Making AI Work
This is the first in a series of five articles on what AI transformation looks like in mid-market businesses, and the role a PE firm should play in making it happen.
I tried to write this article using AI, and it was a complete disaster. It produced some good sentences and paragraphs, but it blindly followed slightly illogical and contradictory instructions, resulting in a messy, barely coherent piece that would immediately read as AI-generated.
Given that this article is about using AI effectively – and that effective use starts with understanding how it works, it’s somewhat ironic that I fell into exactly the trap I’m trying to help you avoid.
You see, the big problems with AI start from trying to use it for the wrong things and in the wrong way. It’s hard to avoid but it starts with understanding fundamentally what an LLM is and how it works. Let me kick off this article by explaining exactly what an LLM is and how it works.
What even is an LLM?
So, you’ve used ChatGPT, Claude, Gemini, Grok or any of a host of other LLMs but do you know what’s happening underneath the hood when you use an LLM? For most of us the answer is no and even for those of us who have taken the time to understand it, it really is still confusing.
What an LLM does is predict what word should come next. Sounds simple right? The reality is that it’s far from simple. To do this well, an LLM needs to “understand” what words mean. For example, the phrase, “I went to the bank” could be followed by “to deposit a cheque” or “to pull myself out of the water”. The meaning of the word bank is completely different, and it will fundamentally change what needs to happen after this.
So, an LLM has a rudimentary understanding of what words mean, given the context of the whole conversation. From this, they can then give you a range of possible words to choose from with a percentage probability that it’s the next word. Based on that probability, a word is chosen and the whole conversation is fed back into the LLM for the LLM to repeat the cycle and predict what comes next.
Somehow, in a completely mindboggling and confusing way, this results in output that is somewhat intelligent.
What are the consequences?
Incoherence and hallucinations are fundamental to a model that randomly chooses potential answers to a problem based on predicting what word comes next. This is because an LLM has some understanding of the context, but that understanding is mapped to a number and doesn’t have the nuance and depth of a human.
In addition, while all the words that have been generated so far form part of the context, the LLM is fundamentally incapable of really making sure that what it’s doing at the end of a piece of output is actually consistent with what was happening in the middle and the end of the output. It struggles with narrative arcs that require consistency because there is the randomness that needs to get injected to make LLMs useful also lead to random diversions that don’t fit the narrative. Also, the only way an LLM can correct itself is by doing a task twice. An LLM can’t delete the words it’s just produced (like I just did while typing this).
This can lead to disappointing and surprising failures. These failures can be more dispiriting because once you start to use LLMs a lot you begin to become reliant upon them and your ability to provide critical thinking can degrade dangerously.
Is AI always an LLM?
Another reason for failures to adopt AI effectively is that the assumption now that we have LLMs is that they are the beginning, ending and everything you need when trying to implement technical solutions. The reality couldn’t be further from the truth and it’s important to understand that there are three distinct things that can get lumped under the AI banner, but they are different, with different outcomes and costs.
An algorithm. This is just a set of rules. This is what every single computer programme was before we started using other more advanced forms of AI. A lot of automated decision making used to be called AI but there is nothing inherently intelligent about these systems. They are as intelligent as the people (or AI model) that programmed them and they can’t do anything they weren’t deliberately designed to do. I can’t reiterate enough that this should be the main tool whenever you want to solve a problem, particularly if the problem is well defined, well understood and has simple inputs and outputs. Don’t try to get LLMs to perform tasks that should be pre-designed and follow a set of fixed rules as then you’ll burn money for inconsistent results.
A machine learning model. These learn patterns from data. Feed it enough historical examples and it can predict things. Churn, fraud, demand. In theory. They require a large amount of training and once trained they perform much like an algorithm because each time you give it the same inputs, it produces the same outputs. The difficulty with machine learning models is that there often isn’t enough data that is well labelled to provide good training. An LLM is essentially a very complex machine learning model, and it needed the entirety of all human produced writing available in digital form to train them how to work properly.
Moreover, ML models often perform not much better than heuristics and algorithms. This is especially true if what you’re trying to do is use AI to predict human behaviour. Human behaviour is classically unpredictable, and the rules of the game change faster than you can train an ML model. Have you tried using a simple algorithm before embarking on this complex road?
An LLM As you now know, is a word prediction engine that injects randomness into the response. Use these where there are not a clear set of rules to be followed and you need help with dealing with ambiguity and uncertainty. DO NOT use them to replace processes which are entirely rules-based. Use an LLM to produce the code that processes the rules but don’t inject LLMs where they don’t add value.
Going back to writing my article with an LLM. The LLM was the right technology, as it’s a problem with an uncertain outcome and without clear rules. However, it failed the first test where LLMs don’t necessarily have narrative consistency and hence the failure.
What is your AI Strategy?
This is a question that gets asked to companies all the time and the problem is that it’s a poorly defined question. An AI strategy is NOT one thing. What people are really asking is ‘how are you going to use LLMs to improve your business?’. Even that doesn’t help because there are (at least) 3 different areas and focuses that companies need to understand when planning out a strategy and these areas are typically distinct and will rarely form a single coherent strategy but actually 3 different strategies with different timelines and different owners, goals and deliverables.
These three areas are:
*AI in your product *AI for operational transformation *AI for product development
These three areas are related and will often (but not always) use similar tools, but if you try to conflate them into one strategy with a single owner and a single set of KPIs then don’t be surprised that everyone is very confused about what you’re trying to do and that the results are underwhelming. You need to split them up and have a separate strategy for each. The rest of the articles in this series will talk more about how you do that.
Is understanding that important?
I can guarantee that if the people who oversee championing and pushing AI within an organisation (including the CEO) don’t understand AI then they will use it to solve the wrong problems and they will create bigger problems than they ever solved.
If you can properly understand what LLMs are, whether you should be using LLMs at all and what strategic area you’re trying to work within then you give yourself a fighting chance.
Next: Part 2 – AI in Your Product: Data Moats and Why Most AI Features Are Worthless
The series:
1. You Need to Understand AI Before You Implement AI
2. AI in Your Product – Data Moats and Why Most AI Features Are Worthless
3. AI for Operational Transformation – Making the Entire Business AI-Capable
4. AI for Product Development – Why Your Engineering and Product Teams Need to Change Now
5. The AI Transformation Toolkit – Frameworks, Checklists, and What We’re Actually Building
