A faulty assumption I see echoed constantly in the current SaaS landscape: throwing generative AI at everything automatically means better outcomes. It doesn’t.
Gartner projects that over 40% of agentic AI projects will be canceled by the end of 2027. That level of failure doesn’t happen because the underlying technology is weak; it happens because teams reach for AI before understanding what kind of AI a problem actually requires. This thinking risks net-negative user experiences and erodes trust in products over time.
The most effective AI systems are built with intent and value first, and a core part of that is using three different methods of AI in an ensemble, rather than simply wrapping an LLM around a dataset. This requires a deeper understanding of the different methods of AI beyond what we call generative and how to best use them together.
When people say “AI” today, they almost always mean generative AI. But AI is much broader than that single subset of models. Understanding how the different methods work together is what separates systems that drive real value from those that just look impressive in demos.
One way to distinguish between different AI methods is to look at it through 3 chronological phases.
Phase 1 is Symbolic Intelligence, which is the earliest form of AI. This form of AI has explicit rules and deterministic logic. If a variable meets condition X, trigger outcome Y. You know exactly what happens, why, and when. It’s auditable, controllable, and predictable. It can’t learn or adapt, but it’s crucial. No hallucinations. The rules are the rules.
Phase 2 is Machine Learning. Here the model flips. Instead of programming knowledge top-down, you infer it from data bottom-up and let the machine find the patterns, and effectively create the rules that humans used to. Feed the system enough data signals, and it will predict churn, score purchase likelihood, or optimize send times. These models are efficient, fast, and mathematically rigorous for the problems they’re designed to solve. They do require clean data and human-defined features, and they don’t handle open-ended reasoning well.
Phase 3 is the Deep Learning and Foundation Model. This is the shift from hand-engineered features to large neural networks that learn their own representations from data. At scale, these networks exhibit emergent capabilities: language understanding, creative generation, and the ability to handle novel situations. Large language models are the most visible examples. They are extremely powerful, the most flexible general-purpose AI we’ve built so far. But they’re expensive to run, non-deterministic, opaque, and prone to hallucinations.
IBM found that only 25% of AI initiatives have delivered their expected ROI in recent years. I would question how often that was a result of a company deploying LLMs to solve a problem without considering whether it is the right tool: when you wield a hammer, everything looks like a nail.
Why Composition Beats Replacement
The industry narrative that phase 3 replaces the others is misguided.The best AI systems in production today, the ones that work at scale, in real time, with real money on the line are compositions of all three phases.
Phase 1 handles deterministic orchestration. When a system sends millions of communications, you need to know exactly what went out, to whom, and why. Regulators and policies like GDPR require this level of control. You don’t replace that with an LLM.
Phase 2 is where the biggest misconceptions live. People assume that because LLMs are “smarter,” they must be more accurate at everything. They’re not. For well-defined optimization problems in marketing like contextual personalization, churn prediction, purchase likelihood scoring, using classical ML is not only cheaper, but it’s also often more accurate. Contextual bandit algorithms, for instance, converge mathematically to optimal allocations with provable guarantees. That’s not what LLMs are designed to do. LLMs also don’t naturally produce calibrated probabilities from client-specific data. They aren’t trained on it, and fine-tuning at that level of granularity isn’t practical or cost-effective.
Phase 3 adds deep learning via neural network training as another layer of intelligence, not necessarily replacements. Trained neural networks can predict next-best actions by considering the order of user interactions, applying deep learning to phase 2-style optimization problems. LLMs handle text generation, copy refinement, and text book knowledge.
The clearest expression of this compositional approach is using generative AI not to replace the underlying systems, but to program them. A user describes a goal in natural language; the LLM translates that into a fully configured workflow that utilizes symbolic predicates, wires the appropriate closed loop ML models to execute, and generates creative content and variants using generative models. An agent backed by an LLM that can write a plan but has no infrastructure to execute it is just generating text. An LLM that can recommend a model configuration but has no ML layer to train and serve it is just hallucinating a roadmap.
Preparing for a Compositional Future
To solve most complex problems, we will need what I call “hybrid intelligence,” an ensemble requiring all three phases of AI. Remove phase 1 and you lose auditability and control. Remove phase 2 and you lose accuracy on the optimization problems that matter most. Ignore phase 3, and you can’t generate, or generalize to adapt to novel situations.
Before buying into hype cycles, recognize that not all AI is the same. Teams that believe throwing an agent wrapper around a generative model and expect this to replace everything will fall short of customer expectations. They may demo well for a few happy path cases, but they won’t work in production, at least not optimally. The organizations that build the most effective and defensible products will be the ones that layer all three types of intelligence into coherent systems.
