“When the facts change, I change my mind. What do you do, sir?”
A 19th century economist, John Maynard Keynes, said that to a critic who accused him of flip-flopping. It is worth borrowing here, because the framework most people still use to answer “will AI take my job” was built on facts that have quietly changed underneath it.
In this article, I’ll explain why I think the facts are different and where we need to correct it, and I’ll begin with the question that a lot of people seem to be asking and posting about right now:
“Isn’t AI going to take my job?”
Personally, I think this is the wrong question. My reason for this, is that I don’t believe that a job is one thing, it’s not a single blob of activity. It’s a bundle of multiple, varied tasks. And so the right, but slightly layered question ought to be:
“What share of those tasks, can AI already do, at acceptable quality and cost, and whether the tasks left over still justify a dedicated, full-time, qualified human?”
In reality, it’s even more complicated than that, because when enough tasks can be automated to not require a full job, that’s not the end of the story, but let’s put that aside for now.
Anyway, this reframe is not new, it goes way back. It’s been debated, written, corrected and re-written about many times, in different circles, including economics and technology.
I’ve partaken first-hand in the debate directly over the past 15 years, and in more recent times, since generative AI models became commonplace.
So, the first correction to make: tasks, not jobs
Quick recent history lesson
In 2013, two Oxford researchers (Frey and Osborne) co-wrote an influential paper titled “The Future of Employment: How Susceptible Are Jobs to Computerisation?”. In this paper they estimated that 47% of US jobs were at high risk of automation. It has became one of the most cited numbers in the automation debate, but the research behind it was admittedly a little limited (some might suggest “shaky”), because it treated whole occupations as either automatable or not.
Jump forward a couple years, and we meet three economists and researchers, Arntz, Gregory and Zierahn, who co-authored another famous paper in 2016, titled “The Risk of Automation for Jobs in OECD Countries”. In this paper, they corrected the shortfall from the 2013 research.
They wrote that automation does not happen at the occupation level. It happens at the task level. Most jobs are a mix of tasks that can be automated and tasks that cannot, which means far fewer jobs disappear entirely than the occupation-level model implied.
This is the version of the argument most people know today, task bundles, not job titles. It is correct, but it’s also not the whole story, and it wasn’t in 2016 either.
Therefore, here comes the second correction: which tasks, and why
In 2003, three learned scholars, Autor, Levy and Murnane had already attempted to answer the “which tasks” question, a whole decade before Frey and Osborne. They developed a framework, abbreviated to ALM, that said:
“a task is automatable if a human can write down the explicit rule a machine then follows”.
This system has nothing to do with difficulty or skill. It’s about whether the task could be “codified”.
Under ALM, a bookkeeper reconciling a bank statement for example, is doing a routine task, however skilled they are at it, because the rule can be written down.
Drowning in manual, repetitive work? Tell us the task and we’ll show you what to automate.
Conversely, a partner sensing that a client is about to walk, or that a set of accounts simply looks wrong before they can say exactly why, is doing a “non-routine” task, however junior the underlying arithmetic is, because nobody could write that rule down.
However, skill and safety from automation were never the same axis. ‘Codifiability’ was believed to be the axis, and for two decades this has roughly been where the safety line sat, for accountants and most other skilled professions.
The break: the rule stopped needing to be written down
Here is the fact that changed. ALM’s whole framework assumed that a human had to be able to articulate what they did, before a machine can execute it. Modern AI does not work that way. Nobody needs to write down the rule for how a language model should draft a management report, or spot an anomaly in a general ledger, or judge that a set of figures looks off.
Most models have not specifically been given these rules. It inferred a statistical approximation of the pattern from other examples, which is a different expression of competence entirely, and it’s a route ALM’s model never accounted for because at production quality it barely existed until recently.
Economists have a name for the gap this exposes. Polanyi’s paradox: we know more than we can tell. Plenty of expert judgement has always been like this, real, reliable, and impossible for the expert to fully explain. That used to be the safest place to stand, precisely because it resisted codification. Autor himself flagged this shift around 2014 and 2015 (around the time businesses were getting to grips with Robotic Process Automation), arguing machine learning was starting to close Polanyi’s paradox, not by finally getting humans to state the tacit rule, but by making the statement unnecessary.
I think this matters enormously for accounting, because so much of what the profession considers “safe” from automation, was safe under the old definition specifically because it was tacit. Tacit is no longer the safe space.
Three categories, not two
So, the old question was binary: automatable or not. The honest question now sorts tasks into three categories.
Routine and automatable. Tasks with explicit, describable rules, that can be executed by machines (and I do not mean AI!). For example, reconciliations, transactions, returns preparation. This isn’t a prediction, it’s reality. Bank reconciliations (should or could) run at over 95% automatable. Bookkeeping and data entry should sit around 80% at least.
Tacit and approximated. Pattern-based judgement that used to be protected because nobody could explicitly write the rule down. This is the category the old framework got wrong, and it is the one causing most of the unease in the profession right now, because it was never supposed to be exposed. Reading a client, spotting that something in a set of accounts smells wrong before you can articulate why, these are being approximated with real, measurable competence, not perfectly, but well enough to change what “safe” means.
Accountable and irreducible. This is the category that actually holds, and it is not defined by whether a task is routine or tacit, simple or complex. It is defined by who carries the liability for being wrong. A model can produce a plausible judgement call. It cannot be struck off. It cannot be sued. It cannot carry a duty of care to a client or to HMRC. That distinction has nothing to do with how well the model performs and everything to do with who is legally on the hook when it does not.
What this means in practice
AI does not eliminate the accountant. It relocates value, away from tasks that were merely hard to codify and toward whoever holds the licensed signature and controls the platform the work runs on. The role does not get safer by being difficult. It gets safer by carrying accountability that cannot be transferred to software.
If you need convincing, run your own version of the exercise as a thought experiment – use AI if you fancy. List the 15 to 25 tasks that actually make up the majority of your role or your team’s roles. Sort each one into the three categories above. For anything landing in “accountable and irreducible” ask yourself honestly whether that is because the task is genuinely hard to delegate, or because nobody has yet configured the structure that would let AI do it and a human sign off on it responsibly.
That second question is the one most firms aren’t asking yet. UK firms are near enough unanimous that AI and automation use will keep rising over the next three years. I don’t feel that many feel confident they could actually assess what that will do to their own workforce. The technology is not the gap. The governance around it is.
This is important because I’m seeing and hearing everywhere how accountants and firms are getting frustrated with the lack of actual genuine and scalable progress with AI. That they’re looking for real, valuable use cases. That they’ve had enough of AI-washing and hype. What I believe organisations and firms need now, is not more tools, they need more clarity and fewer choices, they need genuine understanding of the problem they’re all trying to solve, and to apply that to the way they’re going to solve it – spoiler alert, it’s not going to be what you think.
So, the facts have changed. The old safety line, non-routine equals safe, does not hold the way it used to. What are you doing about it?
