15/04/2026
WHEN ARTIFICIAL IS SUPERFICIAL
As Consultants, we have been pushing the reasoning limits of AI platforms and engines. It’s not an accident therefore that we’ve seen the good, the bad and the really ugly.
The biases. The sycophancy. The hallucination.
Don’t even ask it to predict future outcomes based on specific statistical models you will require.
And no – this doesn’t mean it’s all bad. The ability to digest quickly from files in our NotebookLMs – it’s superb. But far from perfect.
It is the lack of consistency that bothers us and pushes our consulting team farther away from assigning any degree of trust on what it outputs.
It's prompting three times to validate before it recognizes an error. Three times, if you're lucky. The second attempt goes like, "I have reviewed all recent data and my analysis remains the same, etc." The third one? It's suddenly in tears, apologizing for the hallucination.
Guess what, it’s roughly the same experience from non-enterprise customers.
With all the marketing hype on AI, what’s not being given equal airtime is the implementation reversals. YES, REVERSALS.
AI solutions implemented to the detriment of customers. To the disdain of clients. To the utter disappointment of humans who were promised a reliable deputy.
KYC perfectly done, but implementation gone wrong.
When the tool hallucinates with confidence, apologizes on the third prompt, and still gets it wrong — that is not a deployment problem. That is a product defect in an expensive suit. I've sat across from enough failed transformations to know how this goes. The damage lands on the client. The invoice lands on the consultant. And the vendor drops a press release about the next model.
To call it intelligence is an affront to true human intelligence.
But since the word Intelligence is already out there and being widely used, I will simply call it for now SUPERFICIAL INTELLIGENCE.
AI will get there. we somehow still believe that. But "getting better" is not a warranty. And it is not a defense when you've already invoiced the client for something the tool could not deliver at quality and scale.
Until the tool is actually reliable — not impressively capable mostly on Tuesdays — there is no such thing as intelligent deployment.
There is only managed exposure to an unfinished product.
Do we hope it gets better? Absolutely! Purge the training junk, all that it has consumed from Wiki and Reddit junks. That's a good place to start.
Wait, wasn’t this the very approach advocated by IBM Watson?
AI is for now SI.