Number of baby/toddler interruptions in writing this text: 0
One of the main motivators for starting this blog has been to give a nuanced view that is neither of these two:
AI will automate 50% of the knowledge workforce in the next 3 years
SPOILER: It won’t
AI is a hallucinating mediocre gimmick that should be nowhere near real business cases
SPOILER: It’s not
We live in a post-modern world, and a bi-polarisation of seemingly any issue on the planet seems to be a defining feature of postmodernity; a law as strong as gravity by this point. A lot of this is a result of us (society) reading algorithmically curated information pathways, with algorithms being optimised for engagement1. I’ll write more about that in a separate post - but to keep short - algorithms being optimised for engagement are going to be very close to the human subconsciousness - and in doing so - very close to the more primal parts of our brains. Our primal brain parts operate on a fight-or-flight basis and like being either very safe or to defend with ferocity a view that we hold dear - and that has all the ingredients you need for a negative feedback loop into bi-polar information bubbles.
To try to give a more nuanced view - I’ll try to outline, as of the time of writing this - how to think about Large Language Models (LLM’s - these are your GPT’s, Gemini’s, Claude’s and Grok’s of the world) in a business setting. I’ll do this by using the core principles I’ve outlined - so I’d strongly recommend reading that first if you’re going to dive into this content. This post will be an overview of the considerations as you examine your business case - and I’ll follow through with more detailed examination of the three business verticals where I’m seeing the most disruptive potential for an average organisation today - technology (hello CTOs), marketing (hello CMOs) and legal (hello CLOs). They might be separate posts though - just to keep the text size manageable.
Ingredients for success
It’s worth noting that the main reason I don’t believe we’ll automate 50% of the knowledge workforce is because most organisations don’t have a sane and digitised information architecture - and as I already stated - data infused with the most pertinent information is how you build successful AI systems.
This means - any organisation that does have a good information architecture and that has digitally embedded knowledge sharing - can reap AI in a lot more functional verticals than organisations that haven’t done that. My experience of corporations has, however, been, let’s say, less than inspiring in terms of how evolved data and information strategies they have. Many organisations share information in hallways in meetings - which is very socially engaging for humans, but it hides information from AI2. At least until transcription becomes ubiquitous.
However, some verticals will be better suited to get to a sane place faster. To start, let’s decompose the LLM.

An LLM has been trained on the entirety of human public3 written word (more or less) to predict the next word in an unfinished piece of text. Theoretically, you should be able to feed an LLM the first 50% pages of a novel you like, and it will be able to finish it with more or less similar structure and themes as the final product. It won’t be as good though, because LLM’s are regressive - they will tend to go to the most vanilla version of text you can imagine, that recycles tropes, common themes and creates something that looks super passable, but also super bland. I call it the “looks the same if you squint” kind of output.
You can typically make the LLM less regressive and more creative by upping its temperature parameter, or by prompt engineering to get it into the edges of the distributions it’s been trained on, but there be dragons in that creativity - as more temperature also means more hallucination.
That’s because an LLM doesn’t really have any concept of truth or any grounding in the real world. It just “knows” the distributions of language - so by upping the temperature, you’re making it go into the edges of both creativity and truth. If an AI lab was to separate those concepts - we’d be cooking; but to do that, they need a much better model of truth to embed into the LLM’s. Something that’s not necessarily trivial4. AI labs are trying though - but they do it by actually carefully triaging in the background. If your prompt looks like something that needs truth, they will lower the temperature in the background to give you back truth; but if you’re looking for a creative bedtime story for your toddler, they will dial the temperature right back up. That doesn’t solve the problem of being truthful, business aligned and creative at the same time, and I believe there is a ceiling as to how much they can optimise the LLM’s using this hardcoded set of protocols and without changing the core architecture underneath it.
So you might be tempted at this point to think - well, that means that I can either get the most bland output possible, that’s likely of no value, or the most unsafe output possible, that’s likely of no value. As always, there is nuance here.
Let’s start with the first ingredient - evaluate your [business / personal / academic] process.
