Twenty-seven years ago, I sat in front of a green screen terminal watching a colleague post a journal. Line by line, account code by account code. Last week, I watched an accountant do the same thing in Xero. The screen was prettier. The steps were identical.
I’ve been in the technology industry long enough to watch every major shift play out: mainframes to client-server, the internet blowing open the SaaS era, mobile, social, analytics. Each one was supposed to change everything. And yes, accounting is structured by nature. Double entry bookkeeping is always going to be double entry bookkeeping, and nobody is arguing it shouldn’t be. But the point is this: despite all those technology shifts, the productivity gains for accountants have been incremental, not transformative. You’re still pressing the same number of keys. Still staring at a huge list of transactions. Still clicking through the same workflows. Each upgrade has felt more like an iPhone update than a revolution. A nicer interface, a slightly faster process, but fundamentally the same job done in fundamentally the same way.
That is, until now.
Beyond the hype
We are in the early stages of the next major shift, and this one is different. I’m not talking about LLMs you chat to and get wildly inconsistent answers from, different every time, confidently wrong (I heard someone call them “confident bullshitters” and it made me chuckle). And I’m not talking about deterministic workflow tools with a bit of AI sprinkled on top.
I’m talking about properly grounded AI. AI that can securely access your data, your insights, your knowledge, the things that differentiate your practice from every other one out there. AI that gives you reliable, consistent results every time.
The World Economic Forum’s Future of Jobs Report 2025 ranks accountants and auditors among the fastest-declining roles globally, driven by AI and automation. We hear about this wave hitting white-collar workers harder than ever. And I know what most accountants are still thinking: “We’ve heard it all before. It’s just hype.” I get it. You’ve sat through a decade of vendors promising that automation would transform your practice, and you’re still doing the same things you were doing before, just with a slightly nicer dashboard. But beyond the hype, something genuinely different is happening this time. AI will elevate the expert. If you don’t know how to prepare a set of corporate accounts, if you can’t spot the common errors clients make, if you don’t know which clients are costing your firm money, if you lack the instinct to navigate a tricky tax position or read between the lines of a set of financials, then no amount of AI is going to teach you that. That knowledge comes from real-world experience. What AI will do is let you move faster, strip away a lot of the drudge work, and free you up to apply that expertise where it matters most.
But enough of the big picture. Before we get into the practical stuff, it’s worth understanding why small practices are actually better placed than anyone to make this work.
The small practice advantage
Here’s something that doesn’t get said enough: small and medium accounting practices are in a unique position right now. The barrier to entry has never been lower. You can experiment, iterate, and try new things without the constraints, committees, bureaucracy, and decision-making processes that slow large corporations to a crawl.
The tech industry is moving at breakneck speed and nobody can truly keep up. But you can move faster than most.
One word of caution though: don’t lock yourself into one ecosystem. Saying “I’m all in on vendor X” is the equivalent of visiting your favourite restaurant every day, ordering the same thing forever. Yes, it’s comfortable, but the novelty wears off, you miss out on a whole variety of other amazing options, and before you know it you’re locked into a pricing model that doesn’t scale. And here’s the thing that surprises most people: a lot of what you want to achieve with AI doesn’t require going all-in on Sage or Xero or whoever. You can build your own tools, simply by sitting down with AI and talking through how to approach the problem. You don’t need a six-figure software contract. You need a clear workflow and a conversation.
What to actually do about it
Whether you haven’t touched AI yet or you’ve been dabbling but not seeing the results you expected, these are the things that separate average results from genuinely useful ones:
Get curious. Don’t try an LLM once, ask it a question, get a wrong answer, and dismiss it as too risky. That’s like test-driving a car with the handbrake on and concluding cars don’t work. And just as importantly, don’t assume that because Sage or Xero are adding AI features, your practice is “doing AI.” Auto-labelling transactions is a start, but there is so much more to it. Start looking at your actual workflows, where things are repetitive, where you and your team lose time, and then use AI to figure out solutions.
Ground it. Give AI the context it needs. The old mantra “rubbish in, rubbish out” has never been more true. Tell it what to read, where to source information from, where not to, how you want it formatted, what you want the output to look like. Push back. Don’t believe everything. Tell it not to guess. These simple techniques alone can dramatically improve what you get back.
