But while working on enterprise automation, we’ve been arriving at a different conclusion:
Enterprise adoption seems to come from constraint, not intelligence.
In practice, over 80% of enterprise data is unstructured: emails, documents, messages, transcripts, speech-to-text. When LLMs are used freely on this data, results are hard to trust or automate.
We’ve had more success applying strong constraints and what we’d call weak semantic grounding: using LLMs to detect predefined business signals and map them into fixed, verifiable outputs.
Dates. Events. Entities. Status changes.
Under these conditions, LLMs start behaving less like reasoning engines and more like semantic infrastructure — predictable, testable, and usable in real workflows. This insight also changed how we think about tooling. At Genum AI, we’ve been treating prompts as code: versioned, tested, regression-checked, and deployed like software. That discipline made the constrained approach workable in practice. We’re not claiming this replaces creative or open-ended GenAI — it feels complementary. But for automation-heavy environments, this seems to be where things actually scale.
Curious to hear from others here:
- Have you seen constrained LLM setups outperform open-ended ones in production?
- Is this just a modern take on classic NLP, or a new category enabled by LLMs?
- Where do you think this approach fails?