Reducing Hallucination in Customer-Facing AI Agents
The scariest hallucinations aren't the obviously wrong ones — those get caught fast. It's the plausible-sounding wrong answer, delivered confidently, that a customer acts on before anyone notices. Early in a retail client's chatbot deployment, the agent invented a 60-day return window when the actual policy was 30 days, and it took three support escalations before we traced the root cause to a prompt that never explicitly said "only use the policy text provided."
The single biggest reduction we've seen from one change was forcing citation: every factual claim the agent makes has to be traceable to a specific retrieved passage, and if no passage supports the claim, the agent is instructed to say it doesn't know rather than fill the gap. That change cut hallucinated policy claims by roughly 70% in testing against a 200-question adversarial set we built specifically to probe for gaps in the knowledge base.
Temperature matters more than people expect for factual tasks. We run customer-facing agents at temperature 0 to 0.2 for anything involving policy, pricing, or account details, reserving higher temperatures for genuinely open-ended, creative parts of the conversation like small talk or product recommendations where a bit of variation is fine and even desirable.
We also added a self-consistency check for high-stakes answers: the agent generates a claim, then a second pass checks that claim against the retrieved context before it's shown to the user, discarding anything unsupported. It roughly doubles latency on those specific turns, which we accepted as a trade-off only for the categories of question — refunds, cancellations, legal terms — where being wrong is expensive.
None of this gets hallucination to zero, and we tell clients that up front. What it gets you is a system that fails safely — saying "I'm not sure, let me connect you with someone" instead of inventing an answer — which is the difference between an embarrassing gap and a real liability.