Tell it who it is. This is huge, and most people skip it entirely. Before you ask AI anything, set the scene. Tell it: “You are an experienced UK accountant working in a small practice. You are preparing work for a client where accuracy has legal and regulatory implications. Do not guess. If you are unsure, say so.” The difference this makes is night and day. Without that context, you get generic, surface-level answers. With it, you get responses that actually sound like they came from someone who understands the profession. Think of it like briefing a new member of staff on their first day. You wouldn’t just hand them a task and walk away. You’d tell them who the client is, what the stakes are, and how your practice likes things done. Same principle applies here.
Ask AI to help you get better at using AI. This sounds circular, but it’s one of the most powerful things you can do. Most people type in a vague question like “can you help me understand the impact of MTD on accountants?” and get a vague answer back. Instead, try this: “I want to use you as an expert advisor on Making Tax Digital. Write me a detailed prompt I can reuse that sets you up as a UK accountant specialising in MTD compliance for sole traders and landlords. Include instructions on tone, accuracy, and where to source information from.” What you get back is a reusable, structured prompt that transforms every future conversation on that topic. You’re not just getting an answer, you’re building a tool. And once you’ve done it for MTD, do it for corporation tax, for client onboarding, for year-end checklists. Each one compounds. Before long, you’ve got a library of prompts that make AI work the way your practice works.
Understand skills. This is one of the most important leaps in how AI works today, and it’s tailor-made for accounting. A skill is a saved set of instructions that configures AI to operate within a specific professional context. Think of it as giving AI a permanent briefing document rather than re-explaining everything from scratch each time. You could build a skill for month-end close that knows your chart of accounts, your review process, and your reporting format. Another for client onboarding that follows your practice’s exact steps. Another for drafting VAT return summaries in the tone and structure your clients expect. The beauty is you don’t need to be technical to create them. You just describe how you work, and AI helps you build the skill. Anthropic has published a guide to skills for financial services that’s worth a read, and there’s a practical walkthrough for accountants on Medium that covers the sceptic’s perspective. These are one of the lowest-cost, most effective ways of grounding AI, and they will change the game.
Tell it exactly what you want back. AI doesn’t know you want a spreadsheet unless you say so. Be specific about the output format. “Give me this as a table with columns for client name, deadline, and status.” “Format this as a letter I can send to the client.” “Summarise this Companies House filing in three bullet points with no jargon.” “Give me this as a Word document with headings for each section of the accounts.” If you’re reviewing a trial balance, tell it you want variances flagged in a table with prior year comparisons. If you’re prepping for a client meeting, tell it you want a one-page summary, not an essay. The more specific you are about the shape of the output, the less time you spend reformatting and the more likely it is to be immediately useful.
Start with something you already hate doing. Every practice has those repetitive tasks that eat hours but don’t require deep expertise. Chasing clients for missing documents, drafting engagement letters, writing up year-end summaries, preparing board minute templates, responding to HMRC correspondence. Pick the one that makes you groan the most and try getting AI to do the first draft. You’ll see results immediately, and it builds the confidence to tackle bigger things.
Not all models are equal (and neither is the cost). Most people don’t realise that AI platforms offer different models at different price points, and choosing the right one matters. Take Claude as an example. It has three tiers: Haiku, Sonnet, and Opus. Think of them like your team. Haiku is your trainee: fast, cheap, brilliant for high-volume simple tasks like categorising transactions, reformatting data, or drafting a quick chaser email. Sonnet is your senior: capable, reliable, handles the bulk of daily work like summarising client documents, drafting correspondence, or working through MTD queries. Opus is your partner: you bring it in for the complex stuff where nuance and accuracy really matter, like reviewing a set of accounts, working through a tricky tax position, or producing advisory work you’d put your name to. You wouldn’t charge a partner out to do bookkeeping, and you shouldn’t use the most expensive model to sort your inbox. Match the model to the task, and you’ll get better results for less money.
How to approach it without getting overwhelmed
Pick a platform and go. Claude is particularly impressive right now, but Google and OpenAI are strong options too. It doesn’t matter which you choose, just pick one. Understand the different models, the different tools available. If you try to get your head around all of them at once, trust me, you won’t. You’ll end up scattered across so many that it gets in the way of progress.
Start with one. Go deep. Build from there.
Follow for more insight on AI in Practice.
Next up: the Claude Excel add-in. This one is a game changer for accountants.