We’ve been over this many times here and on LinkedIn. And we are going to go over many of the reasons yet again, because it seems we have to. But first, we’re going to back up and give you a fundamental answer as to why Agentic AI cannot exist with today’s technology.

The dictionary definition of an agent is generally either “a person who acts on behalf of another” or “a person or group that takes an active role or produces a specific effect“. It’s about “doing, performing, or managing” for the express purpose of a “desired outcome“.

Classically, an agent is a person, and a person is an individual human and a human, unlike a thing, is intelligent, which today’s technology is not and that should be Q.E.D.

However some “modern” or “progressive” dictionaries (and there are at least 26 publishing houses publishing dictionaries as per Publishers Global), will allow an “agent” to be a “thing” as the term has been used in technical diagrams and patents, so if you define AI as a thing (which is being nice), then this refutation is not enough so we have to consider the rest of the definition.

As for “doing, performing, or managing“, like any automated system, it, unfortunately, does something. However, unlike past “automations” which took fixed inputs and produced predictable outputs, today’s Gen-AI takes an input and produces an unpredictable output that depends on how it was trained, parameterized, implemented, interacted with (down to the specific wording and format of the prompt as they are designed to continuously learn [which is a huge misnomer as they do not learn, but simply evolve their state]), fed the data, etc. As a result, two systems fed the same inputs at any point in time are not guaranteed to give the same, or even similar, output and are NOT guaranteed to produce the “desired outcome” (as even one single “bit” of difference between data sets, configuration, and historical usage can completely change an outcome, just like one parameter can completely change an optimal solution in a strategic sourcing decision optimization model). While this can be argued unlikely (and is for run-of-the-mill scenarios), it’s not infeasible (and, in fact, has a statistically significant chance of happening the more an input is off from the norm, and even 1/100 is significant if you plan on pushing [tens or hundreds of] thousands of processing tasks through the agent). Just like one wrong cost (off by a factor of 100 due to a decimal error) can completely change the recommended award in strategic sourcing decision optimization, because these are essentially super sized multi-layer deep learning neural networks (with recurrent, embedding, attention, feed-forward, and maybe even feed-back layers) based on non-linear and/or statistical activation functions where the parameters change based on every activation and, thus, every input, one wrong input, in the right situation, can completely change the expected outcome.

As for the arguments that “they often work as well as people, so why not“, that’s equivalent to the argument that “dynamite often works as well as excavators, so why not” or “loop antennas work just fine for direction bearing on aircraft, so why not“. In the second argument, the “why not?” is that even the best expert can’t always predict the full extent, direction, or result of the explosion (which, FYI, could also create a shockwave that could damage or take down nearby structures). In the third argument, ask Amelia Earhart how reliable they are if the short range is too short. Oh wait, you can’t! (And that’s why modern planes have magnetic compasses and GPS in addition to radio-based navigation.)

As for the arguments that the issues aren’t unique to AI and show up in other systems or people, while that is fundamentally true, there is no other technology ever invented that collectively has all of the issues we have so far identified in Gen-AI … and I’m sure we aren’t done yet!

[ Moreover, when you think about it logically, it makes no sense that we are so intent on pursuing this technology after having spent almost a century designing and building computing systems to be more accurate and reliable than we could ever be (with stress tested design techniques, hardware, and generations of error correcting coding to get the point where a computer can be expected to do trillions of calculation without fail, whereas an average human, who can do an average of one simple calculation a second, might not get through 10 without an error) while continually enhancing their capacity to the point that the largest supercomputer is now one quintillion times more powerful at math than we are! ]

As per a previous post, for starters, this technology:

1. Lies. It hallucinates every second, and often makes up fake summaries based on fake articles written by fake people with fake bios that it generates in response to your query, because it’s trained to satisfy (even though the X-Files made it clear in Rm9sbG93ZXJz how bad an idea this was back in 2018, four months before GPT-1 was released). And that’s the tip of the iceberg of the issues.

2. Doesn’t Do Math. These models don’t do math as they are built to combine the most statistically relevant inputs, not to do standard arithmetic. Plus, they aren’t even guaranteed to properly recognize an equation, especially when words are used. So it will miscalculate, and sometimes even misread numbers and shift decimal points. (Just ask DOGE, even though they won’t admit it.)

3. Opens Doors For Hackers. Over 70% of AI code has been found to be riddled with KNOWN security holes. (So imagine how many more holes it’s introducing!) It doesn’t generate code better than an average developer, and sometimes generates code that’s even worse than a drunken plagiarist intern who can at least identify a good code example to base the generated code on.)

4. Puts Your Entire Network at Risk. If you’re using it to automate code updates, cross-platform integration and orchestration, or system tasks, all it has to do is generate one wrong command and it could lock up and shut down your entire network, no Crosslake required!

