Joël Collin-Demers recently made a post that correctly stated that real “AI solutions” tell you exactly what they do, in which sequence, and [help you] understand how it solves your exact problem. the doctor totally agrees. Otherwise, they are using buzzwords and trying to cash in on the hype to sell you old-school automation at best, or third party (Gen-AI LLM) wrappers at worst, and don’t have any real AI.

He covered ten different types of technology and attempted to capture the positives, the negatives, and the best uses therefore in source-to-pay. For a relative non-techie (compared to the doctor with a PhD, degrees in CS and Mathematics, and actual experience implementing everything but BS LLMs from scratch), he got a lot right. But he got a few things wrong. As a result, there was a need to correct him (in this post) and ensure the corrections have as much permanence as the original post.

We do recommend you read his original post, but for each tech addressed, integrate the correction below.

RPA – does not “break” when processes change; it breaks when you feed it bad data; when you change processes, it simply loses its usefulness until you change it to match the process — with a good RPA system, that shouldn’t be hard

Machine Learning – requiring clean data is NOT a bad thing; it ensures the algorithm “learns” the patterns you need it to learn to use it effectively

Natural Language Processing – doesn’t struggle with jargon, just context — you define the dictionaries, the grammar, the language — it’s accuracy boils down to that; if you are feeding in documents that use the same words/phrases in multiple contexts, it will always struggle with that to a point

Predictive Analytics – in lay terms, classical predictive analytics is essentially just multi-dimensional curve fitting based on the data available — when something has not been modelled, there is nothing the algorithm can learn to fit against — but to be fair, nothing you’ve listed will succeed with unprecedented events/data

Anomaly Detection – this is based on outlier detection, and well trained outlier detection does NOT have high false positive rates (which are no higher than false negatives), and any “wrong” classifications from a business perspective simply means that the definition of a valid transactions needs to be amended to reduce the “outliers”

Computer Vision – lighting is not as much of a problem as you think as most good algorithmic interpretations will always mathematically adjust the brightness and contrast to a consistent range in pre-processing, and sometimes even greyscale; angles are a problem, because if they can’t be determined, the right transform can’t be applied to appropriately orient the image to maximize identification likelihood

Optimization Algorithms – “requires precise problem definition” is not a bad thing, it’s a good thing — if you don’t have a precise problem definition, you cannot get a precise answer with any technique; one of the best uses is product mix, not just supplier portfolio

LLMs – not “can” hallucinate false info; “will” hallucinate false into — every single time, just a question of the degree; it’s “generative” AI which literally means it makes stuff up, and how accurate what it makes up is with respect to your problem depends on how it was trained, what was asked of it, and how you ask it … very unreliable all around

Pattern Based Recommendations – the whole point is to “filter” to what you would normally buy so that you don’t have to sort through everything, that’s not a problem

Reinforcement Learning – it does not require extensive time, it requires extensive data — computers process mathematical calculations billions of time faster than we do — it’s never time anymore!