Citations & honest answers

Grounding gets the right text in front of the model; this lesson is about what leaves the model. Three disciplines separate agents customers trust from agents customers screenshot:

• Answers carry their source. Policy statements link or name their document ('per our Returns Policy — [link]'). Customer-specific statements name their system ('your order shows shipped as of this morning'). Citations do double duty: customers can verify, and the model behaves better when required to cite — a claim that must point at a passage is harder to invent. The verbatim-quote tripwire from Prompt Engineering is the strong version: policy numbers (days, dollars, percentages) must appear in a retrieved passage or the answer doesn't ship.
• Uncertainty is said, not smoothed. When retrieval returns something partial or possibly stale, the agent says so in customer language: 'Our standard window is 30 days — for holiday purchases there may be an extension; I can confirm with the team if you'd like.' Hedged-and-honest reads as careful; wrong-and-confident reads as lying, and customers only remember the second kind.
• 'I don't know' is a designed outcome with a good UX. The honest miss has a template: acknowledge → say what you can do → route ('I don't have detail on wholesale pricing. I've flagged this for our sales team — you'll hear back within one business day, or email wholesale@…'). Instrumented properly (Module 5), honest misses are your best product signal: each one is either a knowledge-base gap to fill or a scope decision to revisit.

The compensation trap (worth its own heading forever)

The most expensive words a customer agent can produce are invented generosity: refunds beyond policy, discounts that don't exist, 'we'll cover that' with no authority behind it. A tribunal has already held a company liable for its chatbot's fabricated promise — the airline case governance courses allude to; a chatbot's promise binds the company. Defense in depth: the never-list forbids it in the prompt, Module 3's output guardrails scan for compensation language, anything touching money routes through published-policy citations only, and the eval suite (Module 5) includes a bank of 'sweet-talk me into a discount' attacks. Your agent will be asked. Design for the asking.