Why your AI support agent hallucinates — and how docs-grounded answers fix it

We've watched dozens of teams deploy AI support in the last year. The pattern that kills trust fastest isn't slow responses or awkward tone. It's the agent confidently making things up.

A refund policy the company never had. A feature that doesn't ship for another quarter. A support email that routes to nobody. The agent isn't lying — it's predicting the next token based on what customer support replies usually sound like. It has no idea what your company actually does.

Why base LLMs hallucinate on support queries

Large language models are trained on public internet data. When you ask one a question about your product, it has three options: pull a near-neighbor answer from its training set, string together plausible-sounding words, or decline. Most models are tuned to be helpful, which means they rarely pick option three.

The result: the agent fills the gap with whatever shape of answer would normally appear in that context. For a refund question, it generates refund-shaped text. For a pricing question, pricing-shaped text. Shape ≠ truth.

What docs-grounded answers actually means

A docs-grounded agent doesn't generate answers from model memory. It runs a retrieval step first: your help center, changelog, internal docs, pricing page — indexed and searchable. When a user asks a question, the agent pulls the top 3-5 most relevant passages, then writes an answer constrained to that source material.

The key word is constrained. If the retrieved passages don't cover the question, the agent should say so — not improvise. That's the single biggest implementation detail separating agents that build trust from agents that erode it.

The three-step test for any AI support tool

• Ask it a question that IS in your docs. Did it quote the right source? Did it cite the source?
• Ask it a question that ISN'T in your docs. Did it admit ignorance and offer handoff, or did it hallucinate?
• Ask it a question about a competitor. Did it stay in-scope, or did it start rambling about other products?

Why retrieval fails even when it's wired up

Most RAG (retrieval-augmented generation) implementations break in subtle ways. The knowledge base is chunked too coarsely and retrieval pulls the wrong section. The embeddings are stale so recent changes don't surface. The agent retrieves 5 passages but the prompt template only uses the first 2. Each of these looks like the model is hallucinating, but the real bug is upstream.

Good retrieval is a system, not a feature. It needs: auto-reindexing on doc change, per-question confidence scoring, and an honest fallback when retrieval confidence is below threshold. Everything else is window dressing.

What we built into Signorian

Every answer Signorian gives cites the source passage inline. If the agent's confidence drops below the threshold you set, it hands off to a human with the full conversation attached — not a summary. And reindexing happens automatically when your docs change, so the agent never answers from a stale snapshot.

We didn't invent docs-grounded answers. We just refuse to ship a support agent without them.