CSAT (customer satisfaction score)

CSAT asks one question at the end of a conversation: how satisfied were you with this interaction? Responses are aggregated as the share of positive ratings (typically 4 and 5 on a five-point scale). Helpdesk platforms collect it automatically, which makes it the most widely available customer-experience metric in support, though response rates are low and skewed towards the very happy and the very annoyed.

For AI support agents, CSAT is measured separately for AI-handled, escalated and human-handled conversations. The comparison is what matters: an AI agent whose CSAT trails the human team on the same intents is not ready for autonomy on those intents, whatever its deflection number says. It is also read alongside auto-resolution rate and reopen rate, because a satisfied customer whose problem recurs a week later is not a success.

CSAT is not NPS (loyalty to the company, asked periodically) or CES (how much effort the interaction took). It is transactional and immediate, and its main weakness is that it measures the feeling at the end of a conversation, not the outcome.

• Support manager: it is the customer-facing check on any automation change; a drop after AI rollout on a given intent is a stop signal.
• Founder / CEO: it is the metric that keeps cost savings honest, because it comes from customers rather than from the system.
• Product manager: low CSAT on resolved conversations usually indicates tone, speed or clarity problems rather than wrong answers.
• Operations head: segmenting CSAT by AI versus human versus escalated conversations shows exactly where the handover is hurting.

Automation metrics come from the system and can all be inflated by design choices; CSAT comes from the customer and cannot. It exists as the independent check that cost reduction has not come at the expense of the people paying for the product. The trade-off is that it is noisy, low-response and sensitive to survey wording and timing, so it is a guardrail rather than a target: you set a floor per intent relative to the human baseline and stop or roll back automation that breaches it. Eazyware reports CSAT segmented by handling path in every support-agent Care Plan review.

• SaaS: comparing CSAT for AI-resolved configuration questions against the human team before promoting the intent to full autonomy.
• E-commerce: tracking CSAT on AI-handled return initiations during peak season as volume shifts to automation.
• Bank chat: setting a CSAT floor for AI-handled dispute-status queries with automatic rollback to agent assist if breached.
• University helpdesk: measuring student satisfaction with an admissions agent across languages.
• Logistics: post-call CSAT via SMS after AI-handled delivery reschedules.

MetricA tracked experience metric, not a skill. Eazyware uses it as a per-intent guardrail on Customer Service Agents deployments, segmented by AI, escalated and human handling, and reviewed in every Care Plan cycle.

Eazyware service that covers it: AI Customer Service Agents. Starting prices are on the pricing page.