A demonstration can look convincing for five minutes. An artificial intelligence system, however, must be useful when documents change, unexpected questions arise or someone needs to know where an answer came from.
Between the two lies less visible, more decisive work: preparing information, setting boundaries, measuring quality and designing how people will step in.
Quality starts before the model
If information is out of date, duplicated or poorly organised, a model will not magically fix it. A reliable project identifies authoritative sources, decides who maintains them and establishes what content must remain out of scope.
For systems that answer questions using an organisation’s own documentation, evaluation should check more than whether a response sounds plausible. Is it supported by the right sources? Does it acknowledge when it does not know? Does it avoid exposing data the user is not entitled to see?
Measure what matters
Before deployment, build a representative set of questions and agree on evaluation criteria: accuracy, usefulness, citations, response time and escalation rate. Then test the system with real cases, including awkward and exceptional ones.
Human oversight should not be a generic checkbox. Decide which choices need approval, how errors are reported and who has the authority to pause or change the service.
A living product
Launch is not the finish line. Data, team needs and model behaviour change. A production solution therefore needs owners, monitoring, a feedback channel and an update process.
AI moves from experiment to everyday tool when people can use it without being asked to trust it blindly. The “intelligence” matters; so do the operations, evaluation and people who take responsibility for its results.