Evaluating the process where you use AI
Some industries5 are a lot more riskier for LLM-powered AI, and we can call these industries non-starters6. Examples are frontline medical doctors7 , financial advisors8, financial traders9, judges10, accountants11. What all of those have in common is that the penalty of mistake is so high, that any risks makes it hard to even start contemplating an autonomous application. Typically, humans have regulated these industries to a high degree - with a very high barrier to entry and typically a certification needed to participate in the industry. In these industries, AI is currently limited to being a supporting tool to a human, rather than a generator of value through autonomous delivery, and typically, it would be deterministic AI12, rather than LLM’s, if any autonomy was attempted.
Another industry that has struggled with that is the autonomous car industry. Arguably13, the autonomous driving algorithms are already on average achieving a lower failure rate than humans, but the implications of it failing and killing someone are so vast (both ethically, legally and culturally) that it’s a non-starter. If it was deterministic, maybe, but the public mind just can’t live with the idea of a car randomly swerving into you, even if the likelihood is 1:100000000. To be clear, reader, I, personally, don’t think we should have autonomous cars on the roads, yet.
Some industries are a lot less risky - marketing, design, strategy, sales - all industries where lateral thought is appreciated - and success is not measured by producing something passable (which anyone could do) but by producing something extraordinary (which only quality industry professionals could do). Crucially - all of these industries are measured by quality of outcome, but there’s very low penalty for bad work. Bad work in these industries might not make you money - but it’s unlikely to lose you money, at least not on a one-off basis. If you were to go autonomous though, you’d likely start incurring losses as you start pushing out unusable interfaces, offensive marketing and company-busting strategies at scale. These are the industries where we see the most experimentation and free usage of chatbot-for-ideation pathways with a human-in-the-loop14.
And some industries have ingrained systems of command and control that were built to keep distributed human workforces in check and ensure high quality of output. Industries such as software engineering (which has been increasing per-developer productivity steadily for 60 years), customer operations, defence15. This is where we’re seeing the most success with full autonomy in parts of the workflow (i.e. agentic AI).
So the critical thing to understand as you’re thinking about something like an LLM - a very good word predictor - and your industry - is to think how you would ensure its success. Not how would you build it - you get an ML engineer for that - but how you would ensure success. If the penalty is too high unless someone is qualified - use (deterministic) AI as a tool and start generating data for the next wave of models. If the process is very creative with a very low penalty - have your humans use the AI to enhance their creative vision and start digitally encoding it - for the next wave of models. If you have good systems of command and control - start experimenting with autonomy with human operators at scale.
Evaluating the output
To achieve a higher return from AI - ideally you’d want to ensure that the majority of the work value output of an industry is digitally encodable and available at scale. You want to ensure that anything in the essence of the value of output is somehow digitally recorded. In other words - are there any choices that the worker had to make that are not readily apparent from the final digital output? Are those choices encoded as well? If not, the output will not be as informationally rich or available at scale. This could be a difficult concept to wrap your head around, so I’ll use some examples - but the question is philosophical in nature.

Good examples are software engineering and law - two industries where the output of the work (a document of code or a legal document) is extremely informationally rich and encoded as text. LLM’s are good at text - and the output of these industries leaves little to be desired outside of text. This is because in both law and software engineering, you need to make a lot of little choices continuously for each small part of the document. This makes the full document very informationally rich as to the decisioning process of the worker who produced it. In other words, software engineering and law are a product of an incredibly large amount of small relatively simple decisions that form a whole.
A bad example is something like a medical diagnosis - the output is very clear and thorough - but the choice of diagnosis doesn’t speak for itself - most of the output will be detail on the diagnosis, but that detail just flows from the fact of the diagnosis itself - which has been a result of many determinations in the background. A diagnosis document of 2 pages really only contains 1 piece of information about the diagnosis - the diagnosis itself. It would be different if the output document also contained a detailed accounting of the thought process of arriving at the diagnosis (which it doesn’t, typically). In other words, medical diagnosis is one very big decision with a lot of inputs and a complex decisioning process.
How the decisioning process is encoded in the output is extremely important - because AI are decisioning machines. And AI such as LLM’s that has been trained only on the output - will be much better at making decisions, if what it has been trained on includes the diversity of decisioning choices it needs to make.
Existing quality assurance in the process
Business systems that had a high level of deterministic automated quality assurance are business systems that are likely going to benefit most from a new way to automate things. This sounds trivial - but is not necessarily obvious to everyone.
The reason why this is extremely useful is because ML-powered AI loves feedback, and automated QA creates feedback at scale. It’s like a worker who is never emotional and corrects on feedback immediately. Imagine putting such a worker in an environment where any mistake they made has has an automated process telling them to correct it. Such a worker would become very good, very fast.

