“What if we put AI on this?” The phrase has been circulating in demos for three years. It works everywhere because you can effectively plug an LLM into anything. Pluggable doesn’t mean useful. In the majority of cases where it’s used, a regex, a business rule or a SQL join already did the job, better and for a thousand times less.
Five situations where I see AI projects that had no reason to exist.
Classification when categories are stable
Your customer service department receives emails and has to route them to the right departments. A sales representative offers you an LLM to classify the emails.
The question to ask: are the categories stable and well-defined? If so, a classifier trained on 200-500 labeled examples (or even a set of keyword-based rules) does the job with 90-95% accuracy, virtually no cost to inference, and total explicability (you know exactly why an email was classified in such and such a category).
An LLM costs 10 to 100 times more per request for a similar result. And you have less control over border behavior.
Generative AI is useful for classification when categories are unclear, numerous or change frequently. For stable, well-defined categories, classical solutions are better.
Extracting structured data from known formats
Extract amounts, dates and names from documents.
A regex or structured parser does this perfectly on known formats. It’s predictable, testable and deterministic. An LLM can do it too, but with a risk of hallucination on precise numerical values (dates, amounts), and a higher cost.
The exception: unstructured documents in highly variable formats (contracts written in natural language, e-mails). This is where LLM adds value. For invoices or forms with a known format, a structured parser is superior.
A search in an already well-structured database
Your product catalog has 10,000 references. Someone suggests you put in an LLM for research.
If your references are in a relational database with well-defined attributes (category, color, size, price), a SQL or Elasticsearch search is faster, more precise, more transparent, and infinitely less expensive.
LLM brings value for natural language search on unstructured content (documentation, articles, emails). For search on structured attributes, the classic search engine wins.
Generate volume content without human proofreading
Your marketing team wants to “generate product descriptions with AI” to speed up production.
Legitimate use when a human proofreads before publication. A risky use case when the output goes directly online. The difference between the two is not in the tool, it’s in the proofreading process that may or may not exist downstream. Without this step, you’re publishing factual errors on your own products, without even knowing it.
The pitfall: measuring the productivity gain on generation without measuring the cost of proofreading and corrections. Often, the net gain is less than the demo suggested.
The RAG chatbot on already poorly maintained documentation
“Let’s put a chatbot on our knowledge base so employees can find information more easily.”
If your knowledge base is well structured and maintained, an efficient search engine (Elasticsearch, Typesense) is often more accurate, less costly and more predictable than a RAG. RAG chatbots are useful when documentation is voluminous and varied, and questions are open-ended. They are less useful when questions are factual and answers are in clearly identifiable documents.
And if your knowledge base isn’t well maintained, the RAG chatbot will reproduce and amplify the inconsistencies. It’s a problem of organization, not technology.
AI as a solution of last resort
AI is the solution of last resort. Not as a matter of principle: as a matter of auditing discipline. Before signing off on a project, we need to have eliminated two alternatives. The deterministic business rule, which has worked for thirty years in the same kind of problem. And process reorganization, which sometimes solves the problem without a single line of code. If these two options have been examined and ruled out, then LLM is worth talking about. Otherwise, you’re doing AI for AI’s sake, and paying twice: the API bill, plus the technical debt of not looking for something simpler.