The first question every procurement team asks about AI in customer service is some version of: what stops it making something up. The honest answer is architectural, not a reassurance about model quality. A model asked to answer from memory will eventually produce something plausible and wrong. A model asked to answer only from retrieved, approved content will not, because it is not being asked to know anything.
Retrieval is a governance feature
Retrieval-augmented generation gets discussed as a technique. In an enterprise it is really a control: it forces every customer-facing answer to have a source you can point at. That is what lets a compliance officer sign off, and it is what makes an incorrect answer a content problem with an owner rather than a mysterious model behaviour.
The part that is actually hard
Retrieval is the easy half. The hard half is that your knowledge is wrong, duplicated and out of date, and always has been — it simply did not matter while only humans read it, because a human silently corrects for a stale document. An AI does not. Putting retrieval in front of unmanaged content industrialises your existing inaccuracy.
- Ownership. Every article has a named owner, or it will rot.
- Versioning and approval. Nothing publishes itself. What the AI serves is what someone approved.
- Ingestion with a schedule. Sites, documents and APIs re-crawled, so a policy change reaches the AI without anyone remembering to tell it.
- Gap detection. The questions the AI could not answer are the most valuable content backlog you will ever get, and they are free.
One source for people and machines
The failure mode worth naming is running two knowledge bases: one the agents use and one the AI reads. They diverge within a quarter, and then a customer gets one answer from the bot and a different one from a person, which is worse than having no bot. The same articles should serve both, which also means the humans notice when something is wrong — and they are the ones who will.