What “Serious” Actually Looks Like for AI in Accounting
In this series, I wrote that the exploration phase for AI was over. That statement is both true and false all at once, because although that’s true in the sense of the lifecycle maturity of any new, emergent, disruptive technology, AI has genuinely passed the explorative stage and is well into general early majority adoption. However, not every sector and every type of user has made it this far yet. Whether you’ve had your ChatGPT moment yet, or your CoPilot rollout, or your own curious month playing with the latest Claude or Gemini release, depends very much on your own unique circumstances. That includes both your own and your firms’ willingness and openness to experiment, and it’s worth saying plainly that businesses and departments managing more secure and private data have understandably been more reluctant to move quickly here.
But the question of whether you’re using AI yet stopped telling us anything useful a long while back. What’s still valid, what actually matters now, is not whether your firm is using AI, but how it is intending to using it and for what purpose.
And what I’ve tried to do across this series is answer the questions that actually matter, and the questions your business needs to start answering, to get to one ultimate answer: what does it look like to run AI properly at scale in your firm, rather than just using it?
There are five pillars, which are key to any serious or fledgling AI program in any firm, and I’ll be honest, none of them are as cool and exciting as vibe coding. Nobody ever really gets that energised talking about governance frameworks or agent retirement schedules. And that’s rather the point. The serious phase isn’t glamorous. It’s the unglamorous, structural work that separates those who’ll be confidently using AI in three years from the ones who’ll be trying to figure it out or clean up a mess they can’t quite explain.
The five pillars, put together
Governance is where we start, because none of the rest matters if you can’t answer basic questions about what your AI is allowed to do and who’s accountable when it doesn’t. A policy document isn’t a control. Structural governance, configured into access, data permissions, and a proper AI Development Lifecycle, is.
Literacy makes the case that your best AI talent is probably already on the payroll. The accountant who understands exactly where a client gets nervous, and now understands how, why, and where to use AI, will consistently outperform an AI specialist who’s never opened a set of accounts. Domain knowledge plus AI fluency beats the reverse, almost every time.
Tokenomics deals with the bills that always arrive after the pilot is done. Pilots are cheap by design. Scale isn’t, unless you’ve done the work to understand why, before you’re exposed to it. And “cheaper” is never the right frame for AI in the first place, or your clients will expect same and you won’t have a sensible argument for why not.
Lifecycle & Agent Structure makes the case that agents need managing the same way people do: correct selection, onboarded properly, reviewed on a schedule, and retired deliberately when they’ve stopped earning their place. Ten agents is a spreadsheet. A hundred is a business risk, if nobody’s watching.
Control & Orchestration brings it all together. Nobody has achieved genuine competitive advantage from AI yet, but that advantage is coming, and it’ll belong to whoever can see, govern, and manage everything they’ve built, across the whole organisation, rather than in disconnected pockets.
If you look at those five together, a pattern emerges: every single one of them is about knowing, not guessing. Knowing what your AI can and should access. Knowing what your people can actually do with it and why. Knowing what it costs and what it’s worth. Knowing what exists and whether it should still exist. Knowing all of that from one place, continuously, rather than reconstructing it from memory when something forces the question.
That’s what “serious” means in this context. Not more AI. Not more pilots. More knowing.
What I’ve learned about AI in accounting
I said at the start, that none of this series was theoretical, and I want to emphasise that now the series is finished. Every one of the five pillars came from real conversations, with our real customers and prospects who’ve shared their situations honestly, with other people in this industry whom I interact with. There’s still always a little discomfort when I ask a firm how many agents they actually have running throughout the whole firm. The pause and eventual answer no longer surprises me.
Drowning in manual, repetitive work? Tell us the task and we’ll show you what to automate.
But, if there’s one thing that has surprised me, it’s how consistent the gap is. It’s not that firms are being careless. It’s that AI has moved faster than the management discipline around it could keep up, and most firms are still operating on exploration-era habits while running production-era AI. And that’s just where we are, collectively, right now. What’s important is what happens next. It will inevitably happen at pace, but it needs to happen safely and controlled too. We need to slow down, in order to speed up.
What’s next: an AI toolkit built specifically for accounting
We want to help make your AI program extremely practical for your accounting firm.
I’ve put together an AI toolkit built specifically for accounting and accountancy firms, drawing directly on everything in this series. It brings the five pillars together in one place, with direct links to the tools that go with each one: our AI Amnesty Survey, the Governance, Literacy, Tokenomics, and Lifecycle Assessments, and all the practical resources behind them, all live on our website.
The idea is simple: if you’ve read this series and recognised a gap or two, and I’d be surprised if you haven’t, the toolkit is where you go to actually start closing it, without having to work out where to begin on your own.
Where to actually start
If you’ve read all six parts of this series, you already know more about what getting serious means for AI in your firm than most of the market does right now. That’s not nothing, but it’s also not enough on its own.
So here’s the honest, practical starting point I’d give anyone reading this series: pick the pillar that made you most uncomfortable, and start there. Not the one that sounds most impressive or appealing to fix. The one that made you wince slightly when you read it, because that’s almost always the one closest to the real gap that you have. And if you have gaps in all of them, simply begin at the beginning, with Governance.
Run the Amnesty Survey if you genuinely don’t know what’s out there. Get a Governance baseline if you’re not sure what your AI can and can’t touch. Check your literacy gaps if you suspect your team’s AI confidence and competency is thinner than it should be. Model your token costs before scale forces the question. Take stock of your agents before one of them becomes abandonware nobody remembers building.
Whichever one you start with, start. The exploration phase gave everyone a taste of what AI can do, and what it can’t. The serious phase is about proving you can be trusted to run it, and that starts now, not once everything feels comfortable.
Daniel Lawrence is the CEO and co-founder of bots for that, creators of the Automation Operating System (AOS) and Agent Console. He has spent more than a decade deploying enterprise automation and AI in regulated industries including accounting and professional services. AI: The Serious Phase, is a six-part series exploring what AI-native operations should look like for mid-tier and large UK accounting firms.
