By Cat Hall Artificial intelligence How to train your AI 27 Aug 2026 Practical tips on chat bots and automation, and a walk-through on building a Project. We spoke to Rachel Harris FMAAT, Founder, accountant_she and strivex, about the practical ways she uses AI in business. Using chat bots There are really three levels people use AI at, and most people never get past level one. Level one is just using it as a chat tool. You open ChatGPT, Claude, Copilot or Gemini, ask it something, get an answer, close the tab. Fine for quick jobs, but it starts from scratch every single time, so it never learns anything about you or your practice. Level two is where you start “closing the feedback loop”. You give it context about your tone, your clients, your way of working, and correcting it when it gets things wrong, so it improves next time. Level three is building something dedicated. In Claude these are called Projects, in ChatGPT they’re custom GPTs. It’s a separate space with its own memory, its own instructions and its own files, so it stops being a generic chatbot and starts being a proper tool built around one job. That’s the level worth aiming for if you want AI to actually save you time rather than just be a novelty. Task automation Full automation (where it runs on a trigger without you prompting it – say, an email lands and it drafts a reply automatically) usually needs a connector tool like Zapier sitting behind the AI. It’s worth knowing that distinction exists, but I’d always get comfortable at level three first. We use both projects and automated sequences within accountant_she. Case study: Walking through a real build For most accountants, social media falls to the bottom of the pile whenever things get busy. Here’s how I built a project for a practice owner to create compliance-safe social media machine to save time and uncertainty around creating content! We built a dedicated Project inside Claude. Think of it as giving the AI its own dedicated brain for this one job, separate from anything else he uses AI for. Prompting your AI Project Before touching the Project itself, we used an ordinary chat to build the brief. I told Claude what I wanted in plain English: “I run an accountancy firm and want to build a social engine that writes on-brand LinkedIn and Instagram content, with compliance guardrails” and asked it to question me until it had enough to work with. That’s called reverse prompting: rather than trying to write the perfect instruction yourself, you tell the AI the outcome you want and let it ask you the right questions to get there. Claude asked things like: what’s your actual goal here – ready-to-post content, drafts you’ll edit, or just ideas? Which platforms matter? Who’s your audience? What can you never say for compliance reasons? The answers to those became the instructions for the Project. If you’d rather write instructions yourself than be interviewed, there’s a simple four-part structure that works well: tell it who it should be, the situation it’s walking into, what you need, and how you want the answer delivered. Either approach works – use whichever suits how your brain works. Testing your AI Project Before you rely on it, test it on something real. I created this project during a live webinar, so we tested it by running a live prompt through it (for example, write a LinkedIn post for a specific date, in his voice, with a specific goal) and checked the output against what the firm owner actually wanted to say. I’d always recommend running new automations ‘with stabilisers on’. Test manually for a while, side by side with your current process, before you trust it to run unsupervised. Feed back to your AI Project Checking it stays consistent is the bit most people skip, and it’s the difference between AI that gets better over time and AI that stays mediocre. Whatever it gives you, you’ll tweak before you use it… so the job isn’t done until you go back and tell it what you changed: “here’s what you gave me, here’s what I actually posted, please update your instructions and remember this.” That’s the feedback loop. Skip it, and you get the same average output forever. Do it consistently, and it starts sounding more like you every time. Project hygiene The other practical habit worth mentioning is to start a new dedicated space for each distinct job. Upload real material to it (past posts, a note on your tone of voice, examples of things that do and don’t sound like you), and don’t expect one generic chat to hold everything. that’s usually why people end up frustrated that ‘it doesn’t remember’. Cat Hall is Content Specialist at AAT.