Hallucination Mitigation

A practical example

Example: an agent that drafts renewal emails is barred from generating pricing or dates; those fields are pulled directly from the CRM, and any sentence the model marks as low-confidence is replaced with a neutral placeholder for rep review.

What to evaluate before investing

• Ask which specific controls the vendor applies — source restriction, output constraints, confidence thresholds — and where each is configured.
• Test whether the system can generate claims about facts you never provided, such as competitor pricing or customer logos.
• Ask how mitigation interacts with human review: are low-confidence outputs blocked, rewritten, or merely flagged?

Limitations and tradeoffs

Tradeoff: every mitigation layer narrows what the agent can do and adds latency or review workload. Over-constraining can make outputs so generic they lose value, so controls must match the risk of each use case.