AI & decision-making
What should a business clarify before funding an AI initiative?
Before asking which model to buy, decide what the system may do when it is wrong.
Published 25 September 2026
Consider a team that wants AI to answer customer policy questions. The model produces a fluent answer from last quarter's policy.
A bounded first use is drafting from current approved material for a reviewer to check. Automatic sending should not follow until error rates, review effort and exception handling are understood.
Describe the job without saying AI
Is the work finding a document, drafting a reply, classifying a request or making a decision? Identify who performs it now, how often it occurs, what information they use and what a mistake would cost. Some tasks need better search or a rule, not a model.
Draw a firm line between a suggestion and an action. A draft that a person can reject carries different risk from a tool that sends a message, changes a customer record or approves a payment.
Design a test that can fail honestly
Use representative cases, including stale documents, missing facts and awkward exceptions. Record accuracy, time saved or added by review, and how often the tool should abstain. A polished demonstration is not evidence of routine performance.
The funding decision should specify who owns source material, reviews outputs and can stop the use case. If those responsibilities are unclear, fixing them is the first investment.