What Is Contact Center Net Promoter Score (NPS)?

What Is Contact Center Net Promoter Score (NPS)?

Contact center Net Promoter Score (NPS) is a single-question customer-loyalty metric: “On a scale of 0 to 10, how likely are you to recommend us?” Responses split into promoters (9–10), passives (7–8), and detractors (0–6). The NPS is the percent of promoters minus the percent of detractors. Contact centers collect it via post-interaction surveys — IVR follow-up, email, or in-app — and report it weekly or monthly. It is a business KPI, not a model metric, but in 2026 it is increasingly tied to AI behavior since AI handles the interaction.

Why It Matters in Production LLM and Agent Systems

NPS is the most common board-level metric a contact center reports. When NPS drops two points, an executive asks “what changed?” In an AI-augmented contact center, the answer often involves an LLM, an ASR provider, or a routing policy — but the survey itself doesn’t tell you that. The gap between the business signal (NPS down) and the technical signal (which span regressed) is where customer-experience teams spend their week.

The pain shows up across roles. A CX VP defends the NPS number to the executive team without a clear cause. A product manager tries to attribute NPS movement to a recent model swap and has no per-trace correlation. An ML engineer is asked whether the new prompt regressed customer outcomes and can only show ConversationResolution dashboards that may or may not predict NPS. A compliance officer is asked whether AI escalations correlate with detractors and lacks the evaluator data to answer.

In 2026, AI handles 40–70% of first-contact interactions in many contact centers. That makes AI behavior a primary NPS lever. Without per-trace evaluator scores joined to per-customer NPS responses, NPS movement looks like noise — and teams optimize the wrong thing. The fix is to instrument every interaction with FutureAGI’s evaluators and join the eval scores to NPS at the customer-cohort level so cause-and-effect becomes visible.

How FutureAGI Handles Contact Center NPS

FutureAGI does not compute NPS directly — that lives in the survey platform — but it provides the per-trace evaluator scores that explain it. Each interaction emits a trace with ConversationResolution, Groundedness, ASRAccuracy, PII, and CustomerAgentQueryHandling scores; the contact-center system writes the NPS response back as a tag on the same trace ID. The dashboard then joins eval scores to NPS responses and produces a per-evaluator correlation: e.g. interactions with ConversationResolution < 0.6 produce 31 detractors per 100 surveys versus 8 detractors per 100 for >= 0.8.

A concrete example: a SaaS support team sees NPS drop from 42 to 36 over two weeks. They open the FutureAGI dashboard, slice traces by detractor responses, and find two patterns. First, calls where the LLM agent invoked escalate_to_human more than once produce 4× the detractor rate of single-escalation calls. Second, calls where Groundedness scored below 0.7 produce 2× the detractor rate. The team caps escalation count at one inside the agent’s plan, runs RegressionEval against the cohort that had been escalating twice, and tightens the RAG retriever’s top-k. NPS recovers four points the next survey cycle.

How to Measure or Detect It

NPS itself is a survey calculation; what matters is correlating eval signals to it:

• ConversationResolution: the canonical interaction-level outcome score. Strong negative correlation with detractor rate when below 0.7.
• Groundedness / Faithfulness: hallucination scores; ungrounded answers correlate with NPS detractors at higher rates than other failure modes.
• ASRAccuracy: voice-transcript quality; low-accuracy calls produce visible detractor clusters.
• Escalation count per trace: an OTel span attribute that signals customer effort; high escalation counts correlate with detractors.
• NPS-per-evaluator-bucket: dashboard signal that joins per-trace evaluator scores to NPS responses and shows detractor rate by evaluator quartile.
• NPS over time vs deploy markers: overlay model-version and prompt-version markers on the NPS time series to spot AI-driven inflection points.

Minimal Python:

from fi.evals import ConversationResolution

resolver = ConversationResolution()
result = resolver.evaluate(
    input="Customer wants refund on order 7621",
    output=conversation_transcript,
)
# join result.score to nps_response on trace_id downstream

Common Mistakes

• Treating NPS as a model metric. It isn’t — it’s a survey artifact. Use it as the outcome variable, not the optimization target.
• Reporting NPS without confidence intervals. Two-point swings can be noise; always report sample size and CI.
• Not joining NPS to trace IDs. If you can’t attribute NPS to specific interactions, you cannot localize the cause.
• Optimizing for promoters only. A “passive-to-promoter” lift is real; ignoring passives misses the easiest wins.
• Ignoring channel mix. NPS varies by channel — voice typically scores 5–10 points lower than chat. Slice before reporting.