5. Helps Your Employees Commit Fraud. They can generate receipts that look 100% real, especially if they do their own math (and ensure all the subtotals and totals add-up and the prices are actual menu prices), look up the restaurant name and tax codes, and have a real receipt example to go off of. (And as for the claims by Ramp and other T&E firms that they can detect fake receipts generated by Chat-GPT, good luck with that, because all the user has to do is strip the embedded metadata/fingerprint by taking an image of the image or running it through a utility that strips the metadata/fingerprint or converts the image format to one that loses the metadata.)

Can you name another technology that comes with all of these severe negatives (with more being discovered regularly, including addiction and a decrease in cognition as a result of using it). (We can’t! Not even close!)

[ And going back to the “people do all of this too” argument, it’s true we collectively do, but the vast majority of humans are not narcissistic sociopathic psychopathic robber barons with delusions of grandeur and no moral code or ethics. (Most criminals and con artists have a code or a line they won’t cross. To date, Gen-AI has proven such a concept is beyond it.) ]

And for those of you who believe Gen-AI is emergent, it has been refuted multiple times. There is no emergence in these models whatsoever. They were, are, and will always be dumb as a doorknob while being much less reliable. (At least a doorknob, when turned sufficiently, will always open a door.) If you want to believe, go find religion (and keep your religion out of technology) … or at least restrain yourself to the paranormal. When it comes to Gen-AI, THERE IS NOTHING THERE.

As for it doesn’t have to be smart to be useful, that is one of the most useless statements ever — because it’s context free and could be 100% true or 100% false or anything in between, all depending on the context you are referring to. By definition, technology does not have to be smart to be useful. Every piece of technology we use today is, by definition, dumb because it all lacks intelligence. (And some of it is so useful we’d never live without it — like control systems for energy grids and water works, air traffic, and modern communications).

However, all of the dumb technology that we have developed that is useful is ONLY useful with a precise context that defines the problem to be solved, the inputs it will expect, and the outputs that can be generated. An online order system is useless in a nuclear power plant control station, for example. (And while you think that is far-fetched, many of the areas vendors are claiming you can apply Gen-AI to are even more far-fetched.)

Even if the situations where the task you want ultimately boils down to the digestion, search, summarization, and/or generation of outputs based on a very, very large corpus of data, which is essentially what LLMs were designed to do, they are still only useful as a starting point or suggestion generator, or guide. They are not perfect, are totally capable of misunderstanding the question, being too broad in their interpretation of one or more inputs, being incorrect in their interpretation of the prompt for the desired output, generating fake data despite requests not to, excluding key documents or data from a result, or including irrelevant or wild-goose documents or data in the result.

In the best case, an LLM query will be about equivalent to a traditional Google search (that weights web pages based on key words, context, link weight, freshness, etc. and returns only real links in its results) of potentially relevant data. In the worst case, it’s a mix of real relevant, real irrelevant, and a lot of made up results, and you have no clue which is which until you check every one. (Which means it is ultimately less useful than a drunken plagiarist intern who will only refer to and copy existing references when sent on a research project, and all you will have to do is filter out the irrelevant from the relevant, as such an intern will be too lazy to make anything up completely.)

Since Gen-AI LLMs are what are powering all of the Agentic AI claims, it should now be quite clear as who why Agentic AI is a fraud!

Now, this isn’t saying Gen-AI is bad (as it will have some solid use cases where it is dependable with further research; as with all generations of AI tech before, it needs more time; nor is it saying that we won’t have smarter software agents in the future that can do considerably more than they can today), just that we aren’t anywhere close to being there yet and that they won’t be based on Gen-AI as it exists today.

So for now, as Sourcing Innovation has always advocated, switch your focus to Augmented Intelligence and systems that make your humans significantly more productive than they are today. The right systems that automatically

1. collect, summarize, and perform standard analysis on all of the data available;
2. make easily modified suggestions based on those analysis, similar situations, and typical organizational processes (that can then be easily accepted); and then
3. invoke traditional rules-based automations based on those accepted scenarios and invoke the right processes to implement the defined workflow

can, depending, on the task, make a human three (3), five (5), and even ten (10) times more productive than they are today, allowing for a team of two (2) to do the work that used to require a team of six (6), ten (10), or even twenty (20). For example, consider the best AI-based invoice processing applications that can automatically processes standard form invoices, break out and classify all the data, auto-match to (purchase) orders and receipts, auto-grab missing index/supplier data, auto-accept and process if everything matches within tolerance, auto-reject (and send back with reason and correction needed) in the case of a significant mismatch, and automate a dispute resolution process. When these, with the right configuration and training, will get most organizations to 95% touch-less processing (with zero significant errors), that’s a 10X improvement on invoice processing.

While not all tasks/functions will achieve the high efficiency of invoice processing, depending on the time required to collect and analyze data before strategic decisions can be made, most functions can easily see a 3X improvement with the right Augmented Intelligence technology. With the right tech (especially with supercomputing capabilities in modern cloud data centres), we have finally reached the age where you can truly do more with less … and all you need to do is NOT get Blinded By The Hype.

And for those interested, this latest rant was inspired by one of THE REVELATOR‘s Triple-Play Thursday LinkedIn posts and comments thereon (which have some deeper explanations in the comments if you want to dig in even deeper).