Something interesting happens here with LLM’s, explicitly because of the non-deterministic nature. When you give an LLM feedback, it will correct that, but it will also create a change somewhere else randomly - sometimes a change for the better, sometimes a change for the worse. But only good changes are accepted and not returned - and so the LLM reinforces16 itself to an optimal position, but there’s an element of an evolutionary algorithm17 as well, where random mutations might help it get to a more creative optimal position.
There’s not many industries that fit this pattern - but software engineering has famously automated QA to a very polished state in pursuit of greater efficiency and productivity over the last 30 years or so.
Conclusion
If the conclusion we’re building towards is obvious, congratulations reader - feel free to close the browser and start using agentic AI in your coding process.
For the rest - not to be coy, I strongly believe that, as of writing this, the biggest disrupted industry by the current AI wave is software engineering - and in an amazing way. Software engineers are already incredible economic value to the global economy - with tech companies being such a big component part of the global stock market that small chip changes move the entire market more than foundational breakthroughs in other industries18.
I believe agentic AI has product-market-fit - and as the AI labs optimise the agentic systems for lower cost and greater utility, we’ll see software engineering productivity sky-rocket. This might make people nervous - but I also strongly believe this will fuel growth and more hiring, rather than layoffs, in the medium term when the system stabilises. It should in any case, but capitalism has disappointed in the past.
In a way - we’re already seeing what this could bring - as it’s likely that the incredible pace of growth we’re seeing from the AI industry is fuelled by these efficiencies being embedded in the business process - imagine that embedded in the tech industry at large - we’d be seeing 10x the pace of technology evolution. That is equally exciting, and also, scary in its’ societal implications19. But likely inevitable none the less.
I’ll be looking into the software engineering disruption in more detail next week.
Remember that objective’s point from my core principles?
Incidentally, if you’re the kind of worker who either philosophically or out of existential dread, wants to postpone the AI revolution, the easy solution is to force more meetings
Copyrighted or not
More on that in a separate post
To help the text flow - I’ll be using industry to mean both industry and functional vertical, and I’ll use them interchangeably. That might confuse you, and while I do apologise, English is my second language and so my solution is to sod it
They’re absolutely not non-starters, but you might need an AI professional to very carefully, dare I say, with nuance, evaluate how AI should be applied within this domain
AI exists, but it’s deterministic. Experiments with non-deterministic are underway and…not moving very far
AI exists, but it’s deterministic. It’s called algorithmic advice and it’s been here for … actually quite a while. Most financial advice firms don’t actually manually go and figure out the best asset blend for you - that’s done algorithmically, the human is the glue that makes the financial mumbo-jumbo make sense to you
You might’ve heard about algorithmic trading - and this will differ from firm to firm, but it would be very rare to use an LLM for algorithmic trading. Typically, when you give an algorithm access to billions of dollars, you want it to be a bit more deterministic
You might think that AI would actually improve this profession, but probably not in a fully autonomous fashion. Can you imagine the outcry if humans were judged by AI without any humans in the loop?
Accountants are a very interesting one, that I might write a post about - but basically, with an accountant, you’re not buying the skill, which is fairly trivial in the majority of it, you’re buying the trust that things won’t get overlooked
Did you know that deterministic AI has been around for at least 60 years? Ever since the first decision tree was implemented to funnel customers based on a questionnaire, we’ve had deterministic AI optimising human processes; it was just very limited in capability
I don’t actually believe this, as there’s been a lot of data cooking in autonomous driving
Human-in-the-loop is the type of AI usage pattern where the AI is not given any autonomy in solving the task
This one might shock you, but the defence industry has the most refined command and control structures. I for one don’t think we should be rushing into defence AI (looking at you Palantir), but I can see the appeal and the reasoning behind it.
Reinforcement learning is a concept in machine learning, but you can just think of it as the system being rewarded when it does something good and penalised when it does something bad, and it learns through repetition
Evolutionary algorithms are a concept in machine learning, but you can just think of them as algorithms where a random change is made - and evaluated. Over many generations of trial and error, the algorithm is supposed to evolve to an optimal state
Reader, please don’t confuse me using this phrasing as me thinking that the global stock market is a good indicator of societal value of the economy.
More on that in a different post