TL;DR
If you want to watch the full CRMKonvo, please go ahead here (optimized for smartphones) or here (optimized for tablets/computers).
Else, be my guest and continue to read.
Or do both …
The Domain Expertise Imperative: Correlation is Not Causation
One of the most dangerous, and frankly lazy, narratives pushed by AI maximalists is the idea that artificial intelligence negates the need for deep domain expertise. This is a fundamental misunderstanding of how these models work.
As Dr. Michael Wu frequently points out, almost all machine learning and AI systems today are built using supervised or reinforcement learning. They are, at their core, sophisticated correlation engines. They do not understand causality. They can surface 50 variables that move together, but they cannot tell you whether A causes B, B causes A, or if a hidden confounding variable C is responsible for both.
If an LLM correctly states that smoking causes cancer, it is not because it understands the biological mechanisms of cellular mutation; it is because it has been fed enough human-generated text asserting that relationship. It creates the illusion of causal reasoning without the substance.
This is precisely why domain experts, whether in healthcare, supply chain logistics, or financial services, are more vital than ever. The AI can process the data at unprecedented scale, but it takes a human domain expert to identify spurious correlations, recognize hidden causative factors, and make the final judgment calls. The machines are there to augment the experts, not replace them. So far, they simply can’t.
Deconstructing the Anthropomorphic Illusion
We must also fiercely reject the anthropomorphizing of AI. It is a fundamental human flaw to project human traits onto inanimate objects. It’s a psychological quirk that leads people to form emotional attachments to digital chatbots or, absurdly, even "marry" them because a machine, unlike a human, requires no compromise and has no expectations.
In the enterprise, this anthropomorphism manifests in the careless misuse of terminology. We call these systems "intelligent," giving ourselves an excuse to let our guard down. We use terms like "grounding" to imply that an LLM has a fundamental anchor in truth.
Let’s be clear: LLMs are not inherently "grounded." You can feed them contextual data, such as a causal graph or a vector database via Retrieval-Augmented Generation (RAG), to influence their outputs at inference time. This is valuable and important, but it is not true grounding. The underlying architecture remains a Transformer: a probabilistic, non-deterministic engine that samples from a distribution to guess the next word. If you ask it the exact same question twice, chances are it gives you two different answers. Heck, chances are that the answer is plain wrong. That’s why they all have a disclaimer. It is not reasoning; it is calculating probabilities. Until we develop architectures that sit on top of a truly grounded foundation of world logic, we must remain vigilant against the illusion of machine certainty.
The Myth of "Bigger is Better"
The market is currently obsessed with the idea that the path to Artificial General Intelligence (AGI) is simply a matter of scale; more compute, more parameters, bigger models. I disagree.
We are reaching the point of diminishing returns for massive, generalized models. A model that can write a Shakespearean sonnet, pass the bar exam, and debug Python code is intellectually fascinating, but functionally excessive for most enterprise needs. I do not need my supply chain optimization algorithm to “understand” 18th-century poetry, nor to code in Python.
The future of enterprise AI lies in rightsizing. Once we distill these massive models into smaller, highly focused, domain-specific architectures, perhaps utilizing a Mixture of Experts (MoE) approach, the cost to run them drops drastically. Smaller, smarter, and tightly scoped models deliver vastly superior ROI because they solve specific business problems without the bloat and compute costs of a massive frontier model.
The Value Equation: Digitization vs. Transformation
Ultimately, we must confront the value equation. Are we actually transforming our businesses, or are we just using AI to do the same inefficient things – just faster and more often?
If you take a broken, convoluted business process and simply layer an AI agent on top of it to speed up the keystrokes, you have not achieved digital transformation; you have merely digitized a bottleneck. True enterprise value comes from using these technologies to fundamentally reimagine workflows, ask questions of your systems that were previously impossible to query, and empower employees to operate at a higher strategic level.
Governance in the Age of AI
Finally, we cannot ignore the geopolitical and regulatory backdrop, such as the EU AI Act. While some decry regulation as a handbrake on innovation, pragmatic risk frameworks are essential; my emphasis is on pragmatic here. Bad actors will not slow down, so neither can we. However, we must limit our deployments to lower-stakes environments until we fully understand a model's limitations. You cannot hold an AI accountable when it makes a critical error in a clinical diagnosis or a financial audit. Accountability remains—and must always remain—with the humans who deploy it.
The score in 2026 is clear: AI is a powerful utility, but it is not magic. Enterprises that succeed will be those that pair domain expertise with right-sized, cost-effective models, applying them to genuinely transformative use cases while maintaining strict, human-led governance.
Three Things Enterprise Software Buyers Should Take Away
These are the hard truths for CX software buyers, paired with the uncomfortable questions they need to ask vendors to cut through their marketing noise
Call out the "intelligent" BS
Vendors love claiming their CX chatbots are "grounded" and can "reason". Spoiler: they can't. They are just probabilistic auto-completes guessing the next word based on correlation.
Ask the vendor:
Since your model is fundamentally probabilistic and non-deterministic, what hard, deterministic guardrails actually stop it from hallucinating a fake refund policy for our customers?
Are you using the term 'grounded' just to describe a basic RAG (Retrieval-Augmented Generation) setup, or is there an actual foundational world-model logic driving this?
Keep your human experts
AI is completely clueless about cause and effect. If you want to resolve complex customer issues instead of just failing at scale, you need your human domain experts to babysit the AI and catch its inevitable blind spots.
Ask the vendor:
LLMs don't inherently understand causality, they just mimic it based on training text. How does your system allow our domain experts to review, correct, and input hidden causal factors that the AI missed?
Is this tool designed to genuinely augment our tier-3 support agents, or is it just a generic copilot that's going to force them to spend more time auditing its mistakes than actually helping customers?
Stop paying for bloated models
You don't need a massive LLM capable of writing Shakespeare just to route a simple customer complaint. Demand smaller, domain-specific models. They do the job better, and you won't get fleeced on unnecessary compute costs.
Ask the vendor:
Why should we pay the massive compute premium for a generalized frontier model when a smaller, right-sized model trained specifically on CX data and use-cases would be cheaper and more effective?
If we strip away the massive scale of the underlying LLM, what unique, domain-specific value is your company actually bringing to our specific industry's workflows?
Comments
Post a Comment