AI agents for customer support · Part 5 of 14
RAG for support teams, in plain words
What happens between a customer’s question and the agent’s answer, where the lookup goes wrong, and why a limit written in an article is not a check.
The short version
RAG is how an AI support agent answers from your knowledge. It searches for the few pieces that best match the question, adds them to the conversation, and writes its reply from them. That works well for answers. It doesn’t work for rules: a limit written in an article holds only when the search finds it.
RAG in three words
RAG stands for retrieval-augmented generation: the agent looks things up, then writes. Ada, for example, documents that its agent keeps the three most relevant pieces and writes its reply from them. When nothing matches, it asks the customer to rephrase or hands over.
A customer asks
1 Retrieve
The agent searches your knowledge, split into pieces, for the best matches.
2 Augment
It adds the few pieces it keeps to the conversation.
3 Generate
It writes the reply from those pieces. If nothing matches, it asks again or hands over.
Where the lookup goes wrong
When knowledge that is right gives a wrong answer, look first at which pieces the search found. Three things go wrong there.
1 Different words
“Can I send it back?” never matches the article titled Returns policy.
The fix: title articles with the words customers use (part 4).
2 Another customer’s answer
A customer in the US asks how long they can return an item, and the search finds Returns in the EU: 30 days.
The fix: give the agent the customer’s country from your systems (part 6).
3 Yesterday’s copy
The agent reads your website as of its last sync, before this morning’s new delivery fee.
The fix: know when each outside source updates.
Zendesk, for one, reads an outside source as it was at the last sync, which usually happens every 24 hours. If your tool shows the source of each answer, check it on 20 real questions before launch; part 10 covers testing in full.
Why a rule in an article is not a check
Say your returns article says “refunds up to €30 need no approval”. To the agent, that is one more piece to find and weigh, not a limit. It holds only when the search finds it, and it can’t count what this customer has already received.
“Refunds up to €30 need no approval”
In an article
- Applies
- only when the search finds that piece
- Read as
- words the model weighs against the conversation
- Counts
- nothing: it can’t see what this customer already got
- Changed by
- anyone who can edit the help center
As a rule
- Applies
- on every refund, before the tool acts
- Read as
- a yes, a no, or a person’s approval first
- Counts
- this customer’s earlier refunds, across every agent
- Changed by
- a named owner, who confirms each change
Intercom says the same about instructions to its own agent: “You can write phrasing like “Don’t use [article],” but Fin may still use it if it’s the best match.” Its answer is a setting: take the article out of the sources Fin may use.
Intercom, “Provide Fin AI Agent with specific guidance”So keep in the article what the customer needs to know, such as how long a refund takes, and keep the limit as a rule your tools check before they act. Part 8 shows how to write one, and part 3 why the count has to cover every agent.
Next in this series
Part 6: connecting your helpdesk, store and billing. See all 14 parts.
Where Himmel fits
Himmel is the right-hand column: your team confirms each limit in plain words, and before every refund the agent’s tool checks it with Himmel, counted per customer across every agent.
Sources
- Ada, “Content generation”, read on 19 October 2026.
- Zendesk help, “Connecting knowledge sources to power generative replies in AI agents”, read on 19 October 2026.
- Intercom, “Provide Fin AI Agent with specific guidance”, read on 19 October 2026.