When a business owner reaches out about AI, the second sentence is often a model name. “We want to use Claude for our customer service.” “Should we be on GPT-4 or Gemini?” “We need a RAG system for our knowledge base.”
These are good questions, but they are second-order questions. The first-order question — the one that frequently goes unasked, even by paid consultants — is “should we use AI here at all?”
The honest version of that conversation
Most AI tools are good at three things: summarizing text, classifying or extracting structured data, and generating new text under guidance. They are bad at things their proponents do not always disclose: deterministic logic, hard accuracy guarantees, sub-100ms responses without aggressive caching, and operating on data that does not exist anywhere yet.
A workflow that already works as a checklist or a spreadsheet rarely benefits from being replaced by an LLM call. The output is harder to audit, the cost climbs with usage, and the failure modes are subtler. A workflow that involves reading 80 PDFs to find the relevant clause, on the other hand, is one where AI dramatically saves time — and is also where conventional automation already had nothing to offer.
What we tell clients on the discovery call
Three filters help. First, can a human do this in 30 seconds with the right reference material? If yes, AI usually adds value. If a human needs an hour and specialized training, AI usually struggles. Second, is the cost of being wrong low (someone notices and corrects) or high (regulatory consequence, customer-trust damage)? AI is much better suited to the first kind. Third, does the volume justify the integration cost? A workflow run twice a week may not earn back the engineering effort.
This is not the answer the AI press releases promise. It is the answer that keeps clients out of expensive prototypes that go nowhere — and frees up the budget for the integrations where AI does fit.