Definition
A feedback agent is an AI agent — a system that plans, retrieves, and acts — with a defined, narrow job on a feedback corpus. Indellia ships six: Theme Agent clusters open-text feedback into a hierarchical taxonomy; Anomaly Agent detects unusual spikes in volume or sentiment on a theme, SKU, or channel; SKU Agent links unstructured records to the specific product they describe across ASIN, Walmart Item ID, Model#, and UPC; Search Agent — indelliaGPT™ — answers natural-language questions with citations; Defect Agent (Beta) surfaces quality patterns that look like manufacturing or design issues; and Response Agent (Beta) drafts public responses to reviews and tickets for human approval.
Each agent produces structured output, runs on a schedule or on demand, and exposes its evidence. None is a general-purpose chat interface.
Why it matters
Most AI in VoC today is an LLM wrapper that answers whatever the user types into a prompt box. That shape pushes the burden of prompt design, grounding, and verification onto every user, every time. Results vary by operator skill and by prompt phrasing. Product, Insights, and QA each ask differently and get different answers. The tool ends up useful for exploration but untrusted for decisions.
Named agents reverse the shape. The job is defined in the platform, not in the prompt. A theme list is a theme list regardless of who runs it. An anomaly alert fires on defined thresholds. A SKU attribution follows one rule set. The team argues about decisions, not about whose prompt was better phrased. And because each agent does one job, the output is auditable — a reviewer can see exactly which records produced a theme, which threshold triggered an anomaly, and which matching rule linked a record to a SKU.
Example
A personal-care brand runs the Theme Agent weekly across Amazon, Walmart, Target, Bazaarvoice, and Zendesk. The agent returns a ranked theme list per SKU. Overnight, Anomaly Agent flags a 3.4x spike in a "dispenser clogging" theme on one SKU at Target. SKU Agent confirms the records are attributed to the correct UPC and not bleeding in from a similarly named variant. In the morning, the Insights lead opens indelliaGPT™ and asks, "What's driving the dispenser clogging spike on our body wash at Target?" and gets a cited answer drawn from the exact records behind the spike. Defect Agent (Beta) has already clustered the records as a probable fill-line variance and posted a draft note for QA. Response Agent (Beta) has drafted replies for CX to review. Four people, four different jobs, one shared evidence base.