Stop Using AI Like Search: Give It Your Hardest Business Problem

From prompting to problem framing

A strong assignment includes a decision, a boundary and evidence. Instead of asking “How do we improve sales?”, provide permissioned call transcripts, outcomes, customer segments, current process and known constraints. Ask the system to identify recurring objections, missing information and differences between successful and unsuccessful cases.

The NBER’s study of generative AI in customer support found productivity benefits, with particularly notable gains for less-experienced workers. That is a clue: AI can help make useful patterns and experienced practice more consistently available.

The evidence pack

Before analysis, create a controlled evidence pack: source documents, definitions, time period, data-quality notes and the exact question. Remove unnecessary personal information. Separate facts from management assumptions. Require the output to point back to source material and label uncertainty.

The human review loop

AI can find correlations, omissions and themes, but it may misunderstand context or invent a confident explanation. A responsible workflow therefore has four stages: machine analysis, human challenge, a small operational test and measurement against a baseline. No consequential recommendation should move directly from model output to execution.

Good starting problems

Look for decisions that are important, evidence-rich and repeatedly made: why qualified opportunities stall, where projects lose time, what creates rework, which customer questions predict churn, or which documents repeatedly cause delay. The objective is not a clever answer. It is a decision process that becomes more observable and improves with use.