spent fixing it than the manual version would have taken.
- Govern it, from day one, not eventually. Decide who checks outputs, how often, and what happens when something’s wrong. This step runs alongside the other three permanently. It doesn’t switch off once the tool is “working.”
Then go back to step one for the next task. That loop, done consistently, is what separates a business getting genuine value from AI from one accumulating subscriptions and hoping.
The one instruction worth setting up before anything else
If there’s a single piece of setup worth doing before any of the above, it’s telling your AI tool, explicitly, what kind of collaborator you want it to be. Something close to: lead with the answer, don’t bury it in caveats or preamble; skip the filler and the false enthusiasm; make the output something that’s actually usable as it stands, not a draft that needs translating into something usable; and flag real risks plainly rather than softening them into uselessness.
That one instruction, saved and reused at the start of every session, changes the baseline quality of everything that follows. It’s the smallest possible piece of AI architecture, and it’s also, in practice, the one that pays off fastest.
Related reading
For the compliance side of the same picture, see AI Self-Governance for Small Business: A Plain-English Guide to the Rules Nobody’s Written Down Yet. If your business is past the tool-by-tool stage, see Why AI Architecture Is the Business Decision Most Medium Businesses Haven’t Made Yet.
