Maturing beyond LLM hype

Anyone who has built software knows that software is inherently complex, and the initial rush to adopt Large Language Models (LLMs) was no exception.

When generative AI first exploded, the industry treated LLMs as capable of solving almost any problem, applying them to deterministic databases with inconsistent or unreliable results. 

Today, we more widely accept that an LLM is a probabilistic text engine, and not a calculator. We integrate these models as capable components within broader software systems, rather than treating them as a universal fix.

‍

A tiered capability model for LLM integration

We evaluate LLM utility across three distinct maturity tiers to solve complex integration problems:

• Tier 1: Taming unstructured data: Extracting structured formats like JSON from messy PDFs and transcripts provides a better alternative to fragile regular expressions.
• Tier 2: Engineering and search acceleration: Developers use models to refactor legacy codebases and write unit tests, raising baseline productivity. Meanwhile, Retrieval-Augmented Generation (RAG) phases out keyword search by retrieving exact corporate policies and synthesising cited answers.
• Tier 3: Semantic routing: Models categorise user intent in the background, such as routing a customer service ticket to the correct department without user interaction.

‍

Assessing the operational risks of AI

We are selling insurance against failure as much as we are selling software, which means being upfront about where LLMs introduce technical debt.

• Rule-based logic: If a problem can be solved by a Python script or a SQL query, using an LLM increases operational risk and is needlessly expensive.
• Factual authority: Because models predict the next plausible word, they remain inherently indifferent to the truth. Relying on them for unverified compliance advice without strict oversight and grounding creates significant compliance and operational risk.
• Autonomous operations: Models drift off course without tight, deterministic loops guiding their actions, making fully autonomous long-term planning unviable.

‍

Actionable takeaway: Build a translation layer

A successful LLM approach targets the root cause of legacy system friction – the gap between human messiness and machine exactness. Modern architectures do not rely on the LLM for complex computations. Instead, the system uses the LLM to understand natural language, write the SQL query, retrieve exact numbers from a traditional database, and translate the output into an accessible summary. This approach provides a natural interface backed by the rigorous governance of a deterministic database.

We recently published a white paper about what it takes to move beyond experimentation and build practical, production-ready AI tools that create real business value. Read The Business of AI white paper.